Changelog - QWED Documentation

All notable changes to the QWED platform, listed by release. Jump to: v9.1.0 · v9.0.0 · List agents and owner activity access · Session admin auth, key list, and user list · Audit CSV exports and natural-language domain · Owner sign-in from qwed init · Whole-chain trail scope and Verification Context audit rows · Startup hook digest pins · SQL and code verifier false positives · diff() and batch math verdicts · Claude providers stop sending temperature · AgentStateGuard JSON numbers · QWED-MCP v0.2.3 · v8.0.0 · Batch and agent routes enforce attestations · Audit paging and security events · Clients fail closed · Pinned sandbox image · CrossHair in the Docker sandbox · Signed audit chain for API rows · Logic DSL fixes · Proven-false math claims · GitHub Action scan paths · Compliance session sign-in · Security activity summary · Role checks and hashed agent tokens · Docker image runs the API server · Go, TypeScript, and Rust SDK admission gating · Signed Verification Context documents · Contacts moved to qwedai.tech · v7.3.0 · Admin endpoint tenant scoping · verify_fact fails closed offline · Request caps and sandbox limits · v7.2.3 · QWED-A2A runtime hardening · QWED-Tax v0.2.1 · QWED-Legal v0.5.1 · v7.2.2 · Open Responses v0.6.1 · QWED-Finance v3.0.1 · QWED-Legal date and input bounds · Open Responses v0.6.0 · QWED-Legal fail-closed hardening · qwed init server identity handshake · Leaked key auto-revocation · API-key format spec · Signed secret-scanning webhook · v7.2.1 · v7.2.0 · v7.1.0 · v7.0.0

v9.1.0 — Allowlist admission for Python code

Released: November 7, 2026 · GitHub PR #596 · GitHub PR #594 · GitHub Release

Minor release with one security fix. CodeVerifier now admits Python code only when every module, module name, builtin and special attribute it uses is on a list. Code that 9.0.0 admitted may now be refused. The server also reads .env before any of its settings. Read the upgrade steps below before you upgrade.

Security fixes

The advisory affects all versions up to 9.0.0. Upgrading is recommended.

CodeVerifier admits only listed modules, builtins and special attributes (GHSA-9r7q-mx75-hcj7, high). Reported by kta1kri. /verify/code returned VERIFIED · ADMIT for Python that read and wrote files with io.FileIO, sent data with urllib3, unpickled with pickle.Unpickler, started processes with multiprocessing and read os.environ. It refused code by lists of dangerous names and admitted anything those lists did not hold. Python code is now admitted only when:
- every module it imports, and every name it uses from one, is on the list of standard-library modules that compute without files, the network, processes, the environment or the import system. Third-party modules are not admitted.
- every builtin it reads is listed: the public exceptions, the builtins that compute, print, input, exit and quit.
- every special attribute (__x__) it uses is listed, and it reads no attribute’s value by a name made at run time. hasattr is admitted: it only answers whether an attribute exists, although the lookup can run a property.
- no module is used as a value or written to, and no frame or traceback is read.
The code-execution sandbox passes the third-party modules its import guard confines (numpy, pandas, scipy, sklearn, statsmodels, patsy) as trusted, so the Statistics engine still runs that code. The GitHub Action’s scan-code mode takes none of the new finding types, so repository scans don’t change.

What changed

CodeVerifier(trusted_modules=...). Names top-level modules that your own runtime confines, such as a sandbox with an import guard. The module checks skip a trusted module, so don’t rely on them for one. The builtin, attribute and frame checks still apply.
The server reads .env before any of its settings (#594). Before 9.1.0, the maths worker, database, rate limiter and auth settings were read before .env loaded, so a value written only in .env was ignored. For example, DATABASE_URL fell back to sqlite:///./qwed.db. The server now loads .env from the working directory, then ~/.qwed/.env, before any other module. Values set in the process environment (the shell, Docker -e, Kubernetes env, Compose) still win over .env.

Upgrading to 9.1.0. These changes can make /verify/code, QWEDLocal.verify_code and the CLI refuse code that 9.0.0 admitted. Each refusal is a finding that names the module, builtin or attribute.

Imports. Code that imports a third-party module (numpy, pandas, requests), or a standard-library module that is not listed (asyncio, logging, warnings, subprocess, importlib, pathlib and others), is refused even if it calls nothing from it. So are the os.path functions that read the file system, the working directory or the environment; sys.stdin, sys.stdout and sys.stderr; contextlib.redirect_stdout; pprint.PrettyPrinter; and star imports from copy, dataclasses, enum, functools and time.
Builtins. A read of open, vars, globals, locals, getattr, setattr, delattr, help, license, credits, copyright, breakpoint, memoryview or __build_class__ is refused, also where it is the name of your own variable or parameter. Rename it, for example to help_text.
Attributes. Special attributes that are not listed (__func__, __wrapped__, __get__, __reduce__ and others) are refused. .format and .format_map need a template written in the code with no attribute in a field. Use an f-string otherwise. copy.copy, functools.update_wrapper, ast.iter_fields, types.MethodType and dataclasses.make_dataclass are no longer listed. A class pattern with positional parts (case Point(x, y)) needs a class whose __match_args__ is fixed: a builtin or listed class, or a class statement whose bases are such and whose decorators are listed.
SDK cache. QWEDLocal doesn’t serve code results it cached before 9.1.0 (its cache policy version is now code_verifier-v3).
.env. Before upgrading, check .env in the working directory and ~/.qwed/.env for settings the process environment doesn’t set. The server now uses them: a DATABASE_URL there becomes the database the server opens, QWED_AUTH_TRUSTED_PROXIES there decides which proxies’ X-Forwarded-For the server trusts for the client address, and a maths setting the server can’t use stops the start.

v9.0.0 — Eleven advisories and models you configure

Released: November 6, 2026 · GitHub PR #589 · GitHub PR #575 · GitHub Release

Major release. It fixes eleven advisories in code, maths, SQL and logic verification, the GitHub Action and sign-up. QWED no longer names models of its own: each provider runs the model you configure. The CSV exports change their columns, and the Go module path ends in /v9. Read the upgrade steps below before you upgrade.

Security fixes

The advisories affect all versions up to 8.0.0. Upgrading is recommended.

CodeVerifier refuses modules taken by name from the import system (GHSA-qrp4-p44j-m484, high). Code that loaded a module by its name as a string, for example through importlib.util.module_from_spec, pkgutil.resolve_name, pydoc.locate or sys.modules.get(...), was VERIFIED · ADMIT. These names are now refused wherever they appear, and any read of sys.modules is refused.
CodeVerifier refuses eleven more modules that run strings as code, and more file readers (GHSA-jrxf-9jh4-86w7, high). Importing timeit, cProfile, profile, pdb, bdb, code, codeop, trace, doctest, logging.config or idlelib is refused. Archive, compressed-file, tempfile, sqlite3, shelve, glob and other file readers join the file-access checks.
CodeVerifier refuses code hidden in strings that typing evaluates (GHSA-p75q-7rg5-ff46, high). Strings that typing may compile, such as string annotations and ForwardRef arguments, are now parsed and refused when they hold a call, lambda, comprehension or assignment expression.
CodeVerifier refuses a source encoding other than UTF-8 (GHSA-pr6c-cqg8-6w5r, high). Under a utf-7 or unicode_escape coding declaration, text checked as a comment could decode to statements at run time. Such a declaration is now refused.
The Action’s scan-code mode resolves from-imports, aliases, importlib and getattr (GHSA-hg73-j3qv-w4vh, medium). It also runs CodeVerifier on each file. Attribute access by a computed name is reported as a warning in the new warnings_count output and doesn’t fail the scan. A pass means no known dangerous pattern was found, not that the code is safe.
The maths parser refuses unknown functions and gives n no hidden domain (GHSA-5h3v-3463-78hf, medium). sinc(0) = 0 and Abs(n) = n were VERIFIED · ADMIT. A name of two or more letters in call position must now be a supported function, a multi-letter name is one symbol, and n is a plain variable.
/verify/math and batch maths run in worker processes with a time limit (GHSA-qgjc-92m3-fcr6, medium). A job past QWED_MATH_TIMEOUT_SECONDS (default 10) is UNVERIFIABLE (math.timeout) and its worker is killed. A request that finds the pool full is BLOCKED (math.capacity).
verify_math no longer accepts wrong small values or wrong 7th digits (GHSA-8hp7-jqjc-hrhf, medium). 1/1000000000 = 0 was VERIFIED. The comparison is now unrounded, tolerance=0 is exact, and the result’s new match field says how a claim matched.
SQL tautologies written with IN, IS, LIKE, BETWEEN and other operators are refused (GHSA-jvhg-jw63-gr5g, medium). A filter predicate that reads no data and still can’t be evaluated is not admitted (sql_verifier.unevaluated_predicate).
Logic verification keeps numbers exact (GHSA-w73j-4jjq-q53j, medium). Number-only sub-expressions such as 1/3 or 0.1 + 0.2 were computed as floats before Z3 saw them, so prove_theorem proved false statements. Decimal literals and divisions of two numbers are now exact or refused.
Sign-up no longer prints the new user’s password hash (GHSA-mr5j-gm4f-6c27, low). Debug prints in POST /auth/signup wrote the user record, password hash included, to standard output. Treat hashes that earlier logs may have collected as exposed.

What changed

Each provider runs the model you configure (#575). QWED no longer falls back to model names of its own. Each provider takes its model from the argument you pass or from its variable, and raises an error naming the variable when neither is set.
Audit rows store real latency. /audit/logs returns the measured latency_ms of each row, and both CSV exports have a Latency (ms) column. Rows written before 9.0.0 return null.
Other changes since 8.0.0. 9.0.0 also ships the API key list fields (expires_at, rotation_required, owner), GET /auth/users, GET /agents and owner access to agent activity, qwed init --owner-email, org_name on sign-in and /auth/me, the shared CSV export columns, the NATURAL_LANGUAGE audit domain, default sampling for the Anthropic and Claude Opus providers, and one verdict for JSON numbers in AgentStateGuard. It also ships key rotation and the threat summary for a signed-in session, statistics verification and verify_python_deep on Python 3.14, diff() in /verify/math (and refusal of d/dx), SQL queries with a parenthesised subquery or a ; inside a string or comment, exec, eval and compile read as names instead of substrings, an audit row for a signed Verification Context document that verifies, and startup hooks allowed by the digest of their contents.

Upgrading to 9.0.0. These changes can break existing deployments and callers.

Set a model. Set the model variable of the provider you use: OPENAI_MODEL, AZURE_OPENAI_DEPLOYMENT, ANTHROPIC_MODEL, CLAUDE_OPUS_DEPLOYMENT, GEMINI_MODEL, OLLAMA_MODEL or CUSTOM_MODEL. With none set, qwed verify and qwed interactive stop with an error naming the variable unless you pass --model. On the server, ANTHROPIC_DEPLOYMENT still wins over ANTHROPIC_MODEL when it is set. With an explicit --base-url, qwed verify reads no model variable, not even CUSTOM_MODEL, and needs --model. For Anthropic the CLI reads only ANTHROPIC_MODEL or --model. The server still starts, because it builds a provider on first use; logic, statistics and fact requests that need the provider then fail with that error, and the natural-language routes answer UNVERIFIABLE. ACTIVE_PROVIDER, and any provider a request names, must be one of openai, openai_direct, openai_compat, azure_openai, anthropic, claude_opus, gemini, ollama or auto.
Pass model= in Python. QWEDLocal, verify() and qwed_sdk.integrations.langchain.QWEDTool raise ValueError without it. The qwed_sdk.langchain.QWEDTool that calls a QWED server takes no model. If qwed verify ran the saved provider with the key in QWED_API_KEY, set that provider’s key variable or pass --api-key.
Update CSV export readers. GET /admin/compliance/export/csv and GET /audit/export write a new, shared set of columns and at most the newest 1,000 rows. X-Export-Truncated: true is set when more rows matched.
Write * between names you multiply. xy is now one name, and n is a plain variable. A single letter before a bracket is still a product (f(0) is f*0). With no tolerance, verify_math needs a claim to match to 9 significant digits, or to be correctly rounded with at least 6, within 1e-6. VerificationEngine.verify_math’s tolerance defaults to None.
Size the maths worker pool. /verify/math and batch maths run in worker processes set by QWED_MATH_WORKERS (default 2), QWED_MATH_TIMEOUT_SECONDS (10), QWED_MATH_MAX_WAITING (32), QWED_MATH_WORKER_MEMORY_MB (2048, Linux only, 0 turns it off) and QWED_BATCH_MATH_BUDGET_SECONDS (120). A value these settings can’t use stops the server at startup with an error naming the setting. In 9.0.0, a server started with uvicorn qwed_new.api.main:app reads the four QWED_MATH_* settings before .env loads: a value written only in .env is ignored and the default runs, and a value there that the server can’t use makes every maths job BLOCKED (math.worker_error) instead of stopping the start. Set them in the process environment, or upgrade to 9.1.0.
Expect more code refusals. CodeVerifier refuses the modules, names and attributes in the advisories above, and a coding declaration other than UTF-8.
Check Action gates. scan-code also reports CodeVerifier’s code- and process-execution findings, refuses a Python file whose coding declaration isn’t UTF-8 or whose bytes aren’t valid UTF-8, and adds the warnings_count output.
Upgrade every server process together. Each audit row’s latency_ms is now covered by its entry hash, so an 8.x process can’t verify rows that 9.0.0 writes. An existing database gets the column at startup. Natural-language rows are stored with the domain NATURAL_LANGUAGE, not MATH; rows already written keep MATH. The audit row of a signed Verification Context document stores the document as returned, with any personal data in it; its query column is still redacted.
Update callers of changed results.
- A batch maths item proven false is VERIFIED with is_valid false and admission BLOCKED, as on /verify/math.
- Input the maths parser refuses is BLOCKED with math.parse_error, and Python that doesn’t parse is BLOCKED with code_verifier.parse_error.
- /auth/me answers 403 for an inactive user or organisation or an unknown role, and sign-in refuses a user of an inactive organisation.
- A password over 72 bytes is 400 at sign-up and 401 at sign-in, not 500.
Move to z3-solver==5.3.2.0. qwed now requires it (it required 5.1.0.0). The server image moves to sqlglot 30.25.0, anthropic 2.3.0, FastAPI 0.162.0, openai 3.41.0 and python-dateutil 3.0.0.
Update the Go module path. Change imports from github.com/QWED-AI/qwed-verification/sdk-go/v8 to github.com/QWED-AI/qwed-verification/sdk-go/v9, then run go get github.com/QWED-AI/qwed-verification/sdk-go/v9@v9.0.0. The Go, TypeScript and Rust SDKs have no API changes in this release.

QWED Verification — list agents, and owners read agent activity

v9.0.0 · GitHub PR #544 · Closes issue #527

Owners and admins can now list their organization’s agents and read any of those agents’ activity without the agent’s token. The activity route’s limit is now capped at 200.

What changed

New GET /agents. Returns every agent in your organization, oldest first, with id, name, type, status, created_at, and last_active. Suspended and deactivated agents are included. last_active is null until the agent first authenticates. The response never includes the agent token or its digest. Owners and admins can call it with an API key or a signed-in session. Members and viewers get 403 Forbidden, and a request with no credential gets 401 Unauthorized.
Owner access to GET /agents/{agent_id}/activity. The route also accepts an owner’s or admin’s API key or session for an agent in their organization. An agent in another organization, a missing agent, or an out-of-range ID returns 404 Not Found. When you send X-Agent-Token, the token alone decides access, as before. A wrong token returns 401 Unauthorized, even beside an owner’s session.
Bounded limit. The activity route’s limit now accepts 1 to 200 (default 20). It was unbounded. A value out of range returns 422 Unprocessable Entity.
Ordering and timestamps. Activity comes newest first, with ties broken by ID. Timestamps on both routes carry their UTC offset.

curl --fail-with-body http://localhost:8000/agents \
  -H "Authorization: Bearer $QWED_SESSION_TOKEN"

See GET /agents and GET /agents/{agent_id}/activity.

v9.0.0 · GitHub PR #531 · #540 · #541 · #549

A dashboard signed in with a session can now rotate keys, read the threat summary, and list the organization’s users without an API key. The key list and the user object return more of what QWED already stores.

Added

Session sign-in for key rotation and the threat summary. POST /admin/keys/rotate and GET /admin/security/threats/{org_id} accept Authorization: Bearer <jwt> with the same session checks as the compliance endpoints. QWED uses the API key when you send both, and returns 401 when you send neither. The role and organization rules are unchanged. POST /agents/{agent_id}/token/reissue still requires an API key.
Key list fields. Each key from GET /auth/api-keys adds expires_at (null when the key doesn’t expire), rotation_required, and owner ({id, email}). owner is null for a key with no user, or whose user isn’t in the key’s organization. expires_at, created_at, and last_used_at always carry their UTC offset. Who sees which keys is unchanged.
org_name on the user object. Sign-up, sign-in, and GET /auth/me return org_name, the organization’s display name, or null if QWED can’t read it.
GET /auth/users. Owners and admins list the users of their organization, oldest first, with an API key or a signed-in session. Each user has id, email, role, is_active, and created_at. Inactive users are included. Members and viewers get 403.

Fixed

Threat summary. GET /admin/security/threats/{org_id} no longer fails on every call with the pinned SQLModel release. Its generated_at is now UTC with a +00:00 offset.
Inactive accounts on /auth/me. GET /auth/me now applies the user-session checks. An inactive user or organization, or an unknown role, returns 403 instead of an answer from a still-valid token. Sign-in refuses a user whose organization is inactive with 403 "Organization is not active".
Passwords over 72 bytes. POST /auth/signup and POST /auth/signin no longer return 500 for a password over 72 bytes in UTF-8. Sign-up refuses it with 400 "Password must be at most 72 bytes", before checking the email. Sign-in answers 401 "Invalid email or password". An account whose password was set over 72 bytes under bcrypt 4, which used only the first 72 bytes, can’t sign in with that password.

See Session sign-in for admin endpoints, GET /auth/users, and GET /auth/api-keys.

QWED Verification — audit CSV exports carry verdict and admission; natural-language audit domain

v9.0.0 · GitHub PR #542 · GitHub PR #543 · GitHub PR #546

Both CSV exports now write the same columns, including the stored verdict, admission, and latency, and take a period. Natural-language audit rows get their own domain, and every audit row stores its real latency.

What changed

Shared CSV columns. GET /admin/compliance/export/csv and GET /audit/export write ID, Timestamp, User ID, API Key ID, Verifier Version, Domain, Query, Verdict, Admission, Is Verified (legacy), Result, Latency (ms), Entry Hash, HMAC Signature, and Chain Status. Verdict and Admission are the stored values, as GET /audit/logs returns them. A value QWED didn’t record is an empty cell. No cell is truncated.
Formula guard. A cell that starts with =, +, -, @, a tab, a carriage return, or a line feed gets a leading apostrophe, so spreadsheet software doesn’t run it as a formula.
Export period. Both exports take start and end. A time without a zone is UTC, and start after end returns 400. Rows are newest first. The compliance export defaults to the 30 days before end, and end defaults to now. It checks the period after those defaults, so a future start with no end returns 400.
Row cap. Both exports hold at most the newest 1,000 rows and set X-Export-Truncated: true when more matched, otherwise false. CORS exposes the header. The compliance export had no cap before.
NATURAL_LANGUAGE audit domain. /verify/natural_language, /agents/{agent_id}/verify, and natural-language /verify/batch items now write domain NATURAL_LANGUAGE (engine_used on /audit/logs, Domain in the CSV exports) instead of MATH. /verify/math and batch math items still write MATH. Rows already written keep MATH, because the domain is part of each signed, chained entry.
Real latency in the audit log. GET /audit/logs and GET /audit/logs/{log_id} returned latency_ms as a fixed 0. QWED now stores each row’s latency: whole milliseconds from the start of the HTTP request to when QWED began writing the audit row, before it waits for the append lock, including the verifier and any LLM call. Rows written outside a request, and rows written before this change, return null. Both CSV exports carry it in a Latency (ms) column after Result, empty when null. The entry hash covers latency_ms, so changing or clearing it fails integrity verification. Rows hashed before this change still verify. QWED adds the latency_ms column to an existing database at startup. Upgrade every server process together: an 8.x process can’t verify rows that 9.0.0 writes.

Breaking for CSV consumers. Update any client that reads the export columns by name or position.

Before: GET /admin/compliance/export/csv covered only the last 30 days, cut the query to 100 characters and the signature to 16, and had no verdict, admission, or result. GET /audit/export wrote is_verified as verified or blocked under Status, with no admission or caller.
After: Both write the 15 shared columns above in full. Read Admission for the decision. Is Verified (legacy) is the old flag as Yes or No.
Audit domain: Reports that group by engine_used or Domain now see NATURAL_LANGUAGE for new natural-language rows. MATH rows written before the upgrade can include natural-language requests.

v9.0.0 · GitHub PR #547 · Fixes issue #524

qwed init used to sign up with a generated email and password that it never showed or saved. Nobody could sign in to the organization it created, and its records could not be read in qwed-dashboard. You can now name the owner.

What changed

--owner-email. Names the organization owner. QWED_OWNER_EMAIL is read only when the flag is not given. An explicit empty or blank flag is refused, not replaced by the environment variable. Without either, interactive qwed init asks for the email after the organization name.
Hidden password prompt. qwed init asks for the owner password with hidden input and a confirmation. The password is never a flag, and qwed init never prints or saves it.
Non-interactive mode. With an owner email, the password comes from QWED_OWNER_PASSWORD. If it is missing, qwed init stops before the server starts. The local server qwed init starts does not inherit QWED_OWNER_PASSWORD.
No owner email. qwed init keeps the generated owner and warns that nobody can sign in to the organization. Running qwed init again with --owner-email creates a new organization and does not recover the earlier one’s records.
Checks. The email needs a name before exactly one @, a domain with a dot inside it, and no whitespace. The password must be 1 to 72 UTF-8 bytes. An email that is already registered fails with Email already registered and tells you how to continue.
Output. On success, qwed init prints [ok] Owner: <email> (sign in with the password you set).

export QWED_OWNER_PASSWORD="$CI_QWED_OWNER_PASSWORD"
qwed init --non-interactive --provider openai --api-key "$OPENAI_API_KEY" --model "$OPENAI_MODEL" --owner-email owner@example.com

See Organization owner.

QWED Verification — whole-chain trail scope and verifiable Verification Context audit rows

v9.0.0 · GitHub PR #551 · GitHub PR #550

Fixed

The whole-chain check names the whole trail. GET /admin/compliance/verify-trail with no start or end checks every entry of the organization, but its scope said “audit entries recorded in the period”. With neither bound, scope now reads “Signature and hash-chain check of every audit entry recorded for the organisation. It does not show that every request was recorded.” With start, end, or both, it still names the period. The security activity summary always has a period, so its audit_trail.scope is unchanged. Only the text changed. The status, counts, and proof_ref are as before.
The audit row of a signed Verification Context document verifies. POST /verification-context/from-diagnostic signs the document it returns, but its audit row stored a copy with personal data redacted after signing. That copy no longer matched its attestation, so a consumer reading it from GET /audit/logs couldn’t verify it.
- The row now stores the document exactly as returned. Its attestation matches the stored document, so the row verifies with POST /verification-context/validate until the attestation expires (QWED_ATTESTATION_TTL_SECONDS, 24 hours by default) and while the server keeps the signing key.
- The row’s query is still redacted.
- The stored document keeps any personal data you put in the query or the diagnostic, including its agent_message and developer_fields. To keep personal data out of it, mask the data in the query of the original verification and send that query unchanged here: for a VERIFIED diagnostic the attestation covers the query, so masking it only at this call makes the document BLOCKED.
- Rows written before this change keep their redacted copy, which may not verify against its attestation: it fails when the redaction changed a signed field.

QWED Verification — startup hooks can be allowed by content digest

v9.0.0 · GitHub PR #553 · Fixes issue #491

In a Python 3.11 venv, setuptools’ distutils-precedence.pth (which uses __import__() and coverage’s a1_coverage.pth (which uses exec() stopped the server at startup. Allowing them by name didn’t help, because an allowed name is still scanned for those patterns. The only way through was QWED_SKIP_ENV_INTEGRITY_CHECK=true.

What changed

Digest pins. QWED_ALLOWED_STARTUP_PTH_FILES now also accepts NAME.pth:sha256:<64 hex>. QWED allows such a file only while its exact contents hash to that digest, and then doesn’t pattern-scan it. A pin overrides a plain name for the same file.
Mismatches fail closed. A pinned file whose contents differ fails the check and is counted as digest_mismatch in the guard’s counts. A pin that isn’t NAME.pth:sha256:<64 hex> stops startup with a ValueError instead of being read as a name.
Plain names unchanged. A plain name works as before, and the suspicious patterns are unchanged. QWED ships no built-in digests, because these files change between setuptools and coverage versions.
Python SDK. StartupHookGuard takes pinned_pth_digests, and parse_startup_hook_allowlist parses the variable’s value.

See Startup environment integrity for how to take a file’s digest, and StartupHookGuard.

QWED Verification — SQL and code verifier false-positive fixes, and Python 3.14

v9.0.0 · GitHub PR #554 · GitHub PR #555 · GitHub PR #560 · GitHub PR #569

The SQL and code verifiers no longer refuse common safe queries and snippets, and statistics verification works on Python 3.14.

What changed

SQL subqueries are checked, not blocked (#554). A query with an IN (SELECT ...), a scalar subquery, or a subquery in FROM was BLOCKED with sql_verifier.execution_error. It is now checked like any other query. Nesting past max_subquery_depth (3) is BLOCKED with sql_verifier.complexity_limit_exceeded. complexity.subquery_depth now reports 0 for a UNION of two plain SELECT statements. See Subqueries and complexity limits.
A ; in a string or comment is not a stacked query (#555). A ; inside a string literal, a dollar-quoted string, or a comment no longer marks the query as stacked_queries. Two real statements still do. A query with a backslash, a -- not followed by ASCII whitespace, a nested /*, a lone carriage return, or a parse failure still counts every ;. See Stacked queries.
exec, eval, and compile are matched as names (#560). CodeVerifier matches eval, exec, compile, __import__, getattr, and setattr on the AST, not as substrings. cur.execute(...), model.evaluate(x, y), and re.compile(...) in a snippet that imports re are no longer dangerous_function. Any line containing __import__ is still reported as dangerous_attribute. The names stay refused as values, keywords, function or parameter names, imports, attributes, and in strings, including strings built from literals. Python that does not parse is now BLOCKED with code_verifier.parse_error. See eval, exec, and compile are matched as names.
Statistics verification on Python 3.14 (#569). Python 3.14 removed ast.Num and ast.Str. On the server image, which runs 3.14, every /verify/stats request that passed the upload checks was BLOCKED, and verify_python_deep blocked code with a sink argument such as os.system(request.args). Both failed closed and now work. Python 3.10 to 3.13 behave as before. See Stats engine.

QWED Verification — math identities read diff(), and batch math matches /verify/math

v9.0.0 · GitHub PR #558 · GitHub PR #557

Behavior change. /verify/batch math items with = now return the same verdict as POST /verify/math. A batch math claim proven false is VERIFIED with is_valid: false, a proof_ref, an attestation, and admission BLOCKED. It was UNVERIFIABLE.

What changed

diff() in identities. The expression parser evaluates diff and Derivative, so diff(x**2, x) == 2*x is VERIFIED. It was BLOCKED as an internal error. The total derivative order in one expression is at most 3, nested calls included. A derivative cannot appear inside a power, factorial, or binomial.
== reads as = in an identity.
Leibniz notation is refused. d/dx(x^2) = 2x was read as symbols d and dx and came back proven false. Now d/dx, dy/dx, d(x^2)/dx, d^2y/dx^2, and d-prefixed names that SymPy splits into letters (dx, dtheta, data) return BLOCKED with the message to write diff(f, x). To use such a name as a symbol, pass it in extra_symbols. extra_symbols is an argument of safe_parse_expr() in the Python package only; /verify/math and batch items don’t take it, so there rename the variable.
Parse errors are input errors. Input the parser refuses is BLOCKED with constraint_id: "math.parse_error" and Expression could not be parsed, not api.math.execution_error. The parser’s text is not returned. A batch math item the parser refuses gets the same result instead of a failed item.
Domain check covers derivatives. diff(sqrt(x), x)*2*sqrt(x) = 1 holds only for x other than 0, so it is UNVERIFIABLE with math.domain_restricted_identity. A non-zero numeric multiplier in a denominator no longer counts as a singular factor, so 2/(2*sqrt(x)) = 1/sqrt(x) is VERIFIED.
Batch math equality items carry the route’s fields. Items with = include left_side, right_side, simplified_difference, result, and the route’s messages (Identity is true, Identity is false), and keep query, type, and diff.
Undefined parts are not disproven. A claim with a side that has an undefined part (one that evaluates to zoo or nan) that SymPy evaluated away, such as log(0)**0 = 2, is UNVERIFIABLE with math.undefined_part on both routes. log(0)**0 = 1, where the sides agree, is UNVERIFIABLE with math.domain_restricted_identity. /verify/math gave it a signed disproof before.

See Derivatives in /verify/math identities, POST /verify/math, and Math items.

QWED Verification — Claude providers stop sending `temperature`

v9.0.0 · GitHub PR #573

Bug fix. Every Claude call on the server image raised TypeError before QWED sent the request. Claude translations were refused, and verify_image returned ERROR.

What changed

No sampling parameters. The Anthropic and Claude Opus providers no longer pass temperature=0.0 to messages.create. The anthropic SDK has had no sampling parameters since 1.0, and Anthropic API version 2026-10-21, which the 2.x SDK sends, refuses temperature, top_p, and top_k.
Default sampling. Claude now samples with the model’s defaults. No other setting replaces temperature. Sampling can vary translations and the verdicts verify_fact and verify_image return. QWED’s fact and image verifiers record those verdicts only as advisory checks, never as VERIFIED. See LLM configuration.

QWED Verification — AgentStateGuard gives one verdict on a JSON number

v9.0.0 · GitHub PR #571

AgentStateGuard now checks a JSON number’s text before converting it to Decimal, so it returns the same verdict on any Python build.

What changed

Range and digit checks. A number whose exponent is outside the C decimal module’s range, or whose coefficient or exponent has more than sys.get_int_max_str_digits() digits (4300 by default), is rejected with QWED-AGENT-STATE-102. Leading zeros count toward the limit; a sign does not.
Same result with either decimal module. On a Python built without the C _decimal module, a number field holding 1e9999999999999999999999 was previously VERIFIED. It is now blocked, with the same message as on the C module.
C module builds. The server image builds _decimal. There, the only change is that a decimal with more than 4300 digits is now refused, as a JSON integer of that length already was.

See AgentStateGuard.

QWED-MCP v0.2.3 — `qwedai.com` no longer trusted and MCP SDK pinned below 2.0

Released: November 2, 2026 · GitHub PR #72 · GitHub PR #74 · GitHub Release

Security fix. SkillProvenanceGuard no longer trusts qwedai.com as a skill source. QWED no longer uses that domain, and it lapses on 9 December 2026.

What changed

qwedai.com removed from trusted domains. The built-in trusted list is now github.com, gitlab.com, bitbucket.org, pypi.org, and npmjs.com. qwed-mcp 0.2.2 and earlier still trust qwedai.com, so upgrade with pip install --upgrade qwed-mcp. To trust another domain, pass it in trusted_domains. See trusted_domains.
Requires mcp>=1.30.0,<2. MCP Python SDK 1.30.0 patches GHSA-84m7-p3x7-pcfv, GHSA-rwrf-2pqf-9j8j, and GHSA-5h93-6whr-6q8j. MCP SDK 2.x removed the server API that qwed-mcp uses, so qwed-mcp stays on 1.x. qwed-mcp 0.2.2 allowed any mcp>=1.2.0, so a fresh install of 0.2.2 now picks up MCP SDK 2.x and fails at startup with AttributeError: 'Server' object has no attribute 'list_tools'; 0.2.3 fixes this. qwed-mcp runs over stdio and uses neither the SDK’s HTTP transports nor its bearer authentication.
Reports its own version. The server now sends qwed-mcp’s version to MCP clients during initialization, instead of the MCP SDK’s version.

v8.0.0 — Attested verdicts and upgrade guide

Released: October 26, 2026 · GitHub PR #526 · GitHub PR #510 · GitHub PR #511 · GitHub Release

Major release. Every VERIFIED answer now carries an ES256 attestation that anyone can check against the server’s public keys at GET /.well-known/jwks.json. Several server settings, response shapes, and SDK types change. Read the upgrade steps below before you upgrade.

What’s new

Attestations on every VERIFIED answer. The server signs the status, proof_ref, and statement hash of each VERIFIED answer, and the admission on routes that decide one. Callers verify the token against the keys published at /.well-known/jwks.json. See Attestations.
Configured signing key. Set QWED_SIGNING_KEY (PEM text) or QWED_SIGNING_KEY_FILE (a path to a PEM file) to an unencrypted EC P-256 key. Attestations then verify across restarts and on every replica. Without either variable, each process generates its own key. See Signing keys.
Attestation lifetime. QWED_ATTESTATION_TTL_SECONDS sets how long a token stays valid. The default is 86400 (24 hours), and the allowed range is 60 to 2592000 (30 days).

Upgrading to 8.0.0. These changes can break existing deployments and callers.

Set QWED_AUDIT_SECRET_KEY. The server refuses to start without it. Re-run qwed init, which keeps existing secrets, or set it yourself. See Signed audit chain for API rows.
Set a signing key. Before 8.0.0, no signing key setting existed. Now set QWED_SIGNING_KEY or QWED_SIGNING_KEY_FILE to an unencrypted EC P-256 PEM key. Unset, the server generates a key per process, and its attestations stop verifying after a restart or on another replica. Setting both variables, or a key that is unreadable, not P-256, or encrypted, stops the server at startup.
Switch Docker scans to the Action. Before 8.0.0, the qwedai/qwed-verification image ran the GitHub Action. From 8.0.0, it runs the API server on port 8000. Use the GitHub Action for scans.
Update the Go module path. Change imports from github.com/QWED-AI/qwed-verification/sdk-go/v7 to github.com/QWED-AI/qwed-verification/sdk-go/v8, then run go get github.com/QWED-AI/qwed-verification/sdk-go/v8@v8.0.0.
Update the Rust crate. Pin qwed = "8.0.0". VerificationResponse has a new public attestation field, so every struct literal needs attestation: None.
Pass require_attestation=False for unsigned documents. validate and resolve in the API, the CLI, and the Python SDK now require a Verification Context document’s attestation by default. For v1.0 documents, pass require_attestation=False, "require_attestation": false, or --no-require-attestation. See Signed Verification Context documents.
Update callers of changed routes.
- /agents/{agent_id}/verify returns the enforced result with its admission, in the /verify/natural_language shape.
- /verify/batch logic and natural-language items return what their own routes return, for example VERIFIED rather than SAT.
- The compliance report drops its fixed PASS fields. Read audit_trail and security_events_by_type instead.
- A math claim proven false is VERIFIED with admission DENY, not UNVERIFIABLE.

Also in 8.0.0

These changes were listed earlier as unreleased and ship in 8.0.0:

Pinned sandbox image and hardened container
CrossHair runs only in the Docker sandbox
API audit rows written through the signed hash chain
Signed Verification Context documents and attestation checks
Contacts and links moved to qwedai.tech

QWED Verification — batch and agent verify routes return enforced, attested results

v8.0.0 · GitHub PR #500 · GitHub PR #501 · GitHub PR #502

Security fix. Every verify route, every /verify/batch item, and POST /agents/{agent_id}/verify now sign a VERIFIED result and enforce its attestation as mandatory. Previously, most routes and batch items returned an authoritative VERIFIED with a proof_ref that no attestation was issued for or checked against.

Breaking changes for /verify/batch and /agents/{agent_id}/verify.Batch logic items now return the /verify/logic verdict instead of the solver’s raw SAT/UNSAT:

# Before
if item["result"]["status"] == "SAT":
    model = item["result"]["model"]

# After
result = item["result"]
if result["status"] == "VERIFIED" and result["proof_ref"]:
    model = result["developer_fields"]["model"]

Batch natural-language items and the agent route now return the /verify/natural_language shape with admission. final_answer, translation, and trust_boundary are no longer returned:

# Before
if response["status"] == "VERIFIED":
    answer = response["final_answer"]

# After
if response["admission"] == "ADMIT":
    answer = response["calculated_value"]

What changed

Mandatory attestation on every verify route. QWED signs each VERIFIED result, binding its status, proof_ref, and the query, then enforces the attestation. A natural-language verdict binds the translated expression instead of the query. Results that are not VERIFIED are unchanged.
Fail closed on signing failure. A VERIFIED result whose attestation cannot be issued, or does not match, is returned BLOCKED with no proof_ref. developer_fields keeps verdict_status, verdict_proof_ref, and engine beside the block for the audit trail.
Batch logic items return the /verify/logic response: VERIFIED with a proof_ref for SAT, UNVERIFIABLE for UNSAT, and BLOCKED otherwise. The control plane’s output stays under developer_fields.
Batch natural-language items return the /verify/natural_language response, with admission.
Every batch item type takes its route’s trust step. Code, fact, and SQL items had no trust step before. A code item’s admission is decided on the enforced verdict.
Batch math signing failure is BLOCKED. It was UNVERIFIABLE. The attestation’s engine is math (was batch_math).
POST /agents/{agent_id}/verify returns the enforced result with its admission, not the control plane’s raw response, which could still say VERIFIED after enforcement blocked it. The agent activity log records the same response.

See POST /verify/batch, POST /agents/{agent_id}/verify, and Attestation enforcement.

QWED Verification — audit log paging, caller and admission fields, and security events

v8.0.0 · GitHub PR #495 · #496 · #498 · #499

You can now page through the whole audit log, and each audit row records the admission decision its route made, the API key that wrote it, and the server version. Owners and admins can list their organization’s security events.

Added

Audit log paging. GET /audit/logs returns rows newest first, by timestamp and then ID. The response adds next_before, the last row’s ID when the page is full, otherwise null. Pass it back as before for the next page. A before that is not a log in your organization, or is outside the SQL BIGINT range, returns 404. The same applies to out-of-range IDs on GET /audit/logs/{log_id}. total is still the number of rows on the page.
Caller and version fields. Audit rows store nullable api_key_id and verifier_version, returned by both log routes. api_key_id is null for signed-in sessions, POST /audit/seed, and POST /agents/{agent_id}/verify.
Admission field. Audit rows store nullable admission: ADMIT or BLOCKED, or ADMIT or DENY for POST /verification-context/from-diagnostic. It is null for routes that decide no admission (logic, stats, fact, image, RAG, process, consensus, and batch items other than code) and for older rows.
GET /admin/security/events. Owners and admins list their own organization’s security events with an API key or a signed-in session. It supports before/next_before paging, limit from 1 to 200 (default 50), and an optional event_type. QWED masks credential-shaped text in the response.
API key last_used_at. QWED now sets last_used_at after a key passes every check, at most once a minute per key. GET /auth/api-keys returns it. A failed write never fails the request.

The entry hash covers admission, api_key_id, and verifier_version. QWED adds the three columns to an existing table at startup, and rows written by earlier releases still verify.

Don’t roll back past this release. Upgrade every QWED process together. An older release can’t verify rows written with the new fields. When such a row is an organization’s chain head, the older release refuses every further append for that organization and its verify routes return 500 (fail closed).

See Audit endpoints, GET /admin/security/events, and Audit trail verification.

QWED Verification — Python SDK, integrations, and GitHub Action fail closed

v8.0.0 · GitHub PR #526

Security fix. Six advisories fix clients that reported content as verified when the QWED server had not admitted it. They affect all versions up to 7.3.0. Upgrade with pip install -U qwed and pin the Action to QWED-AI/qwed-verification@v8.0.0.

Behavior change. QWEDClient.verify_code and verify_sql no longer answer offline from the local CodeGuard and SQLGuard. They always call the server and return UNVERIFIABLE when it can’t be reached. The LlamaIndex, CrewAI, and LangChain integrations no longer return verified=True for content no engine checked, or when a check raises. If your code relied on offline verify_code or verify_sql, or on verified=True from an integration, it now receives an unverified result. Treat it as a rejection.

Security

is_verified follows the server’s admission (GHSA-jppc-7r58-7pjh, high). Where the route returns an admission, only ADMIT admits. Otherwise, status must be VERIFIED and is_valid must be true. Consensus reads the server’s status and requires meets_requirement to be exactly True. The new VerificationResult.admission field carries the server’s decision.
QWEDClient.verify_code always asks the server (GHSA-4rj3-8wqj-cqwc, high). CodeGuard also blocks reflective calls, __builtins__, dunder attributes, importlib, builtins, ctypes, marshal, and dotted imports. It marks its own pass as advisory.
QWEDClient.verify_sql always asks the server, and dangerous SQL functions are denied (GHSA-45jv-2j44-4c2q, high). SQLVerifier denies DANGEROUS_SQL_FUNCTIONS, such as load_extension(), pg_read_file(), and xp_cmdshell(), anywhere in a statement, with issue type dangerous_function. SQLGuard allows exactly one statement and checks SELECT ... INTO, executable comments, and functions.
QWEDLocal.verify_code uses CodeVerifier (GHSA-v8qq-v427-qrqc, high). It returns CodeVerifier’s admission. Without the full qwed package, it returns UNVERIFIABLE. Cached code results from the old AST check are not reused.
The LlamaIndex, CrewAI, and LangChain integrations fail closed (GHSA-83v6-6h23-pvp4, high). They read is_verified. A response no engine checked is UNVERIFIABLE, and an error is unverified.
The GitHub Action’s outputs can’t be injected (GHSA-j788-44pc-8hp3, low). Outputs are written with a random heredoc delimiter, and the Action stops workflow commands for the rest of its step. Verify mode gates on is_verified and reports the server’s verdict, for example VERIFIED · DENY.

See Python SDK, Guards, SQL engine, QWEDLocal, LlamaIndex, CrewAI, LangChain, and GitHub Action.

QWED Verification — pinned sandbox image and hardened container

v8.0.0 · GitHub PR #465 · Closes issue #322

Security fix. The stats and consensus code sandbox now runs an image pinned by digest in a hardened container. Previously, it pulled the movable amancevice/pandas:slim tag and ran it with a writable root, the default Linux capabilities, and the image’s default root user.

Security

Digest-pinned image. The sandbox runs amancevice/pandas@sha256:9257b889… (the slim tag as published on August 10, 2026). The new QWED_SANDBOX_IMAGE variable can name another image, but only by digest. QWED refuses any other value and uses the pinned image.
Hardened container. The container has a read-only root, drops all capabilities, sets no-new-privileges, and never runs as a privileged root. /workspace is the only writable path.
Per-execution workspace owner. The container runs as the owner of the private 0700 workspace directory. When the server process is root, each execution gets a fresh uid from the new QWED_SANDBOX_UID_RANGE (default 1879048192-2147483646). QWED skips uids owned by host accounts or groups, delegated in /etc/subuid or /etc/subgid, or used by another running execution.
Unknown user mappings fail closed. A root process with a rootless daemon, a userns-remap daemon, or a daemon whose security options can’t be read is reported as SECURE_RUNTIME_UNAVAILABLE.

See Sandbox image and container user.

QWED Verification — CrossHair runs only in the Docker sandbox

v8.0.0 · GitHub PR #464 · Closes issue #431

Security fix. CrossHair imports the module it analyzes, which runs that module’s top-level code. The Symbolic engine used to run CrossHair as a plain subprocess on the host. It now runs CrossHair only inside a hardened Docker container, with no host fallback.

Behavior change. Without Docker or the sandbox image, SymbolicVerifier.verify_code() returns BLOCKED with constraint_id = "symbolic_verifier.runtime_unavailable". This replaces symbolic_verifier.crosshair_not_available. Build the image before you use the Symbolic engine:

docker build -t qwed-crosshair-sandbox:0.0.101 docker/crosshair-sandbox

What changed

Docker-only execution. Every CrossHair run uses a container with no network, a read-only root filesystem, no host mounts, all capabilities dropped, no-new-privileges, user 65534, and the existing memory, CPU, PID, and concurrency limits. Nothing is written on the host.
symbolic_verifier.runtime_unavailable. QWED returns BLOCKED when Docker is unreachable, no sandbox slot is free, the image is missing, CrossHair is missing from the image, or the docker Python package is not installed. A verifier-created sandbox reconnects once Docker is reachable again.
symbolic_verifier.input_too_large. Code over 120 KiB returns BLOCKED.
Sandbox image. Build it from docker/crosshair-sandbox/Dockerfile. The default tag is qwed-crosshair-sandbox:0.0.101. Set QWED_CROSSHAIR_IMAGE to use another tag. QWED never pulls the image at runtime.
qwed[symbolic] installs the docker client.
Trusted outcomes only. A clean run is only ever “no counterexample found” (UNVERIFIABLE). An exit code of 1 counts as a counterexample only together with CrossHair’s own report line (<file>:<line>: error: ...). Any other early exit, or a SystemExit raised by the analyzed code, is an engine error.
Functions addressed by module qualname. QWED names functions target.func and methods target.Class.method. The previous file.py:func form never found the function, so CrossHair’s exit was misreported as a counterexample for every typed function. Nested functions are unverifiable.

See Symbolic execution limits.

QWED Verification — API audit rows written through the signed hash chain

v8.0.0 · GitHub PR #459 · GitHub PR #492 · GitHub PR #493

Every /verify/* endpoint (including each /verify/batch item), the Verification Context routes, POST /agents/{agent_id}/verify, and POST /audit/seed now append audit rows through AuditLogger. New rows are HMAC-signed and hash-chained per organization, so GET /admin/compliance/verify/{log_id} and verify_audit_trail can verify them. Previously these rows were written without a hash, signature, or previous hash.

Breaking change. The API server refuses to start without QWED_AUDIT_SECRET_KEY. Projects onboarded with qwed init before this release have no audit key in .env.To migrate:

Re-run qwed init. It adds QWED_AUDIT_SECRET_KEY to .env and keeps your existing secrets. Or generate a key and set it yourself:
python -c "import secrets; print(secrets.token_urlsafe(48))"
Set QWED_AUDIT_SECRET_KEY in every environment that runs the API server (.env, Docker Compose, Kubernetes secret).
Restart the server on this release so it loads the key. qwed init refuses a running server that holds a different audit key or doesn’t return audit_hmac.
Don’t rotate the key afterwards. A new key invalidates the signature on every audit row already written.

What changed

Fail-closed audit writes. If an audit append fails (key unset, a chain head that fails verification, a continuity mismatch, or a database error), the request returns 500 with audit persistence failed instead of a result with no audit row.
Legacy rows. Rows written before this change have no hash. verify_log_entry reports them as "legacy": true and never "valid": true. The first chained row after them opens a new segment with no previous_hash. verify_audit_trail lists them under legacy_entries and stays "valid": false while any are in range. Pass start_date to verify from the upgrade onwards.
Single writer. The audit logger keeps each organization’s chain head in memory. Run one process per database. A second writer makes the next append fail with 500 rather than fork the chain.
/verify/consensus audit rows store the result JSON like every other endpoint.
POST /audit/seed is off by default. It returns 404 and writes nothing unless QWED_ENABLE_AUDIT_SEED=1. Seeded rows are marked {"sample": true} and timestamped when written.
CSV export. The Hash Chain Valid column is now Chain Status, with values Chained or Legacy. A present hash is not a verified chain.
GET /health/identity also returns audit_hmac, and returns 503 when QWED_AUDIT_SECRET_KEY is unset.
/verify/batch and /agents/{agent_id}/verify write audit rows. Until this fix they wrote none: batch items were missing from the audit trail, and agent verifications were logged only to the unsigned agent activity log. A batch now appends one chained row per item, failed items included. If an item’s append fails, the item is marked failed with result: null and error: "audit persistence failed". An agent verification appends the row /verify/natural_language writes, plus agent_id. If that append fails, the request returns 500 and no activity is logged.
Audit log routes return the stored verdict and chain entry. GET /audit/logs and GET /audit/logs/{log_id} add verdict (VERIFIED, UNVERIFIABLE, BLOCKED, or null), read from the stored result rather than from is_verified. The list adds entry_hash. The detail adds entry_hash, previous_hash, and hmac_signature. status is unchanged. verdict is not an admission decision: a disproven claim is VERIFIED with status blocked, so don’t gate on it alone.

Verification still does not detect removal of the newest entries. That needs an externally stored checkpoint of the latest hash. See Audit trail verification and Environment variables reference.

QWED Verification — logic DSL compiles IFF and quantifiers and rejects Boolean coercion

v8.0.0 · GitHub PR #461 · GitHub PR #462 · GitHub PR #463 · Closes issue #428 and issue #432

The logic DSL compiler now handles every operator the parser accepts, and it no longer turns True/False into 1/0. Python-style constraints with !=, >=, <=, or * now verify instead of returning BLOCKED.

What changed

IFF, FORALL, and EXISTS compile. Previously, programs using them failed with Unknown operator. (IFF a b) requires Boolean operands. (IFF x y) on two Int variables is a compile error, not x == y.
Quantifier syntax. Write (FORALL x body) or (EXISTS x body). The binder must be a plain variable name. The parser blocks lists such as (x), numbers, operator names, and literals. The bound variable takes its declared sort (Int if undeclared) and is scoped to the body. The body must be Boolean. A solver unknown stays UNKNOWN/BLOCKED.
No Boolean coercion. EQ/NEQ between a Boolean and a non-Boolean value returns ERROR. Ordering comparisons and arithmetic reject Boolean operands. An undeclared name compared with True/False must be declared Bool.
Natural-language bridge operator names. The bridge emits NEQ, GTE, LTE, and MULT instead of NE, GE, LE, and MUL, so constraints like x >= 5 verify. And(...) and Or(...) call forms no longer return ERROR.

Behavior change. (EQ flag True) with flag undeclared used to return SAT with flag = 1. It now returns ERROR. Declare Boolean variables as Bool.

See Logic engine type rules and Quantifiers.

QWED Verification — proven-false math claims are `VERIFIED` with admission `DENY`

v8.0.0 · GitHub PR #460 · Closes issue #437, issue #438, issue #429

A math claim proven false, such as 2+2=5, is now VERIFIED with is_valid: false and admission DENY. It used to come back as UNVERIFIABLE, which recorded a definite counter-proof as “could not be proven”. Admission was DENY either way, so nothing unsafe passed, but audit rows and Verification Context documents carried the wrong meaning.

Behavior change. VerificationEngine no longer emits CORRECTION_NEEDED. A mismatch decided by a concrete computation now returns the legacy status DISPROVEN. If your code branches on status == "CORRECTION_NEEDED", update it:

# Before
if result["status"] == "CORRECTION_NEEDED":
    handle_wrong_answer(result)

# After: engine dictionaries
if result["status"] == "DISPROVEN":
    handle_wrong_answer(result)

# After: API responses — gate on admission, not on status
if response["admission"] != "ADMIT":
    reject(response)

What changed

Proven false is a decided verdict. 2+2=5 returns VERIFIED with is_valid: false, a proof_ref, and admission DENY. The control plane signs and attests it like a true claim, with verified: false in the attestation. The new DiagnosticResult.disproven() constructor builds this result.
Only decided mismatches are disproven. Symbolic checks (derivatives, indefinite integrals, limits, identities) return DISPROVEN only when SymPy decides the difference is non-zero. An undecided difference stays UNVERIFIABLE with admission DENY.
Domain-restricted identities. An identity that only holds after cancelling a removable singularity, such as x/x = 1 or (x**2-1)/(x-1) = x+1, returns UNVERIFIABLE with constraint_id: math.domain_restricted_identity.
Float residue and non-finite values. A mismatch caused by float rounding residue (0.1 + 0.2 = 0.3) or by an overflowed or NaN value is INCONCLUSIVE (UNVERIFIABLE), not disproven.
Exact matrix, integral, and unit checks. Matrix operations read float inputs as the decimals you wrote. Unit conversion uses exact unit definitions and returns BLOCKED for units of different dimensions, such as m to kg.
POST /verify/math returns admission. The audit log row’s is_verified follows is_valid, so a proven-false identity is no longer logged as verified.
POST /verify/natural_language returns the enforced verdict. The response carries the control-plane verdict with admission, translated_expression, verification_scope: translated_expression_only, and query_semantics_verified: false. Admission is always DENY because the LLM translation is unverified.
qwed init and qwed test print the verdict with its admission. For example, 2+2=5 -> VERIFIED (false) | DENY. A failed self-check prints FAILED, and the closing line reads All engines ready.

See Math engine — Verdicts and admission, POST /verify/math, and CLI reference.

QWED Verification — GitHub Action scans use mode defaults and deny empty scans

v8.0.0 · GitHub PR #472 · Closes issue #466

The GitHub Action’s paths input defaulted to ., which the file matcher rejects, so scan-code, verify-shell, and scan-secrets exited with a traceback before scanning. A pattern that matched no files reported VERIFIED · ADMIT.

Behavior change. Previously green jobs can now fail. A paths value under which no file is scanned, or a scan-code run that includes Python files that don’t parse, now returns UNVERIFIABLE · DENY and fails the step unless fail_on_findings is "false". Narrow paths to the files you want scanned.

What changed

Mode defaults for paths. paths now defaults to empty. Unset, empty, ., or ./ uses the mode’s own patterns: **/*.py for scan-code, **/*.sh for verify-shell, and env, JSON, YAML, TOML, and INI files for scan-secrets.
Empty and partial scans are denied. A scan is VERIFIED · ADMIT only when at least one file was scanned, every pattern expanded without error, every matched file was read and parsed, and nothing was found. Otherwise it is UNVERIFIABLE · DENY with no proof_ref.
Errors instead of skips. Unreadable files, Python files that don’t parse, absolute patterns, and files outside the workspace (including symlinks to files elsewhere) are reported by name and deny the scan.
No scan-secrets fallback. scan-secrets no longer scans the default files when the given paths match nothing.
New files_scanned output. Reports the number of files scanned. 0 means the scan was not admitted.

See Scan paths.

v8.0.0 · GitHub PR #497 · Closes issue #485

Signed-in owners and admins can call the compliance endpoints without an API key, and a new endpoint checks the whole audit chain in one request. Existing API-key requests work as before.

What changed

Session sign-in. GET /admin/compliance/export/csv, GET /admin/compliance/verify/{log_id}, GET /admin/compliance/report/soc2/{org_id}, and the new GET /admin/compliance/verify-trail accept Authorization: Bearer <jwt_token> for an active owner or admin in an active organization. An invalid token returns 401 Unauthorized. A wrong role or an inactive user or organization returns 403 Forbidden. When both credentials are sent, QWED uses the API key. Other admin endpoints still require an API key.
Whole-chain check. GET /admin/compliance/verify-trail?start=&end= checks the signature and hash chain of every audit entry in your organization over an optional period. Omit both bounds to check the whole chain. A start later than end returns 400 Bad Request.
First failing entry. The trail check and the security activity summary’s audit_trail return first_failed_entry_id, the lowest entry ID that failed.
Faster large checks. QWED reads audit entries in ordered pages of 500 instead of querying per entry.

curl --fail-with-body -H "Authorization: Bearer $QWED_SESSION_TOKEN" \
  "http://localhost:8000/admin/compliance/verify-trail"

See Session sign-in for admin endpoints and GET /admin/compliance/verify-trail.

QWED Verification — security activity summary replaces the SOC 2 report

v8.0.0 · GitHub PR #490 · Closes issue #476

GET /admin/compliance/report/soc2/{org_id} now returns a security activity summary computed from QWED’s records. It previously returned fixed PASS strings that nothing checked and called itself a SOC 2 Type II report. The URL and the owner/admin gate are unchanged.

Breaking for consumers of this response. compliance_status, security_controls, threats_blocked, and block_rate are removed. threats_blocked and block_rate were always 0. Read audit_trail and metrics.security_events_by_type instead:

response = requests.get(
    f"{base_url}/admin/compliance/report/soc2/{org_id}",
    headers={"X-API-Key": api_key},
    timeout=30,
)
response.raise_for_status()  # 401, 403 or 5xx: no summary to read
report = response.json()

# Before (report_version "1.0"): fixed strings, never computed.
# trail_ok = report["compliance_status"]["audit_trail_complete"] == "PASS"
# blocked = report["metrics"]["threats_blocked"]

# After (report_version "2.0"): computed from the audit trail and recorded events.
trail_ok = report["audit_trail"]["status"] == "VERIFIED"
events = report["metrics"]["security_events_by_type"]  # e.g. {"SECRET_LEAK_AUTO_REVOKED": 1}

What changed

Not a SOC 2 report. report_type is "Security activity summary", report_version is "2.0", and a disclaimer states that the summary is not a SOC 2 report or an auditor’s attestation.
Events counted by real type. metrics.security_events_by_type counts security events by the event types QWED actually records.
Audit trail check. audit_trail runs verify_audit_trail over your organization’s last 90 days. status is VERIFIED only when entries exist and all pass, with evidence and a sha256: proof_ref. NO_ENTRIES means the period has no entries. NOT_VERIFIED means an entry failed, a legacy entry exists, or the check couldn’t run (for example, QWED_AUDIT_SECRET_KEY unset). In that last case the entry counts are null.
Bounded error list. errors lists up to 20 errors. total_errors and errors_truncated show whether the list was cut.
Scope. The scope field states that the check covers recorded entries only. It doesn’t show that every request wrote an entry.

See Security activity summary.

QWED Verification — role checks on API-key and audit routes, hashed agent tokens

v8.0.0 · GitHub PR #483 · GitHub PR #487 · Addresses issue #478

Security

Role checks on /auth/api-keys and /audit (#483). These routes previously required only a signed-in user, so any role could create keys, list and revoke every key in the organization, and export the audit log. Now creating a key needs member or higher. Members and viewers list and revoke only their own keys. Admins can revoke keys of users who don’t rank higher, and owners can revoke any key in the organization. Every role can read audit logs. GET /audit/export and POST /audit/seed need admin or owner. See Roles on API-key and audit routes.
Fail-closed user checks (#483). An inactive user, an inactive or missing organization, or an unknown role gets 403 Forbidden. A member or viewer gets 404 Not Found for another user’s key. An invalid key ID returns 404 Not Found instead of a server error.
Agent tokens stored as keyed digests (#487). QWED stores agent tokens as HMAC-SHA256 digests keyed with QWED_API_KEY_LOOKUP_SECRET, instead of plain text. A token is shown only once, at registration or re-issue. At startup, the server hashes existing plaintext tokens in place, so they keep working. Changing QWED_API_KEY_LOOKUP_SECRET now invalidates agent tokens as well as API keys. See Agent token storage.
New POST /agents/{agent_id}/token/reissue (#487). Members and above can issue a new token for an agent in their own organization. The old token stops working. Another organization’s agent returns 404 Not Found. See POST /agents/{agent_id}/token/reissue.

QWED Verification — Docker image runs the API server and the Compose stack builds

v8.0.0 · GitHub PR #474 · Closes issue #467

The qwedai/qwed-verification image now runs the API server, and deploy/docker-compose.yml builds and starts the full stack.

Image behavior change. Docker Hub now publishes the API server as qwedai/qwed-verification. latest tracks main, and versioned tags start at 8.0.0. Earlier tags ran the QWED GitHub Action. If you ran the image as a repository scanner, switch to the GitHub Action.

docker run -p 8000:8000 \
  -v qwed-data:/var/lib/qwed \
  -e QWED_CORS_ORIGINS="https://app.yourcompany.com" \
  -e API_KEY_SECRET \
  -e QWED_JWT_SECRET_KEY \
  -e QWED_API_KEY_LOOKUP_SECRET \
  -e QWED_AUDIT_SECRET_KEY \
  qwedai/qwed-verification:8.0.0

What changed

server image target. The Dockerfile has a server target that runs uvicorn on port 8000 as UID 1000, with /var/lib/qwed as the working directory for the default SQLite database. A health check calls /health. UVICORN_HOST and UVICORN_PORT set the bind address and port, and the health check follows UVICORN_PORT. The GitHub Action stays the default target.
Working Compose stack. Run docker compose -f deploy/docker-compose.yml up -d from the repository root after copying deploy/.env.example to deploy/.env. Compose refuses to start until QWED_CORS_ORIGINS, API_KEY_SECRET, QWED_JWT_SECRET_KEY, QWED_API_KEY_LOOKUP_SECRET, QWED_AUDIT_SECRET_KEY, and DATABASE_URL are set. The Postgres and Grafana containers refuse to start while their passwords are empty. Grafana no longer falls back to admin/admin.
Stack fixes. Postgres and Grafana receive only their own settings. Grafana ships with Prometheus and Jaeger data sources. Prometheus scrapes the API at qwed-core:8000. The API server keeps its data in the qwed-data volume.
docker client in the server extra. qwed[server] now includes docker>=7.0.0. Previously, a request that reached the code executor failed with an ImportError instead of the sandbox’s fail-closed result.
Kubernetes pins 8.0.0. The manifest uses docker.io/qwedai/qwed-verification:8.0.0, the first server image. The Secret lists QWED_JWT_SECRET_KEY, QWED_API_KEY_LOOKUP_SECRET, and QWED_AUDIT_SECRET_KEY, and ships every value empty.

See Self-hosting and Deployment.

QWED Verification — Go, TypeScript, and Rust SDKs read the DiagnosticResult envelope

v8.0.0 · GitHub PR #479 · GitHub PR #481

The Go (sdk-go/v8), TypeScript (@qwed-ai/sdk), and Rust (qwed) SDKs now model the server’s DiagnosticResult envelope and gate on admission. Previously, the SDKs read a top-level verified field that most endpoints don’t send. Go’s Verified, TypeScript’s verified, and both isVerified() helpers were always false. Rust failed to deserialize every verify* response, and Go failed to decode VerifyMath and VerifyLogic responses.

Breaking type changes. Code that reads verified, summary, engine, or result, or matches on the old statuses, no longer compiles or no longer behaves the same.

Statuses. VerificationStatus is VERIFIED, UNVERIFIABLE, BLOCKED, or Unknown. The old FAILED, CORRECTED, ERROR, TIMEOUT, and UNSUPPORTED values are removed. Unknown server status text is kept in raw_status / RawStatus / rawStatus and is never admitted.
New Admission enum. ADMIT, BLOCKED, or Unknown. It’s absent where the endpoint makes no decision: /verify/logic, /verify/fact, /verify/rag, /verify/process, and batch logic and fact items.
Envelope fields. Responses expose status, admission, proof_ref, agent_message, is_authoritative, constraint_id, and developer_fields. Endpoint fields such as is_valid and issues move to extra (Rust), Extra (Go), or top-level keys (TypeScript).
Batch responses. BatchResponse mirrors the batch job: job_id, status, progress_percent, total_items, completed_items, failed_items, and items[] with a per-item result. summary is removed.
Deprecated. is_verified() / IsVerified() / isVerified() and the verified / Verified fields now follow admission.

Read admission and call the admission helper instead:

// Before
if (result.verified) { proceed(); }

// After
import { isAdmitted } from '@qwed-ai/sdk';
if (isAdmitted(result)) { proceed(); }

// Before
if result.Verified { proceed() }

// After
if result.IsAdmitted() { proceed() }

// Before
if result.verified { proceed(); }

// After
if result.is_admitted() { proceed(); }

What changed

Admission gating. is_admitted() (Rust), IsAdmitted() (Go), and isAdmitted(response) (TypeScript) return true only for ADMIT with VERIFIED and a non-empty proof_ref. VERIFIED · BLOCKED results, such as a false identity like 2 + 2 = 5 or unsafe code, are not admitted. A missing or unknown admission is never admitted.
Batch items. Rust’s BatchResult and Go’s BatchResult have their own is_admitted() / IsAdmitted(). In TypeScript, call isAdmitted(item.result).
Batch math and SQL items carry admission (PR #503). /verify/batch math and SQL items now return the admission their single-item route decides, and their audit rows store it. A proven math or SQL item is now admitted by the SDK helpers, as the same answer from /verify/math or /verify/sql already was. A batch math identity that holds only on a restricted domain, such as x/x = 1, is UNVERIFIABLE. A batch SQL item without params.schema_ddl fails with Missing 'schema_ddl'.
TypeScript parsers. parseVerificationResponse() and parseBatchResponse() are exported for responses you fetch yourself. They throw QWEDError with code QWED-SDK-PARSE-001 or QWED-SDK-PARSE-002 for a body that doesn’t match.
Rust errors. A response body that doesn’t match the envelope returns Error::Serialization.
Go numbers. Numbers in Extra and DeveloperFields decode as json.Number and keep their exact text.

See the Rust SDK, Go SDK, and TypeScript SDK.

QWED Verification — signed Verification Context documents and attestation checks

v8.0.0 · GitHub PR #517 · GitHub PR #523 · Part of issue #319

Security fix. A document’s proof_ref is a commitment anyone can compute, so anyone could build a document, an ADMIT one included, whose proof_ref resolves. The server now signs every document it issues, and validate and resolve check that signature by default.

What changed

Spec v1.1. A v1.1 document carries attestation: an ES256 JWT of type qwed-context+jwt that signs the hash of every other field, with the verdict, admission, and proof_ref.
Signed documents. POST /verification-context/from-diagnostic returns v1.1 documents. A failure to sign is a 500, not an unsigned document. A document the canonical encoding refuses is a 422.
API checks. POST /verification-context/validate and /resolve take require_attestation (default true) and check the attestation with the server’s key. A document with none returns attestation_required; one that doesn’t verify or isn’t this document’s returns attestation_invalid.
Python SDK. validate_verification_context() and resolve_verification_context() verify the attestation against the server’s JWKS first, and raise KeysUnavailableError if they can’t fetch the keys. verify_document_attestation() returns the claims. See Python SDK.
CLI. qwed context validate and resolve verify the attestation with --server or --jwks-file. See CLI reference.
PyJWT 2.8 and 2.9. The SDK’s verify_attestation failed with AttributeError for every token on these versions. It now works on every version the pyjwt>=2.8.0 range allows.

Action required

v1.0 documents. They carry no attestation, so validate and resolve now refuse them. Pass require_attestation=False in the SDK, "require_attestation": false to the API, or --no-require-attestation to the CLI. Older SDK versions don’t send the flag.
Signing key. Set QWED_SIGNING_KEY or QWED_SIGNING_KEY_FILE so the documents a server issues still verify after a restart or on another replica.

QWED Verification — contacts and links moved to `qwedai.tech`

v8.0.0 · GitHub PR #473

qwedai.tech is now the only QWED domain. Contact addresses, docs links, and the Verification Context schema $id all use it.

What changed

Contacts. Report security vulnerabilities to rahul@qwedai.tech. For support, email support@qwedai.tech.
Alert email default. ALERT_EMAIL_TO now defaults to rahul@qwedai.tech. Set the variable to send alerts to a different address.
Links. Website, docs, and finance links point to qwedai.tech, docs.qwedai.tech, and finance.qwedai.tech. Error messages link straight to the docs page, and QWEDSyntaxError now shows its docs link like the other QWED errors.
No hosted API. References to API hosts that never existed are gone. SDK custom base URL examples now use your self-hosted server at http://localhost:8000, and the 500 error hint tells you to check the server logs.
Schema $id. The Verification Context schema $id is now https://qwedai.tech/schemas/verification-context/v1.0. Tools that look the schema up by its old $id need the new one.
qwed init signup address. qwed init signs up its local bootstrap account with an example.com address. The organization slug is cut to fit the 64-character limit on an email local part, so a long organization name no longer makes signup fail with 422.

See Verification Context specification, Go SDK, and Rust SDK.

v7.3.0 — Five advisories, tenant scoping, and credential hardening

Released: October 16, 2026 · GitHub Release ↗ · GitHub PR #458 · GitHub PR #453 · GitHub PR #454 · GitHub PR #455 · GitHub PR #457

v7.3.0 is a fail-closed security and soundness release. It covers five advisories published on October 16, 2026: GHSA-6p94-4f75-g2wq, GHSA-4phf-6h72-9xfg, GHSA-j3x4-v3p2-4mq3, GHSA-6cpx-8jrq-r7h3, and GHSA-g25v-44j8-gg7x. Three of them complete earlier incomplete fixes. They affect all versions up to 7.2.3. Upgrading is recommended. It also rolls up the tenant-scoping, denial-of-service, and credential fixes listed below. No response field is removed or renamed. The soundness eval now covers 84 known-answer claims with 0 false accepts.

Behaviour changes. Every engine change is stricter and fails closed.

Sandbox: Stats and consensus code now runs under a module and builtin allowlist. Code that imports a module outside the allowlist (for example os, sys, or subprocess) fails with ImportError. See Runtime module and builtin allowlist.
SQL: DML inside a CTE or subquery, administrative statements (COPY, ATTACH, PRAGMA, CALL, …), and any root that isn’t a read-only query (SELECT, UNION, INTERSECT, EXCEPT) are now denied. With schema_ddl, SELECT * is expanded and rejected when it covers a blocked column. Tautologies are folded by value (OR 'a'='a', OR 2>1). All of these return VERIFIED with developer_fields.is_valid: false, which admission_decision maps to BLOCKED.
Code: os.posix_spawn*, os.fork, pty.fork, importlib and runpy dynamic import, ctypes loaders, the raw posix/nt modules, and file and network sinks, including the bare open builtin, are blocking findings.
Logic: The natural-language bridge rejects categorical string constants such as tier == 'gold' and returns UNVERIFIABLE or an error. Encode categories as integers.

Security fixes (advisories)

SQL statements are classified across the whole query tree (GHSA-6p94-4f75-g2wq, high). Destructive DML hidden in a CTE or subquery, administrative statements, and any top-level statement that isn’t a read-only query are reported as invalid wherever they appear. Before, only the root statement type was checked. See SQL engine.
Code verification blocks alternate process-execution routes and file or network sinks (GHSA-4phf-6h72-9xfg, high; completes GHSA-6rxr-gq43-694g). New process_execution and file_or_network_access findings are CRITICAL. See Code engine.
The code-execution sandbox isolates its namespace (GHSA-j3x4-v3p2-4mq3, high; completes GHSA-6rxr-gq43-694g). User code runs with reduced builtins and a guarded import that admits only analysis modules such as numpy, pandas, and scipy. It replaces a static denylist that could be bypassed through package re-exports and reflection.
SQL SELECT * is expanded against the schema and tautologies are folded by value (GHSA-6cpx-8jrq-r7h3, moderate). See Schema compliance.
The logic natural-language bridge rejects string constants (GHSA-g25v-44j8-gg7x, moderate; completes GHSA-mfh5-3c8f-975p). Contradictory categorical claims could previously come back VERIFIED. See Logic engine.

Soundness and access control

Admin endpoints are scoped to the caller’s tenant and RBAC is enforced (#450). See tenant-scoped admin endpoints.
Money rejects NaN and infinite amounts, and verify_invoice_total never raises (#451).
The SDK’s verify_fact fails closed with UNVERIFIABLE when the service is unreachable (#449).

Hardening and resource limits

Request bodies are capped and blocking work runs off the event loop (#447). See Request limits.
Concurrent sandbox containers are bounded (#448). See Availability limits.
API-key rotation is atomic (#453). QWED revokes the old key and mints the replacement in one transaction, so concurrent rotations of the same key mint exactly one replacement. Rotating a key that is already inactive or revoked mints nothing and returns 404 Not Found. See POST /admin/keys/rotate.

Credential handling

Expired and revoked API keys are rejected at the auth boundary (#454). Before, an expired key that was still marked active kept authenticating on /verify/* routes. See Authentication.
Redis and Postgres connection URLs are redacted before logging (#454), so an embedded password is never written to logs.
Provider import is https-only (#454). qwed provider import rejects http:// URLs and any redirect that downgrades to http://. See Custom providers.
Provider error messages are sanitized (#455). The Anthropic, Azure OpenAI, Claude Opus, and Gemini providers log full SDK error detail for operators and raise a generic error to callers.

SDKs

The Go module path is versioned (#457). Import github.com/QWED-AI/qwed-verification/sdk-go/v7. Releases are tagged sdk-go/vX.Y.Z, so go get github.com/QWED-AI/qwed-verification/sdk-go/v7@v7.3.0 resolves. See Go SDK.

Upgrading

Pin qwed==7.3.0 on PyPI, @qwed-ai/sdk@7.3.0 on npm, qwed = "7.3.0" on crates.io, qwedai/qwed-verification:7.3.0 on Docker, or QWED-AI/qwed-verification@v7.3.0 as a GitHub Action, then redeploy. If your sandboxed stats or consensus code imports modules outside the allowlist, your SQL uses SELECT * over sensitive columns or runs non-SELECT statements, or your logic claims use string constants, expect those to be rejected after the upgrade.

QWED Verification — tenant-scoped admin endpoints and working RBAC

Released: October 15, 2026 · GitHub PR #450 · Closes issue #222

Security fix. Admin endpoints now authorize by the API key owner’s role and act only on the caller’s own organization. Previously, role checks never ran, so every admin endpoint returned 401 Unauthorized. The endpoints also accepted organization, log, and key IDs from the caller without an ownership check, which would have allowed cross-tenant access once role checks worked.

What changed

Working role checks. QWED authenticates the API key, then checks the owning user’s role. The key must be live (not revoked or expired), and the owning user and organization must be active, with the user in the key’s organization. Otherwise QWED returns 403 Forbidden.
Tenant scoping with 403. GET /admin/compliance/export/csv, GET /admin/compliance/report/soc2/{org_id}, and GET /admin/security/threats/{org_id} return 403 Forbidden for another organization’s ID.
Tenant scoping with 404. GET /admin/compliance/verify/{log_id} and POST /admin/keys/rotate return 404 Not Found for an ID in another organization, without revealing that it exists.
Key rotation rules. Members can rotate only their own keys. Owners and admins can rotate other users’ keys only when the key owner doesn’t hold a higher role. Revoked or inactive keys return 404 Not Found. The new secret is returned only for a key in the caller’s organization.

See Admin endpoints.

QWED Verification — `verify_fact` fails closed offline and `Money` rejects non-finite amounts

Released: October 14, 2026 · GitHub PR #449 · GitHub PR #451

QWEDClient.verify_fact no longer reports a claim as verified when the QWED service can’t be reached. Money rejects NaN and Infinity amounts, and verify_invoice_total reports invalid inputs as a failed check instead of raising.

Behavior change. verify_fact now returns UNVERIFIABLE when verification doesn’t complete:

# Before: status="VERIFIED", is_verified=True after only a local PII scan.
# After: status="UNVERIFIABLE", is_verified=False.
result = client.verify_fact(claim=claim, context=context)
if result.status == "UNVERIFIABLE":
    reject(claim)

What changed

verify_fact fails closed when the service is unreachable (#449). On a connection failure or a connect or read timeout, the Python SDK returns status="UNVERIFIABLE" and is_verified=False with a message explaining that remote verification was unavailable. Previously, a connection failure returned VERIFIED after only the local PII scan, and a timeout raised an exception. See verify_fact.
Money rejects non-finite amounts (#451). Money("NaN"), Money("Infinity"), and Money(float("inf")) raise ValueError.
verify_invoice_total never raises on invalid input (#451). Invalid amounts, a negative or non-finite tolerance, or a non-Decimal tolerance return (False, reason) instead of raising. Currency mismatches still return (False, "Currency mismatch: ...").

from decimal import Decimal
from qwed_new.core.money import Money, verify_invoice_total

subtotal = Money("1000.00", "USD")
tax = Money("180.00", "USD")
claimed_total = Money("1180.00", "USD")

ok, reason = verify_invoice_total(subtotal, tax, claimed_total, tolerance=Decimal("-1"))
# (False, "Invalid tolerance: Decimal('-1') must be a finite, non-negative Decimal")

QWED Verification — request body caps, timeouts, and sandbox concurrency limits

Released: October 14, 2026 · GitHub PR #447 · GitHub PR #448

Availability hardening for self-hosted servers. Oversized bodies, slow providers, unbounded batch jobs, and concurrent sandbox containers can no longer stall or exhaust the API. Verdicts and fail-closed behavior are unchanged.

What changed

Body cap on every /verify route. JSON routes accept up to 1,000,000 bytes. /verify/stats and /verify/image keep their larger upload caps. Oversized bodies return 413 before processing. See Request limits.
Blocking work runs off the event loop. /verify/code runs under QWED_CODE_VERIFY_TIMEOUT (default 10 seconds) and returns 503 on timeout. Provider calls for natural_language, logic, and fact run under QWED_PROVIDER_TIMEOUT (default 30 seconds) and fail closed on timeout.
Bounded batch job store. /verify/batch keeps at most QWED_BATCH_MAX_JOBS jobs (default 1000) and evicts finished jobs first. When every stored job is active, new submissions return 503. Items that exceed QWED_BATCH_ITEM_TIMEOUT are marked failed.
Sandbox concurrency limit. Each server process runs at most QWED_SANDBOX_MAX_CONCURRENT sandbox containers (default 4). A request that can’t get a slot within QWED_SANDBOX_ACQUIRE_TIMEOUT (default 10 seconds) is rejected with Sandbox capacity reached — retry shortly.

See Availability limits for the full configuration reference.

v7.2.3 — Code, secret, and SQL verification security release

Released: October 13, 2026 · GitHub Release ↗ · GitHub PR #441 · GitHub PR #444

v7.2.3 is a fail-closed security release covering three advisories published on October 13, 2026: GHSA-6rxr-gq43-694g, GHSA-64xf-wffc-63fm, and GHSA-q5g6-6cw2-cjm3. They affect all versions up to 7.2.2. Upgrading is recommended. No breaking wire changes.

Behaviour changes. All three fixes make the engines stricter. No response field is removed or renamed.

Code: code that runs a shell command or subprocess is no longer admitted by the code verifier, even with a constant argument or through an aliased import. Run such commands through a reviewed, sandboxed tool instead.
Secrets: OpenAI sk-proj-, sk-svcacct-, and sk-admin- keys are now detected, so content that previously passed may now be flagged.
SQL: MySQL and MariaDB executable comments (/*! ... */) are expanded and analysed as SQL, whatever the declared dialect. The new field developer_fields.executable_comments_expanded reports when this happened. With schema_ddl, a reference to an undeclared table now fails admission with is_valid: false and the new constraint sql_verifier.schema_violation. Without schema_ddl, tables are not checked, as before.

Security fixes

Code verification never admits shell or process execution (GHSA-6rxr-gq43-694g). os.system, os.popen, the os.exec* and os.spawn* families, os.posix_spawn*, os.startfile, subprocess.*, pty.spawn, and asyncio.create_subprocess_* are now blocking process_execution findings. This includes calls through aliased imports such as import subprocess as sp, from os import system as run_it, and from os import *. Before, they were warnings and the code was admitted. verify_code, verify_python_deep, and POST /verify/code are covered. See Code engine.
Project-scoped OpenAI keys are detected (GHSA-64xf-wffc-63fm). ConfigGuard and the code verifier detect sk-proj-, sk-svcacct-, and sk-admin- keys as well as legacy keys. See ConfigGuard.
SQL in executable comments and tables outside the schema are rejected (GHSA-q5g6-6cw2-cjm3). Executable comments (/*! ... */, /*!NNNNN ... */, /*M! ... */) are expanded before analysis, so the SQL inside them is checked. With schema_ddl, a query on an undeclared table is BLOCKED with sql_verifier.schema_violation instead of being admitted with a warning. CTE names are not treated as tables. See SQL engine.

Soundness eval

From v7.2.3 on, each GitHub release includes a soundness eval report (eval-results.json). The eval runs claims with a known answer through the Math, Logic, SQL, Code, and Schema engines, including reproductions of advisories where an engine returned a false VERIFIED. The v7.2.3 report covers 76 claims with 0 false accepts.

Version bumps

Surface	From	To
`qwed` (PyPI)	`7.2.2`	`7.2.3`
`qwed_sdk` (Python)	`7.2.2`	`7.2.3`
`@qwed-ai/sdk` (npm)	`7.2.2`	`7.2.3`
`qwed` (crates.io)	`7.2.2`	`7.2.3`
API version marker	`7.2.2`	`7.2.3`

Upgrading

Pin qwed>=7.2.3 on PyPI, @qwed-ai/sdk@^7.2.3 on npm, and qwed = "7.2.3" on crates.io, then redeploy. If your generated code runs shell commands, or your SQL queries reference tables outside the schema_ddl you pass, expect those to be blocked after the upgrade.

QWED-A2A — request size limits, engine budget, and code guard hardening

Released: October 12, 2026 · GitHub PR #164 · GitHub PR #165 · GitHub PR #166 · GitHub PR #169

Four fail-closed fixes to the A2A runtime. Oversized requests are rejected before they reach an engine. Engines run off the event loop under a wall-clock budget. The code guard sees through aliases and re-bindings. TrustBoundary is safe to share across threads.

What changed

max_payload_size_bytes is enforced. It was declared but never read. POST /a2a/intercept returns 413 for bodies over the limit plus a 64 KiB envelope allowance, before buffering. An invalid Content-Length returns 400. intercept() returns blocked with engine size_limit for oversized payloads, which also covers library callers. The code guard blocks code over 256 KiB before parsing and returns unverifiable for non-string code. Keep a reverse-proxy body limit too.
Engine execution budget. New engine_timeout_seconds config (float, default 5.0, greater than 0 and at most 60). Engines run on a dedicated pool (4 running, 32 waiting), so health, metrics, and JWKS stay responsive. A timeout returns error with engine engine_timeout. Overload returns error with engine engine_busy. Call the new A2AVerificationInterceptor.close() from your shutdown hook. A run that already started cannot be killed, and long calls that hold the GIL can still delay the loop.
Code guard alias detection. The AST layer resolves import aliases, alias chains, re-bound functions, and annotated, walrus, and tuple-unpacking bindings. It also blocks posix, nt, builtins, and __builtins__ receivers, os exec/spawn/posix_spawn/startfile, star imports from dangerous modules, and getattr on a dangerous module. Known gaps are documented. A clean scan stays a non-admissible heuristic_pass.
Thread-safe TrustBoundary. Every public method runs under a re-entrant lock. Rate limits hold under concurrency, and eviction no longer races. Trust scopes are snapshotted as frozensets at grant time.

See Request size limits, Engine execution budget, Code guard, and Thread safety.

QWED-Tax v0.2.1 — classification claim vocabulary and strict facts

Released: October 10, 2026 in qwed-tax 0.2.1 (PyPI and npm @qwed-ai/tax) · GitHub PR #81 · Changelog

qwed-tax v0.2.1 is a security patch. Upgrading is recommended for all users of NexusGuard (Python and npm) and ClassificationGuard. The release also ships the structured nexus claim described in the August 31 and September 1 entries.

Behavior change. Classification claims and worker facts that earlier versions interpreted are now rejected:

# Before: "1099 contractor" was matched by substring and treated as 1099.
# After: verified=False with an INVALID_CLAIM audit trace outcome.
guard.verify_classification_claim(
    llm_claim="1099 contractor",
    facts={"provides_tools": False, "reimburses_expenses": False, "indefinite_relationship": False},
)

# After: pass a canonical claim and boolean facts.
guard.verify_classification_claim(
    llm_claim="1099",
    facts={"provides_tools": False, "reimburses_expenses": False, "indefinite_relationship": False},
)

What changed

Closed claim vocabulary. ClassificationGuard.verify_classification_claim maps the claim through a fixed set of values instead of substring matching. W2, W-2, and EMPLOYEE map to employee. 1099, 1099-NEC, CONTRACTOR, and INDEPENDENT CONTRACTOR map to contractor. Matching ignores case, spaces, hyphens, and underscores. Any other text fails closed with INVALID_CLAIM.
Strict worker facts. provides_tools, reimburses_expenses, and indefinite_relationship must all be present booleans. Otherwise the result fails closed with INVALID_FACTS. Strings such as "no" are rejected.
TaxPreFlight hire intents. The same vocabulary and fact rules apply to worker_type and worker_facts.
npm NexusGuard fails closed. The @qwed-ai/tax NexusGuard blocks states that are not in its threshold table and rejects invalid sales data, including non-object, non-finite, and negative values. The transaction-count prong applies only to states that have one.
npm classification unchanged. The npm SDK still compares worker_type exactly. Pass "W2" or "1099".

Migration

Translate the AI’s classification into "W2" or "1099" and pass each worker fact as a real boolean — map known values explicitly ("yes"/"true" → True, "no"/"false" → False) and reject anything else. Do not wrap a string in bool(): bool("no") is True, which would recreate the bug this release fixes. Nexus callers that relied on llm_decision strings must pass claimed_collects_tax as a boolean. See ClassificationGuard and NexusGuard.

QWED-Legal v0.5.1 — DeadlineGuard signing-anchor allowlist

Released: October 9, 2026 in qwed-legal 0.5.1 (PyPI and npm @qwed-ai/legal) · GitHub PR #102 · GitHub PR #103

Security patch. DeadlineGuard now computes a deadline from signing_date only for terms measured forward from signing. Upgrade if you use DeadlineGuard through the Python API, the npm SDK, or the GitHub Action.

Behavior change. Terms that are measured backward or anchored to another event now return UNVERIFIABLE instead of VERIFIED:

# Before: computed signing + 30 days and could return verified=True.
# After: UNVERIFIABLE with an UNSUPPORTED trace step.
guard.verify(signing_date="2026-01-15", term="30 days before signing", claimed_deadline="2026-02-14")
guard.verify(signing_date="2026-01-15", term="30 days after closing", claimed_deadline="2026-02-14")

# Migration for a forward term anchored to another event:
# confirm the event date, pass it as signing_date, and rewrite the term.
guard.verify(signing_date="2026-03-02", term="30 days", claimed_deadline="2026-04-01")

Changing signing_date alone does not make an event-anchored term supported. You must also rewrite the term to a bare-relative form such as "30 days".

What changed

Signing-anchor allowlist. DeadlineGuard computes from signing_date only when the term is bare-relative ("30 days") or explicitly anchored to signing or execution ("30 days from signing", "15 days after execution"). Every other term fails closed with UNVERIFIABLE and an UNSUPPORTED trace step.
Directional terms fail closed. Terms using before, prior to, ahead of, or preceding were previously computed forward from signing, which could certify the opposite direction.
Any non-signing event anchor fails closed. Anchors outside the earlier event list, such as "after closing", "after completion", or "within 30 days of request", are now rejected. The open-ended event-noun list is replaced by an allowlist of signing-anchored phrasings.
Check runs before date arithmetic. Rejected business-day terms never consult the holiday calendar.
npm engine version gate. DeadlineVerifier in @qwed-ai/legal rejects its promise when the importable Python qwed-legal is older than 0.5.1. An npm-only upgrade cannot silently keep the old deadline engine. Upgrade both packages:

pip install "qwed-legal>=0.5.1"
npm install @qwed-ai/legal@0.5.1

See Supported term forms and Event-anchored terms.

v7.2.2 — Verification soundness and safety hardening

Released: October 9, 2026 · GitHub Release ↗ · GitHub PR #434 · Advisories GHSA-4v5r-g7f4-vvgc, GHSA-mfh5-3c8f-975p, GHSA-j622-qffp-rc27, GHSA-mxwv-x5qm-mrmf

v7.2.2 is a fail-closed security release covering four coordinated advisories. They affect all versions up to v7.2.1. No new capability and no breaking wire changes. Upgrade is recommended for any deployment on v6.0.0 through v7.2.1.

Security fixes

Local SDK math verification. The SymPy expression validator in QWEDLocal now rejects string literals nested inside container arguments. This closes an allow-list bypass that could route untrusted model output to SymPy’s string-evaluation path.
Logic verification soundness. The Z3 constraint evaluator no longer accepts Python boolean operators (and, or, not) or chained comparisons (a < b < c) on symbolic objects. The DSL compiler now rejects string literals instead of materializing them as variables. Both cases could previously yield a VERIFIED verdict that did not follow from the input.
Large-number math verification. verify_math now compares results at a precision scaled to the operands. Values beyond 15 significant digits are no longer rounded before the comparison, so a wrong claim on a large number can no longer pass as VERIFIED.
Computational cost bounds. The math parser applies its compute-cost gates to implicit-multiplication input (2x, 3(x+1)). The logic and DSL engines bound statically known integer powers. Short requests can no longer exhaust a worker.

Version bumps

Surface	From	To
`qwed` (PyPI)	`7.2.1`	`7.2.2`
`qwed_sdk` (Python)	`7.2.1`	`7.2.2`
`@qwed-ai/sdk` (npm)	`7.2.0`	`7.2.2`
`qwed` (crates.io)	`7.2.1`	`7.2.2`
API version marker	`7.2.1`	`7.2.2`

Upgrading

Upgrading from before v7.2: re-issue your API keys. Self-hosted deployments must also set QWED_API_KEY_LOOKUP_SECRET to a value distinct from QWED_JWT_SECRET_KEY before starting the server. Without these steps, old keys return 401 Unauthorized and a server without the secret refuses to start. See Authentication.

Patch release with no code changes required for valid inputs. Pin the new version and redeploy:

pip install "qwed==7.2.2"
npm install @qwed-ai/sdk@7.2.2
cargo add qwed@7.2.2

Inputs that relied on Python and/or/not, chained comparisons on symbolic variables, or string literals in DSL expressions now fail closed instead of returning VERIFIED. Rewrite Z3 constraints with And(...), Or(...), Not(...) and separate comparisons, and write DSL logic with the AND/OR/NOT operators.

QWED Open Responses v0.6.1 — ToolGuard verifies client action items

Released: October 9, 2026 in qwed-open-responses 0.6.1 (PyPI and npm) · GHSA-xhq6-w3f2-m5w6 · PR #71

ToolGuard now checks Responses API items that ask your client to run shell commands, apply patches, or control a computer. Before v0.6.1, these items read as tool-free content and passed. Upgrade from v0.6.0.

Breaking change. The default blocklist adds local_shell, apply_patch, and computer_use. A default ToolGuard now blocks local_shell_call, shell_call, apply_patch_call, and computer_call items. To allow them, turn off the default blocklist and allowlist the tools you run:

# Before: local_shell_call items passed a default ToolGuard
ToolGuard()

# After: opt out of the defaults and allowlist the action tools you run
ToolGuard(
    use_default_blocklist=False,
    blocked_tools=["bash", "sh", "powershell"],
    allowed_tools=["local_shell", "apply_patch"],
)

new ToolGuard({
  useDefaultBlocklist: false,
  blockedTools: ["bash", "sh", "powershell"],
  allowedTools: ["local_shell", "apply_patch"],
});

use_default_blocklist=False drops every default entry, shells included. Re-add the names you still want blocked.

What changed

Client action items are tool calls. ToolGuard normalizes local_shell_call, shell_call, apply_patch_call, computer_call, custom_tool_call, and mcp_approval_request items into tool calls. Blocklist, allowlist, dangerous-pattern, and validator checks apply. This covers direct items, output[] of a response object, tool_calls[] entries, and the streaming middleware.
Canonical tool names. Action items are checked as local_shell, shell, apply_patch, and computer_use. Custom tool and MCP approval items use their own name. local_shell argv lists are also scanned as a joined command line.
Unknown executable items fail closed. Any other *_call or *_request item in a streamed item, output_item event, or output[] of a response object is blocked. Provider-hosted items (web_search_call, file_search_call, code_interpreter_call, image_generation_call, mcp_call) keep their existing behavior.
Ambiguous action items fail closed. An action item that also carries a tool_call, function_call, function, or tool_calls envelope is blocked. So is a named item whose tool_name or toolName disagrees with name.

What this means for you

If your agent runs local shell, patch, or computer-use tools, configure ToolGuard with use_default_blocklist=False and an allowed_tools allowlist.
If you rely on the default blocklist for shells, re-add those names through blocked_tools when you opt out.

See Responses API client action items in the guards reference.

QWED-Finance v3.0.1 — ISO 20022, sanctions, and QueryGuard security fixes

Released: October 9, 2026 in qwed-finance 3.0.1 (PyPI, npm @qwed-ai/finance, and GitHub Action QWED-AI/qwed-finance@v3.0.1) · Changelog

qwed-finance v3.0.1 is the current release. It fixes three security advisories. Upgrade from v2.1.0 or v3.0.0.

What changed

ISO 20022 business rules read the parsed document (GHSA-mrrj-6m2q-jch9). CrossGuard.check_business_rules, verify_iso20022_with_rules, and UCPIntegration.verify_iso20022_payment take amounts and currencies from the parsed XML tree instead of regex over raw text. Amounts with child nodes, exponents, or grouping separators fail closed.
Sanctions screening resists in-word perturbations (GHSA-mv2c-jwm9-pfrq). Diacritics and look-alike letters are folded, and separator-free names are compared, so E.V.I.L. CORP or ÉVIL CORP no longer clears an EVIL CORP entry.
QueryGuard rejects MySQL executable comments (GHSA-q8r4-6gpp-5fx2). Queries containing /*! ... */ or /*M! ... */ are unsafe in every QueryGuard check and in CrossGuard.verify_query_with_pii_protection.

Included from v3.0.0 (September 27, 2026)

Breaking (npm): the @qwed-ai/finance bridge sends JSON over stdin with runtime validation. Malformed inputs are rejected instead of falling back to defaults.
VerificationReceipt.get_signature(key): keyed HMAC-SHA256 over all 15 receipt fields, with no default key. The unkeyed get_signature() was removed.
ISO currency and amount business limits in receipts and UCP. UCPIntegration.verify_iso20022_payment takes kyc_verified and requires it to approve.
Fail-closed AML, sanctions, amount, and message-validation fixes.

See QWED Finance overview, the 11 guards, and GitHub Action.

QWED-Legal — fail-closed date handling and input bounds

Released: October 7, 2026 in qwed-legal 0.5.0 · GitHub PR #89 · GitHub PR #90 · GitHub PR #91 · GitHub PR #92 · GitHub PR #93

DeadlineGuard and StatuteOfLimitationsGuard no longer guess what a date means. Order-ambiguous dates, partial dates, and event-anchored terms now return UNVERIFIABLE instead of a confident verdict. ClauseGuard and ContradictionGuard gain input bounds.

Behavior change. Some inputs that used to verify now return UNVERIFIABLE:

# Before: "03/04/2026" was read month-first (March 4) and the claim verified.
# After: UNVERIFIABLE with an AMBIGUITY_NOTED trace step.
guard.verify(signing_date="03/04/2026", term="30 days", claimed_deadline="2026-04-03")

# Before: computed 30 days from signing.
# After: UNVERIFIABLE, because the clock starts at receipt, not signing.
guard.verify(signing_date="2026-01-01", term="30 days after receipt", claimed_deadline="2026-01-31")

Pass ISO YYYY-MM-DD dates. ISO year-first and named-month dates still verify.

What changed

Order-ambiguous numeric dates fail closed (#90). DeadlineGuard and StatuteOfLimitationsGuard, including computation-only mode, return UNVERIFIABLE with an AMBIGUITY_NOTED trace step for dates such as 03/04/2026 that read differently month-first and day-first.
Partial dates fail closed (#91). Inputs such as "March 2024", "Friday", or "23:00" would be completed from the system clock, so both guards reject them. In both cases, parsed datetimes in the result are None and the raw strings stay in the trace. DeadlineGuard.calculate_business_days_between() raises ValueError for these dates.
Event-anchored terms fail closed (#92). DeadlineGuard returns UNVERIFIABLE for terms that start the clock at receipt, notice, delivery, occurrence, demand, invoice, breach, termination, service of process, or payment. It no longer computes those deadlines from the signing date.
Deadline comparison by calendar date (#92). Claimed and computed deadlines are compared by date, not timestamp. Timezone-aware claims are converted to the computed deadline’s timezone. Mixed timezone-aware and naive inputs return UNVERIFIABLE. tolerance_days stays in whole days.
ClauseGuard input bounds (#89). check_consistency() returns invalid_input for more than 200 clauses or any clause over 4,096 characters. It reports at most 100 conflict pairs and adds a truncation notice when it stops.
Digit-run bound (#89). Digit runs longer than 18 characters fail closed in DeadlineGuard terms, ClauseGuard day counts, and ContradictionGuard operands.
Faster US citation patterns (#89). CitationGuard US patterns no longer slow down on adversarial input.
Added — npm SDK (#93). JurisdictionVerifier.verifyChoiceOfLaw() in @qwed-ai/legal accepts an optional fourth contractType argument and passes it to the guard as contract_type. JurisdictionResult adds optional contractType and contractClassification (goods, declared_non_goods, or unclassified) fields. Omitting contractType keeps the existing partial-coverage warning. The 0.5.0 npm package is published as @qwed-ai/legal.

What this means for you

Normalize dates to ISO YYYY-MM-DD before calling the deadline and statute guards.
For event-anchored deadlines, pass the event date as signing_date and drop the anchor from the term. The result reports that date as signing_date, so keep the real signing date separately.
Treat invalid_input from check_consistency() on oversized clause lists as a rejection, not a pass.

See ambiguous and partial dates, event-anchored terms, date-granularity comparison, StatuteOfLimitationsGuard, and ClauseGuard input validation.

QWED Open Responses v0.6.0 — fail-closed hardening

Released: October 2, 2026 in qwed-open-responses 0.6.0 · PR #48 · PR #49 · PR #50 · PR #51 · PR #52 · PR #55 · PR #57 · PR #58 · PR #59

SchemaGuard now rejects undeclared root fields by default, verification is resource-bounded, and ToolGuard, the streaming middleware, and the Express middleware close several bypasses. The Python and npm packages both move to 0.6.0.

Breaking change. SchemaGuard adds additionalProperties: false to the root object when your schema doesn’t set it. Responses with undeclared top-level fields now fail. An explicit additionalProperties in your schema still wins. To keep the old behavior, opt out in the constructor:

# Before: extra root fields passed silently
SchemaGuard(schema)

# After: opt out explicitly to allow them
SchemaGuard(schema, allow_additional_properties=True)

new SchemaGuard(schema, { allowAdditionalProperties: true });

What changed

SchemaGuard cross-language parity. JSON Schema format validation for email, uuid, date-time, uri, and ipv4 now matches across Python and npm.
Resource-bounded verification. JSON responses over 100,000 characters or 100 nesting levels, and responses with oversized integers, return a failed verdict instead of raising an exception. SafetyGuard scanning fails closed above 10,000 nodes, 100,000 characters, or 12 nesting levels. Email PII detection uses a linear-time scan.
SchemaGuard reports the first error only. Failed results carry one error, with the total error count reported as unknown (total_errors=None, errors_truncated=True).
ToolGuard raw argument scanning. Dangerous patterns run on the raw parsed argument strings, including nested values and JSON-encoded calls, so custom patterns match across tabs and newlines. Cyclic or too-deep arguments still fail closed.
ToolGuard name and command normalization. Blocked tool names are normalized for case, surrounding whitespace, dashes, underscores, and camelCase. The default blocklist adds run_command, execute_command, command_line, and terminal. Command checks catch comment-separated keywords, spaced dots, and reordered rm flags, including inside JSON- and base64-encoded arguments. Allowlist entries still require the exact configured separators. Custom validator keys match case and whitespace but don’t merge distinct names.
Streaming tool-call verification. OpenResponsesMiddleware.verify_stream verifies each tool item in its original shape before yielding it. Tool-call types are normalized for case and whitespace. Unknown, ambiguous, or nested tool-shaped items are blocked. tool_result and function_call_output items pass through when their correlation metadata is valid. ArgumentGuard and TaxGuard checks cover batched and nested calls.
SafetyGuard scans field names and quoted credentials. Object field names are now checked for PII and prompt injection, not only values. Quoted JSON credentials such as "api_key": "sk-..." are blocked. Documented placeholders stay exempt.
npm SafetyGuard configuration parity. The TypeScript SafetyGuard accepts customPatterns and piiAllowList. Each matched custom pattern blocks and reports its own custom_pattern finding. Invalid patterns throw at construction, and a pattern that exceeds its 25 ms execution limit fails the check. piiAllowList accepts camelCase and snake_case names, such as ipAddress and ip_address.
ToolGuard envelope hardening. Envelopes that carry conflicting tool identities at the root, in a nested call, or in a function wrapper are blocked. Configured ToolGuards now check structured_output responses too. Nested scans that reach the depth limit block instead of reporting “No tool calls”.
Verification context propagation. A context with usage caps now flows through verify_tool_call, verify_structured_output, and the OpenAI, streaming, LangChain, and LlamaIndex integrations. The OpenAI wrapper takes it as verification_context and doesn’t send it to OpenAI. The streaming middleware carries running total_cost and total_tokens forward, so later items can hit SafetyGuard budget limits.
Express middleware fails closed on unparsed bodies. The npm verifyRequestBody middleware returns HTTP 422 when a request declares a body but has no parsed body, even with blockOnFailure: false. Bodyless requests and parsed empty JSON continue.

What this means for you

Review SchemaGuard schemas for responses that carry extra root fields, and declare them or opt out.
If you use the Express middleware, register a body parser such as express.json() before it.
If ToolGuard now blocks your structured outputs or wrapped tool calls, give each call a single tool identity.
To enforce budgets across a stream or agent run, pass usage totals in the verification context.

See the guards reference, streaming middleware, OpenAI integration, and troubleshooting.

QWED-Legal — fail-closed hardening for guards, diagnostics, and GitHub Action

Released: October 7, 2026 in qwed-legal 0.5.0 (merged September 29, 2026) · GitHub PR #83 · GitHub PR #84 · GitHub PR #85 · GitHub PR #87 · GitHub PR #88

Five QWED-Legal paths that could report success over empty, blank, or unclassified input now fail closed. The QWED Legal GitHub Action now passes only on deterministic proofs.

Breaking change for the QWED Legal GitHub Action. Every block is now gated on to_diagnostic().status == VERIFIED. These runs previously passed and now fail with verified=false and a non-zero exit:

Unknown or misspelled mode values.
Runs where missing inputs mean zero checks execute.
Clause runs, including heuristic "consistent" results and single-clause inputs (now insufficient_input, consistent=False).
Citation runs, including format-valid citations.

Before: mode: citation with a well-formed citation, or mode: clause with no detected conflicts, returned verified=true.After: Only deterministic deadline and liability exact matches keep verified=true. If limited coverage is acceptable for your workflow and the run should still finish green, mark the QWED Legal step continue-on-error: true. That matters because a failed step both skips the later steps and fails the job: adding if: always() to a later step lets it read the output, but it cannot make the run pass. With continue-on-error: true, branch on the results JSON using steps.<id>.outputs.results — results.citation.valid for citation format checks and results.clause.status for clause results. Use verify_using_z3() from the Python API for formal clause proofs.

What changed

LiabilityGuard.verify_tiered_liability: empty tier lists. tiers=[] returns UNVERIFIABLE with total_computed=None and no proof_ref, whether claimed_total is zero or not.
ProvenanceGuard: strict required fields. content_hash, model_id, and generation_timestamp must be non-empty strings. None, 0, False, [], and whitespace now give INCOMPLETE_PROVENANCE. Checks that cannot run are recorded as hash_integrity_skipped, timestamp_valid_skipped, and model_allowed_skipped instead of CONTENT_TAMPERED or UNAUTHORIZED_MODEL.
JurisdictionGuard: blank party entries. check_convention_applicability() returns UNVERIFIABLE for blank entries such as [""]. Genuine negatives such as ["XX"] stay deterministic.
JurisdictionGuard: three-way contract_type classification for CISG. Goods contracts (separator and case variants accepted, bare "sale" excluded) get the CISG warning. Declared non-goods contracts such as "services" get no CISG warning. Missing, blank, or unknown values get a partial-coverage warning and are not verified. The US and non-US party pair no longer triggers the CISG warning by itself. The classification is recorded in trace inputs.
Diagnostics: no proof_ref without a deterministic step. LegalDiagnosticResult.verified() raises unless the evidence trace has a DETERMINISTIC step. Traceless or transport-stripped "consistent" verdicts and empty claim_texts are demoted to UNVERIFIABLE. Blank messages fall back to defaults instead of crashing.

See LiabilityGuard, CISG detection, ProvenanceGuard, ClauseGuard input validation, and No proof without a deterministic step.

QWED Verification — `qwed init` server identity handshake

Released: September 23, 2026 · GitHub PR #388 · Closes issue #376

qwed init now proves that the local server holds this project’s secrets before it bootstraps an API key. Previously, a healthy /health only proved that something answered on the port. A foreign process or a stale server running older secret values could receive the bootstrap and issue a key that stops working after a restart.

What changed

New GET /health/identity endpoint. Takes a caller-supplied nonce (1–128 characters) and returns HMAC-SHA256 proofs over it under QWED_JWT_SECRET_KEY and QWED_API_KEY_LOOKUP_SECRET. Returns 400 for a missing or oversized nonce and 503 when server secrets aren’t configured. The secrets never leave the server. See GET /health/identity.
Nonce-bound, replay-safe handshake. qwed init generates a fresh nonce on every run and compares the proofs against locally computed values in constant time. A captured response is useless for any later handshake.
Runs on reused and freshly spawned servers. A reused server can be foreign or stale even when secrets were retained. A fresh spawn can lose a port race after the health poll.
Fails closed. A mismatch, a 404 from an older server, any other non-200 status, a malformed response, an unreachable server, or missing local secrets stops qwed init with exit status 1 before any key is issued. Each message tells you to restart the local server or re-run qwed init.

See qwed init and troubleshooting.

QWED Verification — leaked API key auto-revocation and false-positive feedback

Released: September 22, 2026 · GitHub PR #380 · GitHub PR #386 · Closes issue #368 and issue #369

Verified reports from POST /webhooks/secret-scanning now trigger enforcement. QWED automatically revokes any matching active API key, records a security event, and emails the key owner. An opt-in reporter can send true-positive and false-positive labels back to the scanning service.

What changed

Automatic revocation. A verified match against an active key sets is_active=False and revoked_at, and writes a SECRET_LEAK_AUTO_REVOKED security event. The event records only the key_preview. The plaintext token is never persisted, logged, or emailed.
Owner notification. QWED emails the key’s user, falling back to the organization’s first active owner or admin. SMTP phases time out after 10 seconds. Revocation is committed first and is never rolled back if notification fails.
Idempotent handling. Already-revoked keys and unknown tokens are not acted on. Redelivered or concurrent reports revoke and notify exactly once. Registered scanner patterns cover qwed_live_ keys only, so qwed_test_ keys are never reported.
False-positive feedback reporter (opt-in). Set QWED_LEAK_FEEDBACK_ENABLED=true and QWED_LEAK_FEEDBACK_URL to POST one label per match after the revocation attempt. Items carry SHA-256 token hashes by default. QWED_LEAK_FEEDBACK_SEND_RAW=true is a separate explicit opt-in to send plaintext tokens, and only works with an HTTPS URL.

See Security hardening — Automatic revocation of leaked keys for the full behavior and configuration.

QWED Verification — canonical API-key format spec

Released: September 22, 2026 · GitHub PR #387 · Closes issue #370

QWED now publishes a canonical, machine-readable specification of its API-key formats, with lookaround-anchored scanning regexes and synthetic test vectors. Use it to build secret scanners, leak detectors, and SDK-side format validators.

What changed

v2 checksummed format documented. Current keys match qwed_live_[A-Za-z0-9]{36}: 30 random Base62 characters followed by a 6-character Base62 CRC32 checksum that you can validate offline.
Legacy v1 still accepted. Keys matching qwed_live_[A-Za-z0-9_-]{43} keep working and are never force-rotated.
qwed_test_ exclusion. Scanners must exclude test keys from high-confidence production-secret matching.
Lookaround-anchored regexes. Published patterns use charset-complement lookarounds instead of \b, because a v1 body can end in -. RE2-compatible delimiter-class alternatives are included.
Masking guidance. Masked previews show the prefix, asterisks, and only the last 4 characters.
Machine-readable spec and test vectors. spec/api-key-format.json and spec/api-key-test-vectors.json in the qwed-verification repository.

See API key format for patterns, checksum details, and scanning rules.

QWED Verification — signed secret-scanning webhook endpoint

Released: September 20, 2026 · GitHub PR #374 · Closes issue #367

New POST /webhooks/secret-scanning endpoint receives signed reports of leaked QWED API keys from GitHub secret scanning and compatible partner scanners. The endpoint is unauthenticated at the transport layer by design — external scanners have no QWED credentials — and the ECDSA-NIST-P256V1-SHA256 signature over the raw request body is the only credential. Verification runs before any JSON parsing, logging, or database work, and every failure mode fails closed to a distinct status code.

What changed

New public endpoint. POST /webhooks/secret-scanning accepts a JSON array of leaked-token matches, verifies the Github-Public-Key-Identifier and Github-Public-Key-Signature headers against GitHub’s published signing keys, and hands verified batches to the revocation sink in the background. See POST /webhooks/secret-scanning for the full request/response contract.
Raw-body signature verification. The signature is verified over the exact bytes received; the body is never re-serialized. Nothing in the payload is parsed or logged until the signature is proven genuine.
P-256 pinning at key-load. Signing keys are pinned to secp256r1 when the PEM is loaded, so a future keys-list entry on another curve is dropped at load rather than silently trusted at verify time.
6-hour key cache with rotation retry. Signing keys are cached in-process for 6 hours with an ETag-conditional refetch. On an unknown key_identifier, QWED refetches the list exactly once (throttled to one forced refetch per 60 s after a success, 10 s after a failure) and retries. A dead or throttled key service fails closed to 503.
Timing-safe verdict comparison. Success and failure branches run through hmac.compare_digest on a fixed-shape verdict, and every failure returns the same generic detail so the failing step is not observable.
Fail-closed status separation. 401 for missing headers, 403 for a proven-bad signature against a fresh anchor, 503 for any state where verification was impossible (signing-key service down, anchor unusable, rotation retry throttled). 413 for over-cap bodies (1 MiB) or batches (500 matches); 400 for a malformed envelope. Tokens never appear in logs.
Optional keys-fetch token. GITHUB_KEYS_TOKEN (or GITHUB_TOKEN as a fallback) can be set to raise the outbound rate limit on the keys request. Neither token grants any authority over the incoming webhook — only the signature does.

See Security hardening — Secret-scanning webhook intake for the full trust model and operator configuration.

QWED Verification — `qwed init` provisions `QWED_API_KEY_LOOKUP_SECRET`

Released: September 21, 2026 · GitHub PR #375 · Closes issue #372

Fresh qwed init runs no longer crash the child server at boot on the fail-closed QWED_API_KEY_LOOKUP_SECRET import check. The CLI now generates the lookup secret alongside QWED_JWT_SECRET_KEY, writes it to .env, and threads it into the local server it starts.

What changed

Auto-provisioned on first run. qwed init generates a URL-safe random QWED_API_KEY_LOOKUP_SECRET (at least 48 bytes) whenever one is not already set, persists it to .env with the same 0600 permissions as other credentials, and exports it into the child server’s environment.
Retained on re-runs, never rotated. If .env, the process environment, or the project-root .env already holds a value, qwed init reuses it verbatim. Rotating this secret invalidates every stored API-key lookup digest, so the CLI will not do it for you.
Refuses to persist an unbootable value. If the existing QWED_API_KEY_LOOKUP_SECRET equals QWED_JWT_SECRET_KEY, qwed init exits with an error instead of writing a value the server would refuse to start with. Remove the offending line (a fresh distinct value is generated) or set a distinct value manually.
Stale-secret guard around the local server. If qwed init regenerates a missing secret but a QWED server is already running with older values, the CLI stops before bootstrapping a new API key — issuing one against a server that cannot verify it would produce a key that dies on the next restart.

See qwed init and API key storage and migration for the full contract.

QWED-A2A — `/a2a/intercept` requires per-agent API keys (breaking)

Released: September 9, 2026 · GitHub PR #106

The FastAPI gateway’s POST /a2a/intercept endpoint previously trusted the body-declared sender_agent_id — any caller could impersonate any agent for trust evaluation, rate limits, and attestation. The endpoint now authenticates every request with a per-agent API key in the X-API-Key header, validated against the QWED_A2A_API_KEYS environment variable. The authenticated identity overrides the body’s sender_agent_id everywhere it matters. The dead AgentMessage.signature field, which was never verified, is removed.

Fail-closed behavior change. Deployments that do not set QWED_A2A_API_KEYS now deny every /a2a/intercept request with 401 API authentication is not configured for this service. There is no anonymous access and no opt-out. Configure a key for each calling agent before upgrading.

What changed

Per-agent API-key authentication — QWED_A2A_API_KEYS holds a JSON object mapping each key to its agent ID, e.g. {"key-abc123": "procurement-agent"}. Missing, unknown, or unconfigured keys return 401. Keys are compared in constant time.
Authenticated identity wins — trust-boundary evaluation, token-bucket rate limiting, telemetry, and the attestation JWT all bind to the key’s agent ID. If the body’s sender_agent_id differs, the gateway logs a warning and overrides it. Direct library callers of intercept() get no override and own identity binding themselves.
No-restart key rotation — the key map is read fresh on every request. Update the environment variable and old keys stop matching immediately.
Agent-ID normalization — map entries are trimmed and validated against the AgentMessage contract (max 256 chars, no control characters); invalid entries are dropped and their keys denied.
AgentMessage.signature removed — the optional field was documented as “JWT signature for tamper detection” but was never read by any verification path. Attestation integrity comes from the verdict JWT, not a caller-supplied field.

Migration

Before (now returns 401):

curl -X POST http://localhost:8000/a2a/intercept \
  -H "Content-Type: application/json" \
  -d '{"sender_agent_id": "agent-A", "receiver_agent_id": "agent-B", "payload_type": "general", "payload": {"message": "Hello!"}}'

After — configure a key map on the gateway and send the header:

export QWED_A2A_API_KEYS='{"key-for-A": "agent-A"}'

curl -X POST http://localhost:8000/a2a/intercept \
  -H "Content-Type: application/json" \
  -H "X-API-Key: key-for-A" \
  -d '{"sender_agent_id": "agent-A", "receiver_agent_id": "agent-B", "payload_type": "general", "payload": {"message": "Hello!"}}'

If any client still sends signature in the message body, remove the field. See Gateway authentication for the full key format and rotation guidance.

QWED-A2A — cross-deployment trusted issuers and jti replay lifecycle (security release)

Released: September 9, 2026 · GitHub PR #108 · GitHub PR #109 · Closes issue #84 and issue #85

Two fixes to A2ACryptoService attestation verification. Peer deployments can now be registered as trusted issuers, so verifying another deployment’s attestations no longer tempts operators into sharing one private key across deployments. And the jti replay registry now separates issuance from consumption: an issuer can verify its own token once, duplicate trace_ids are rejected at signing, and the consumption registry is injectable for cross-worker replay protection.

Cross-deployment verification (PR #108)

Trusted issuers. verify_attestation accepts an optional trusted_issuers mapping of issuer IDs to {"deployment_id": ..., "jwks": ...}, with the QWED_A2A_TRUSTED_ISSUERS environment variable as fallback. Register a peer by copying its /.well-known/jwks.json entry verbatim. The env var is re-read on every verification, so key rotation needs no restart.
Fail-closed on every mismatch. Unknown issuers, kid mismatches, and deployment_id mismatches all deny. Routing never trusts unverified claims: header and payload are read only after a candidate key verifies the signature, and one public key registered under two issuer identities keeps only the first identity.
Verifier-only nodes. A node with no local signing key can verify peer tokens against trusted issuers. Only signing, JWKS publishing, and the interceptor still require QWED_A2A_SIGNING_KEY_PEM. With neither a local key nor trusted issuers, verification returns No verification keys available instead of raising.
Peer jti namespacing. Replay tracking for peer tokens is namespaced per issuer, so two deployments reusing the same trace ID never shadow each other.

jti registry lifecycle (PR #109)

Issuance/consumption split. sign_verdict records the trace_id in a separate per-instance issuance record instead of the consumption registry, so an issuer can verify its own token once — the second presentation is a replay.
Duplicate trace IDs rejected at signing. Reusing a trace_id within its validity window raises ValueError at sign_verdict instead of minting two tokens under one jti.
Injectable consumption registry. A2ACryptoService(jti_registry=...) accepts a shared registry (Redis, database) for cross-worker replay protection. The default remains process-local: two workers in two processes share nothing. Injected registries must provide atomic insert-if-absent, retention covering token lifetime, and a numeric ttl_seconds at least validity_seconds — anything less raises ValueError at construction.

Upgrading

Behavior changes. Code that relied on signing twice with the same trace_id now gets ValueError at the second sign_verdict. Code that verified the same token twice on one instance now gets Replay detected: jti already seen on the second call — previously issuers could not self-verify at all, while consumers could sometimes replay. Verifier-only nodes that expected verify_attestation to raise RuntimeError without a signing key now get a (False, None, ...) tuple or, with trusted issuers configured, successful peer verification. Multi-worker deployments should inject a shared jti_registry — the default registry does not detect replays across processes.

See Crypto attestations for configuration and the full replay-scope documentation.

v7.2.1 — Math answer comparison security patch

Released: September 14, 2026 · GitHub Release ↗ · GitHub PR #365 · Advisory GHSA-xmm6-8r3x-j567

v7.2.1 is a fail-closed security patch. The math answer comparison path previously fed the untrusted model answer to raw sympy.sympify, which is an eval sink that can execute dunder-traversal gadgets smuggled in the answer turn. The answer is now parsed only through the hardened safe_parse_expr, and unparseable input fails closed. No new capability and no breaking wire changes. Upgrade is recommended for any deployment on v6.0.0 through v7.2.0.

Security fix

Hardened answer parsing. QWEDLocal._math_answers_match no longer calls sympy.sympify on the model’s answer. It now parses through safe_parse_expr, which applies a character allow-list, an AST allow-list, strips builtins, and enforces the compute-cost bounds introduced in v7.2.0. Unparseable answers and a missing parser both return False, so the comparison never treats an attacker-controlled string as “equal” by accident.
Release boundary gate covers the SDK. scripts/check_boundary.py now scans qwed_sdk/ and forbids raw sympify outside approved wrappers, so the eval sink cannot reappear in a future refactor.

Version bumps

Surface	From	To
`qwed` (PyPI)	`7.2.0`	`7.2.1`
`qwed_sdk` (Python)	`7.2.0`	`7.2.1`
`@qwed-ai/sdk` (npm)	`7.2.0`	`7.2.1`
`qwed` (crates.io)	`7.2.0`	`7.2.1`
API version marker	`7.2.0`	`7.2.1`

Upgrading

Patch release with no code changes required. Pin qwed>=7.2.1 on PyPI, @qwed-ai/sdk@^7.2.1 on npm, and qwed = "7.2.1" on crates.io, then redeploy. Wire formats and public APIs are unchanged.

v7.2.0 — Precision advisory and security hardening batch

Released: September 7, 2026 · GitHub Release ↗ · GitHub PR #362

v7.2.0 unifies the precision advisory capability and the security hardening batch across all SDKs. One additive capability — an advisory flag for binary floating-point constants in math and stats verification — plus fail-closed fixes: bounded engine-call waits, sandbox containment, and event-loop offload. The release is a semver minor with no breaking wire changes.

Precision advisory (new capability)

Math and stats verification inputs may contain binary floating-point constants (0.1 + 0.2, 1e9) whose evaluation can be inexact relative to decimal arithmetic. v7.2.0 surfaces that signal as a DiagnosticResult advisory without ever affecting the verdict (PR #348):

AdvisoryCheck.float_precision() parses the verification source and, when float or complex constants are present, returns an advisory with constraint_id: "precision.float-constants" listing the offending constants and suggesting decimal.Decimal or SymPy exact rationals where exact arithmetic matters.
/verify/math attaches the advisory to the result on all branches, before trust enforcement.
StatsVerifier.verify_stats carries the advisory on completed analyses when the generated code uses floats — the expected shape for numpy/pandas code, so it flags without disrupting.
Advisory, never a gate. Execution-safety gates decide what may run, not what is exact. Documented inputs like 1000 * (1 + 0.05)**2 legitimately contain floats and still verify. Unparsable input returns no advisory; parse failures belong to the security gates.

{
  "developer_fields": {
    "advisory_checks": [
      {
        "name": "floating-point-constants",
        "advisory_only": true,
        "constraint_id": "precision.float-constants",
        "details": {
          "constants": ["0.1", "0.2"],
          "note": "Binary floating-point values can be inexact; results may differ from exact decimal arithmetic.",
          "suggestion": "Use decimal.Decimal or SymPy exact rationals (sympy.Rational) where exact arithmetic matters."
        }
      }
    ]
  }
}

Engine-call bounds (PR #354)

Every engine-call wait is now bounded so a single request cannot stall or exhaust the service:

SymPy compute-cost gate. The safe expression parser adds a computational-cost layer: integer-literal cap (10^300), static exponent bound (10^4), exact-expansion bounds on factorial/binomial calls, caret-chain fold bounds, and a nested-power base-magnitude budget. Expressions like 2**factorial(10000) or ((9**9)**9999)**9999 are rejected before they can reach SymPy’s eager exact expansion. All cost comparisons use exact arithmetic, never binary floats.
Provider HTTP timeouts. All six LLM provider clients are pinned to a 30-second timeout with zero SDK-level retries. The previous SDK defaults allowed roughly 30 minutes per stalled call. Retry policy belongs to the caller.
Stats upload cap. /verify/stats gains a byte-counting body-limit middleware with a 30-second read deadline and an in-flight concurrency cap, a 10 MB file cap with header preflight, and an early abort at 1 million cells.

Sandbox containment (PR #351)

The Docker code-execution sandbox can no longer leak host resources (CWE-400/401):

Log rotation and process caps. Sandbox containers run with json-file log rotation (10 MB, single file) and pids_limit=128.
Guaranteed container removal. Containers are force-removed in a finally block covering the whole post-creation lifecycle, with create-then-start ordering so start failures cannot leak, one retry for transient daemon races, and a warn-only fallback so cleanup failure never discards a valid verification result.
Result size caps. result.json is capped at 2 MB both inside the container wrapper (streamed, aborts at cap) and at host read-back before parsing. observed_result and every VerificationLog.result site are bounded by a shared pre-encode traversal with per-string caps, an aggregate budget, and cycle markers, and remain valid JSON for the audit integrity verifier.

Event-loop offload (PR #352)

/verify/consensus and /verify/stats declared async def but ran synchronous verification chains inline on the event loop, so a single low-rate tenant could stall the whole service:

Consensus now awaits an async path with a per-call executor sized to the engine list, one aggregate deadline, and per-engine timeouts. Timed-out, errored, or circuit-open engines degrade to partial BLOCKED results with correct circuit-breaker recording instead of hanging the request.
Stats offloads CSV parsing and the verification chain to a worker thread, and every Docker daemon call is bounded at 30 seconds.

Version bumps

Surface	From	To
`qwed` (PyPI)	`7.1.0`	`7.2.0`
`qwed_sdk` (Python)	`7.1.0`	`7.2.0`
`@qwed-ai/sdk` (npm)	`7.1.0`	`7.2.0`
`qwed` (crates.io)	`7.1.0`	`7.2.0`
API version marker	`7.1.0`	`7.2.0`

Included PRs

PR	Summary
#348	Advisory flag for binary floating-point constants
#351	Sandbox containment: log rotation, pids_limit, container removal, result size caps
#352	Consensus and stats verification offloaded from the event loop
#354	Bounded engine-call waits: sympy compute-cost gate, provider timeouts, stats upload cap
#362	v7.2.0 release preparation

The v7.2.0 release also rolls up the security fixes released individually since v7.1.0. See the entries below for the metrics operator allowlist, the math parser and sandbox gate hardening, and the API-key lookup migration.

Upgrading

No action required for the changes in this entry: the precision advisory is additive, and the hardening fixes restore intended behavior with no wire changes. Requests that previously stalled or over-consumed resources — unbounded exponent expressions, oversized stats uploads, oversized sandbox results — now return BLOCKED or 413-class rejections instead. If you upgrade from v7.1.0 directly, also review the API-key migration notes in the September 2 entry.

QWED Verification — all-tenant metrics restricted to explicit platform operators (security release)

Released: September 4, 2026 · GitHub PR #349 · Closes issue #337 (CWE-284, CVSS 7.1)

GET /metrics and GET /metrics/prometheus exposed every organization’s request volumes, latencies, provider usage, and per-tenant breakdowns to any self-service signup. The access gate treated the organization-scoped owner/admin role as platform-wide authority, and since signup is the only user-creation path — hardcoding role="owner" for every new account — the gate degenerated to “possess any account,” one anonymous request away. Access is now an explicit operator allowlist, fail-closed when unset.

What changed

Operator allowlist. All-tenant metrics require the caller’s user ID to be listed in QWED_METRICS_OPERATOR_USER_IDS (comma-separated). The list is read per request, so granting or rotating operators takes effect without a restart.
Org roles are no longer platform authority. owner and admin remain meaningful inside their own organization; they confer no cross-tenant access anywhere.
API-key path hardened. A key resolves through its owning user, who must be on the allowlist. Expired (expires_at) or revoked (revoked_at) keys are no longer usable credentials — but they also no longer preempt a valid operator JWT presented in the same request. Key-only callers without a valid credential are still denied.
Fail-closed by default. With the variable unset, both endpoints deny every caller. Per-tenant metrics (GET /metrics/{organization_id}) are unchanged.

Behavior changes

Deployments that scrape the Prometheus endpoint (/metrics/prometheus) without the operator allowlist and valid scraper credentials will start receiving 401/403 after upgrading. Create a dedicated operator account, mint it an API key, and add its user ID to QWED_METRICS_OPERATOR_USER_IDS before deploying. The deploy/prometheus.yml in the qwed-verification repository documents the scrape bootstrap steps. See Security Hardening.

QWED Verification — float precision advisory for math and stats verification

Released: September 3, 2026 · GitHub PR #348 · Closes issue #347

Math and stats inputs may contain binary floating-point constants (0.1 + 0.2, 1e9) whose evaluation can be inexact relative to decimal arithmetic. A new advisory surfaces that signal without gating anything: it is a precision advisory, never a rejection, because execution-safety gates decide what may run, not what is exact.

What changed

precision.float-constants advisory. When an expression or generated statistics code contains float or complex constants, the result carries an AdvisoryCheck in developer_fields.advisory_checks listing the offending constants and suggesting decimal.Decimal or exact SymPy rationals.
/verify/math. The advisory is attached on all result branches. The submitted expression is checked lexically, so equations (0.1 + 0.2 = 0.3, 0.5x = 0.5x) and collapsing forms retain their float literals even when symbolic simplification would eliminate them. Scientific notation and complex literals (1.0j) are flagged too.
Stats verification. The completed-analysis result carries the advisory when the generated code uses floats — the expected shape for numpy/pandas code, so it flags without disrupting.
Structurally non-verdict-affecting. AdvisoryCheck enforces advisory_only=True at construction. The advisory cannot change the verification status or proof_ref. Unparsable input yields no advisory; parse failures belong to the security gates.

See Math engine — Float precision advisory and Stats engine — Float precision advisory.

QWED Verification — structural math parser and sandbox gate hardening (security release)

Released: September 3, 2026 · GitHub PR #344 · GitHub PR #346

Two security fixes close sandbox-escape bypasses in math expression parsing and sandboxed code execution. The math parser now validates expressions structurally (NFKC normalization, ASCII charset gate, AST node allowlist) instead of relying on a denylist, and the consensus and stats execution paths gained a single-expression structural gate plus a module-indirection blocklist.

Math expression parser (PR #344)

The safe expression parser applies layered structural validation before any SymPy evaluation:

NFKC normalization first. CPython normalizes identifiers at compile time, so Unicode look-alike codepoints (mathematical bold letters, fullwidth forms) could previously slip past string filters. All checks now see exactly what the compiler sees.
ASCII charset allowlist. Only letters, digits, underscore, whitespace, and + - * / ( ) . , ^ % are accepted. A . is legal only as a decimal point with digits on both sides, which makes attribute access structurally impossible on the implicit-multiplication path.
AST node-type allowlist. Python-parseable input may contain only arithmetic operators, function calls, names, and numeric constants. Attribute access, subscripts, string constants, lambdas, comprehensions, and comparisons are rejected before evaluation.
Caret exponentiation fixed. convert_xor joined the default transformation pipeline, so x^2 parses as x**2 instead of failing at evaluation time.
The regex denylist is retained as defense in depth.

Execution gates (PR #346)

Single-expression gate on translated math. A translated math expression must parse as exactly one Python expression (ast.parse in eval mode). Multi-statement, import-bearing, and multi-line-smuggled input fails by construction. Expressions over 500 characters are rejected before parsing.
Module-indirection blocklist. The pre-execution AST check on sandboxed code now inspects every dotted segment of imports and attribute chains. Blocked module roots were extended (posix, nt, importlib, ctypes, builtins), OS-primitive call names (system, popen, import_module, the exec*/spawn* families, fork) are matched on bare names and attribute targets, and package-internal re-export gadgets through pandas/numpy internals are caught.
sys read-only allowlist in the stats executor. Only named read-only interpreter metadata is accessible; frame introspection and every other sys member fail closed.

Behavior changes

Math expression inputs that previously parsed may now be rejected before evaluation:

Bare decimals are rejected — write 0.5 and 2.0, not .5 and 2..
Non-ASCII symbols are rejected — write alpha, not α.
String constants, attribute access, and comparisons inside math expressions are rejected.
Multi-statement or import-bearing expressions fail by construction.

Legitimate arithmetic, implicit multiplication (2x, sin x), and pandas/numpy public APIs in generated stats code (np.linalg.norm, pd.Timestamp.now) are unaffected.

What this means for you

A character denylist can never defend an eval sink — these releases replace pattern-matching known attacks with structural guarantees that reject entire vulnerability classes by construction. See Math engine — Expression input rules, Consensus engine — Secure execution gates, and Stats engine — Pre-execution security validation.

QWED Verification — API-key lookup migration and per-IP auth throttling

Released: September 2, 2026 · GitHub PR #345

API-key lookup digests move from PBKDF2 to HMAC-SHA256 keyed by a new required QWED_API_KEY_LOOKUP_SECRET, removing ~67ms of CPU cost from every unauthenticated request. Anonymous /auth/* routes gain a per-IP rate limit, and password hashing no longer blocks the event loop.

Breaking changes

API keys issued before v7.2 stop working. There is no PBKDF2 fallback — a legacy fallback would re-introduce the denial-of-service the migration fixes. Each key must be re-issued once; the old raw key is never required:

# Before (pre-v7.2 key): now rejected
curl -X POST http://localhost:8000/verify/math \
  -H "X-API-Key: qwed_live_old_key" \
  -H "Content-Type: application/json" \
  -d '{"expression": "1 + 1 = 2"}'
# -> 401 Unauthorized

# After: sign in with email/password (no API key needed), then mint a new key
curl -X POST http://localhost:8000/auth/signin \
  -H "Content-Type: application/json" \
  -d '{"email": "user@example.com", "password": "..."}'

curl -X POST http://localhost:8000/auth/api-keys \
  -H "Authorization: Bearer eyJhbG..." \
  -H "Content-Type: application/json" \
  -d '{"name": "Production Key"}'

Keys can also be rotated by key ID through /admin/keys/rotate using any already-working key. Newly issued and rotated keys are HMAC digests automatically.

What changed

HMAC-SHA256 API-key lookup. The previous PBKDF2 digest (100,000 iterations, ~67ms) ran on every request carrying an x-api-key header, valid or not, letting ~15 garbage requests per second saturate the service. API keys are high-entropy random tokens, so the KDF cost bought no brute-force resistance. The replacement keyed MAC costs microseconds.
Dedicated lookup secret (required). Digests are keyed by QWED_API_KEY_LOOKUP_SECRET. Self-hosted servers fail closed at startup if it is missing or equal to QWED_JWT_SECRET_KEY — reusing the JWT secret would silently break every API-key lookup on the next JWT-secret rotation. Set it before issuing v7.2 keys; changing it later requires a one-time re-issue.
Per-IP rate limit on /auth/*. Anonymous auth routes (signup, signin) are throttled per client IP: QWED_RATE_LIMIT_PER_IP, default 10 requests per minute. Over-limit requests get 429 Too Many Requests with a Retry-After header of at least 1 second. X-Forwarded-For is honored only when the direct peer is listed in QWED_AUTH_TRUSTED_PROXIES (comma-separated CIDRs, default empty), and only the rightmost hop is used, so clients cannot choose their own bucket. The IP table is hard-bounded against spoofed-address floods.
Password hashing off the event loop. Signup and signin bcrypt calls run in a worker thread instead of blocking the server. Signin also performs a dummy verify on unknown emails, so response timing no longer reveals which addresses are registered.
Atomic signup. The organization and user rows commit in one transaction; a failure mid-signup rolls back both instead of stranding an orphaned organization.

What this means for you

Re-issue any API key created before v7.2. Self-hosted operators must add QWED_API_KEY_LOOKUP_SECRET to their environment (distinct from QWED_JWT_SECRET_KEY) and, if running behind a proxy, set QWED_AUTH_TRUSTED_PROXIES so per-IP throttling keys on real client addresses. See Authentication for the migration path and Rate limits for the throttle contract.

QWED-Tax — npm verifier enforces structured nexus claims fail-closed

Released: September 1, 2026 · GitHub PR #65

The @qwed-ai/tax npm verifier now requires the structured boolean claimed_collects_tax claim for economic-nexus checks, matching the Python guard’s contract. Unknown states, non-boolean claims, and malformed sales facts fail closed instead of passing as no-nexus.

What changed

Structured claim required — TaxPreFlight.audit and NexusGuard.checkNexus verify the boolean claimed_collects_tax against the independently computed nexus. Non-boolean values (including strings like "false") are rejected with "Computed nexus liability only. Provide claimed_collects_tax as a boolean for deterministic verification.".
Legacy fallback narrowed — tax_decision: 'no_tax' maps to claimed_collects_tax: false only when the intent does not contain a claimed_collects_tax key. A present-but-invalid structured claim is blocked even when the legacy string is also set.
Unknown states fail closed — states not in the npm threshold table (currently NY, CA, TX, FL) return verified: false with a block-pending-configuration error instead of being verified as no-nexus.
Input hardening — nexus validation runs whenever sales_data is present (including falsy or explicitly undefined values). Non-string state, non-finite amounts (NaN, Infinity), and negative sales or transaction counts are rejected before threshold math.

What this means for you

If your frontend or Node integration relied on unknown states passing, on truthy strings as claims, or on omitting sales_data fields to skip the nexus check, those intents now block. Pass claimed_collects_tax as an explicit boolean and treat unknown-state blocks as hold-and-escalate signals. See Economic nexus in the TypeScript SDK for the full contract.

QWED-Tax — explicit boolean claim required for economic nexus verification

Released: August 31, 2026 · GitHub PR #62

Economic nexus verification now fails closed unless the caller provides an explicit claimed_collects_tax boolean. Free-form tax_decision / llm_decision text is no longer interpreted as a verification claim — free-form model output cannot be a verification substrate.

What changed

NexusGuard.check_nexus_liability — Gains a keyword-only claimed_collects_tax: bool parameter. The guard computes the state’s threshold independently and verifies the boolean against the computed nexus. When the parameter is omitted, the guard returns {"verified": False, "computed_only": True, "has_nexus": <bool>, "error": "Computed nexus liability only. Provide claimed_collects_tax as a boolean for deterministic verification."}. Non-boolean values (including truthy strings) return verified=False with an invalid-claim error.
llm_decision deprecated — The parameter is retained for positional compatibility but is never parsed. Passing it has no effect on the verdict.
TaxPreFlight economic_nexus intents — The required claim field changed from tax_decision (string) to claimed_collects_tax (boolean). Intents that still send tax_decision are blocked as incomplete claims before NexusGuard runs.

Breaking changes

Callers that previously passed a free-form decision string — check_nexus_liability(state, sales, count, llm_decision="No tax needed") or a TaxPreFlight intent with "tax_decision": "No tax needed" — now receive a fail-closed result instead of a verified verdict. Translate the AI’s decision into an explicit boolean before calling:

# Before
guard.check_nexus_liability("NY", 600000, 10, llm_decision="No tax needed")

# After
guard.check_nexus_liability("NY", 600000, 10, claimed_collects_tax=False)

What this means for you

An LLM can no longer “verify” its own nexus decision by phrasing it in text — the claim must be a machine-checkable boolean that the guard compares against the deterministically computed threshold. See NexusGuard for the full parameter contract and TaxPreFlight integration for the intent migration.

QWED Open Responses — fail-closed verification and expanded guard coverage

Released: August 30, 2026 · GitHub PR #33

verify() now fails closed when no guards are configured, ToolGuard recognizes and polices more tool-call envelope shapes, and SafetyGuard scans nested content. All three changes ship in both the Python and TypeScript packages.

Breaking behavior change. verify() with zero guards now returns verified=False with the guard result message “No guards configured”, and blocked=True in strict mode. Previously it returned verified=True because no guard failed. If you use ResponseVerifier() or VerifiedOpenAI without configuring guards, those calls now block: set default_guards (Python) / defaultGuards (TypeScript) or pass guards per call.

What changed

Zero-guard verify fails closed. Absence of verification is not success. A verify() call with no guards returns a ResponseVerifier guard result with severity="error" and the block reason “No guards configured — fail-closed (zero-guard verify).”
ToolGuard envelope coverage. ToolGuard now extracts tool calls from Anthropic content[].type=tool_use blocks and matches type values case-insensitively, so a Tool_Use block is policed instead of ignored. OpenAI function wrappers and JSON-encoded argument strings are normalized before blocklist and dangerous-pattern checks.
ToolGuard fail-closed rejections. Tool-like content in an unrecognized shape, non-object entries in tool_calls/choices/content, hybrid envelopes that mix a direct tool call with a sibling collection, nameless tool calls, and argument payloads over 10,000 characters or 128 nesting levels are all blocked instead of passing silently.
SafetyGuard recursive extraction. SafetyGuard scans string content nested up to 12 levels deep, including the canonical choices[].message.content shape, so PII, injection, and harmful patterns inside nested structures are detected.

What this means for you

If your agent relied on verify() passing with no guards, or on responses in unrecognized tool-call shapes passing as “no tool calls”, those calls now block. Configure the guards you need and emit tool calls in a recognized envelope. See the guards reference and troubleshooting for the updated contracts.

Released: August 31, 2026 · GitHub PR #62

Economic nexus verification now fails closed unless the caller provides an explicit claimed_collects_tax boolean. Free-form tax_decision / llm_decision text is no longer interpreted as a verification claim — free-form model output cannot be a verification substrate.

What changed

NexusGuard.check_nexus_liability — Gains a keyword-only claimed_collects_tax: bool parameter. The guard computes the state’s threshold independently and verifies the boolean against the computed nexus. When the parameter is omitted, the guard returns {"verified": False, "computed_only": True, "has_nexus": <bool>, "error": "Computed nexus liability only. Provide claimed_collects_tax as a boolean for deterministic verification."}. Non-boolean values (including truthy strings) return verified=False with an invalid-claim error.
llm_decision deprecated — The parameter is retained for positional compatibility but is never parsed. Passing it has no effect on the verdict.
TaxPreFlight economic_nexus intents — The required claim field changed from tax_decision (string) to claimed_collects_tax (boolean). Intents that still send tax_decision are blocked as incomplete claims before NexusGuard runs.

Breaking changes

# Before
guard.check_nexus_liability("NY", 600000, 10, llm_decision="No tax needed")

# After
guard.check_nexus_liability("NY", 600000, 10, claimed_collects_tax=False)

What this means for you

QWED-MCP v0.2.2 — hardened math expression sandbox (security release)

Released: August 27, 2026 · GitHub PR #48 · GHSA-2p69-jpm6-jrxh ↗

qwed-mcp v0.2.2 fixes GHSA-2p69-jpm6-jrxh (CWE-94), a residual sandbox-escape bypass in the math expression parser. The previous regex denylist could be evaded through unlisted dunder names, string-literal splitting, and NFKC-equivalent Unicode identifiers. The parser now validates expressions structurally before any evaluation.

What changed

The math expression sandbox in qwed-mcp now applies layered validation before an expression reaches SymPy’s parse_expr:

AST node allowlist. Only arithmetic node types survive (binary and unary operations, calls, names, numeric constants). The allowlist rejects attribute access, subscripts, lambdas, and comprehensions before parsing. Every sandbox-escape traversal needs an attribute or subscript node; legitimate math never does.
String and bytes literals rejected. The validator fails expressions containing string or bytes constants with SafeParserError, closing the concatenation-gadget class (for example '__glo'+'bals'+'__').
NFKC normalization before validation. CPython NFKC-normalizes identifiers at compile time, so Unicode look-alike characters (mathematical bold letters, fullwidth forms) could previously bypass the ASCII denylist. The sandbox now normalizes input first, so the checks see exactly what the compiler sees.
Implicit-multiplication charset guard. Inputs that are not valid Python (such as 2x or sin x) cannot pass AST checks, so the guard restricts their character set instead: it bans quotes, brackets, and separators, and allows . only as a decimal point. 2x, sin x, x^2, 2.5x, and 2(x+1) still parse; the guard blocks 2x.__class__.
Regex denylist retained as defense in depth.

What this means for you

Upgrade to qwed-mcp 0.2.2 or later:

pip install --upgrade qwed-mcp

No API changes. Legitimate mathematical expressions parse exactly as before. Expressions containing attribute access, subscripts, string literals, or other non-arithmetic syntax now fail with a SafeParserError instead of reaching the evaluator. See the MCP tools reference for the current tool surface.

QWED-Infra v0.3.0 — Verification Context v1.0 and attestation trust boundary

Released: August 24, 2026

Every qwed-infra guard now emits a portable Verification Context v1.0 document via to_verification_context(), and ADMIT decisions now require a cryptographically valid ES256 attestation bound to the exact claim and evidence.

Added

to_verification_context() on all four guards — IamGuard (#45), NetworkGuard (#46), CostGuard (#48), and ArtifactBoundaryGuard (#50) each produce a schema-valid, tamper-evident VC document with claim, verifier identity, sha256-bound evidence, and an ADMIT/DENY admission decision, built on the shared bridge module from PR #44.
Attestation trust boundary (PR #54) — ES256 (ECDSA P-256) JWT attestation service with a never-None fail-closed AttestationResult contract, revocation registry, and a single consumption-side gate validating signature, issuer, expiry, and revocation plus claim bindings (status match, query_hash == sha256(formal_statement), proof_hash == proof_ref).
mint_diagnostic_attestation() — issues a token bound to a VERIFIED diagnostic’s own evidence commitment.

Breaking changes

Signature migration (pre-1.0 API). Guard adapters no longer accept pre-computed result objects. to_verification_context() takes raw verification inputs and runs the guard’s own deterministic solver internally. A result-accepting signature is forgeable, so it was removed before release. Pass raw inputs: NetworkGuard.to_verification_context(resources, source, destination, port, ...), IamGuard.to_verification_context(policy, action, resource, context=None, ...), CostGuard.to_verification_context(resources, budget_monthly, ...), ArtifactBoundaryGuard.to_verification_context(package_dir, ...).

Admission semantics — VERIFIED results admit only with a cryptographically valid attestation bound to the exact claim and evidence. Arbitrary non-empty attestation strings are rejected as forged tokens and produce BLOCKED. A missing token demotes VERIFIED to UNVERIFIABLE/DENY. Callers previously passing placeholder tokens must mint via create_verification_attestation() or mint_diagnostic_attestation().
New runtime dependencies — pyjwt and cryptography for attestation signing and validation.

Fixed

Fail-closed on malformed inputs at every VC boundary (PR #51 and follow-ups) — undecimal budgets, non-string build backends, malformed topology, policy, or package inputs, symlink escapes and loops, and wheel entries outside the scanned boundary all map to BLOCKED/DENY documents instead of exceptions or guessed approval.

See the guards reference and usage examples for the updated contracts.

v7.1.0 — Verification Context v1.0 Rollout

Released: August 16, 2026 · GitHub Release ↗ · GitHub PR #317

v7.1.0 ships Verification Context v1.0 end-to-end: a formal specification with a machine-readable JSON Schema, a typed document model with fail-closed invariants, public proof_ref generation and resolution, and exposure across the API, Python SDK, CLI, verifiers, and the Docker GitHub Action. The release is additive — a semver minor with no breaking wire changes.

Spec and core model

Verification Context v1.0 specification — the atomic JSON record of a verification: verified object, four context layers (interpretation / proof / evidence / decision), verdict, and admission, with canonical RFC 8785 encoding for the proof_ref evidence commitment (PR #302). See the Verification Context specification.
VerificationContextDocument model + schema validation — typed VerificationContext, VerificationContextDocument, Verdict, Admission, and nested layer types with fail-closed invariants enforced at construction: VERIFIED requires a sha256:<64-hex> proof_ref; UNVERIFIABLE/BLOCKED require proof_ref: null with admission DENY (PR #308).
Public proof_ref generation and resolution — compute_document_proof_ref() and resolve_document_proof_ref() derive and verify the content-bound SHA-256 commitment over the canonical document. Resolution fails closed: a proof_ref that cannot be resolved confers no authority (PR #309).

Surface exposure

API endpoints — POST /verification-context/from-diagnostic, /validate, and /resolve convert DiagnosticResult records into VC documents, validate documents against the schema, and resolve proof_ref commitments (PR #311).
CLI commands — the qwed context group: validate, resolve, and from-diagnostic (PR #311).
SDK re-exports — all VC v1.0 types and helpers importable directly from qwed_sdk, plus create_verification_context_from_diagnostic(), validate_verification_context(), and resolve_verification_context() on both sync and async clients. See the Python SDK (PRs #311, #315).
to_verification_context() across the verifier classes — implemented on the Code, Consensus, DSL Logic, Fact, Graph Fact, Image, Logic, Reasoning, Schema, SQL, Stats, and Symbolic verifiers plus the BatchFactVerifier and MultiVLMVerifier helpers, each mapping a DiagnosticResult to a VC document. Math and the secure code executor have no such method; their results convert through POST /verification-context/from-diagnostic (PRs #310, #316).
Docker action VC outputs — the GitHub Action emits verdict, admission, proof_ref, and verification_context outputs in every mode (PR #313).

Fail-closed behavior

A VERIFIED diagnostic without an attestation token fails closed to BLOCKED (trust_gate.mandatory_attestation_missing) when converted to a VC document — attestation is enforced through the same trust boundary as /verify/*.
Malformed diagnostic payloads convert to BLOCKED documents instead of crashing, so the audit trail records the failure.
resolve returns true only for a schema-valid VERIFIED document whose stored proof_ref matches the re-derived commitment.

Version bumps

Surface	From	To
`qwed` (PyPI)	`7.0.0`	`7.1.0`
`qwed_sdk` (Python)	`7.0.0`	`7.1.0`
`@qwed-ai/sdk` (npm)	`7.0.0`	`7.1.0`
`qwed` (crates.io)	`7.0.0`	`7.1.0`
API version marker	`7.0.0`	`7.1.0`

Included PRs

PR	Summary
#302	Verification Context Specification v1.0 + JSON Schema
#308	`VerificationContext` model and schema validation helpers
#309	Public `proof_ref` generation and resolution
#310	`StatsVerifier` `DiagnosticResult` → Verification Context mapping
#311	Verification Context exposed across SDK, API, and CLI
#313	Docker action emits VC v1.0 outputs
#315	VC v1.0 types re-exported from `qwed_sdk`
#316	`to_verification_context()` on all remaining verifiers
#317	v7.1.0 release preparation

Upgrading

No action required. Every change is additive: existing /verify/* responses, SDK methods, CLI commands, and action outputs are unchanged. Adopt Verification Context by calling the new endpoints, importing the new types from qwed_sdk, or reading the new action outputs.

v7.0.0 — Full DiagnosticResult engine conformance

Released: August 8, 2026 · GitHub Release ↗ · GitHub PR #300

v7.0.0 completes the engine migration to the unified DiagnosticResult contract (META #216). SchemaVerifier, SQLVerifier, CodeVerifier, SecureCodeExecutor, and StatsVerifier — plus the fact and image batch entry points — now all return DiagnosticResult (status / agent_message / developer_fields / proof_ref). The release also completes the truth-vs-admission separation: verification answers “is this claim provably true?” while a separate AdmissionDecision answers “should this be allowed at this boundary?”.

Two breaking wire changes. POST /verify/code now returns status: "VERIFIED" for proven-unsafe code, with admission driven by the new admission field. POST /verify/stats now returns status: "UNVERIFIABLE" on successful execution instead of the legacy SUCCESS shape — execution is not verification. Migration examples below.

Breaking: `POST /verify/code` — `VERIFIED` is truth, `admission` is policy

Proving a snippet is unsafe is a successful proof. Unsafe code is therefore VERIFIED (previously BLOCKED), with developer_fields.is_valid: false, a bound proof_ref, and a non-null critical_count. BLOCKED is reserved for cases where verification itself failed (empty code, non-string language, internal errors), and blocked results carry no proof_ref. The response attaches an explicit admission field (ADMIT / BLOCKED) so authority-only consumers reading status == "VERIFIED" cannot admit unsafe code. Before (v6.x) — unsafe code:

{
  "status": "BLOCKED",
  "agent_message": "The code failed security verification and is not safe to use.",
  "developer_fields": { "is_safe": false },
  "proof_ref": null
}

After (v7.0.0) — unsafe code:

{
  "status": "VERIFIED",
  "agent_message": "The code failed security verification and is not safe to use.",
  "developer_fields": {
    "constraint_id": "code_verifier.code_unsafe",
    "is_safe": false,
    "is_valid": false,
    "critical_count": 1,
    "issues": [ ... ]
  },
  "proof_ref": "sha256:...",
  "is_authoritative": true,
  "admission": "BLOCKED"
}

Migrating: consumers that branched on status == "BLOCKED" or status == "VERIFIED" for safety gating must switch to the admission field or developer_fields.is_valid. status == "VERIFIED" alone must never be treated as “safe to execute”.

# Before (v6.x)
if response["status"] == "VERIFIED":
    execute(code)

# After (v7.0.0)
if response["admission"] == "ADMIT":
    execute(code)

Breaking: `POST /verify/stats` — execution success is never `VERIFIED`

A run that executes cleanly in the Docker sandbox and returns an observed statistic is UNVERIFIABLE (stats_verifier.claim_not_verified) — the engine has no deterministic claim-proof, so it cannot attest the original natural-language claim. Execution evidence (observed_result, generated_code, columns, a deterministic dataset_sha256, sandbox type, timing, and security checks) is retained in developer_fields for audit. BLOCKED is reserved for failure states: stats_verifier.validation_error, stats_verifier.execution_failure, and stats_verifier.runtime_unavailable. Before (v6.x) — successful execution:

{
  "status": "SUCCESS",
  "result": 52340.5,
  "code": "df['salary'].mean()"
}

After (v7.0.0) — successful execution:

{
  "status": "UNVERIFIABLE",
  "agent_message": "Statistical analysis completed, but the claim could not be deterministically verified.",
  "developer_fields": {
    "constraint_id": "stats_verifier.claim_not_verified",
    "is_valid": false,
    "observed_result": 52340.5,
    "generated_code": "df['salary'].mean()",
    "dataset_sha256": "9f2c...",
    "sandbox_type": "docker"
  },
  "proof_ref": null,
  "is_authoritative": false
}

Migrating: read the computed value from developer_fields.observed_result instead of result, and treat the verdict as advisory — a legitimate execution is not a proven claim. compute_statistics() and get_sandbox_info() are utilities, not verification boundaries, and keep their existing dict return shape.

`SchemaVerifier` → `DiagnosticResult` (#294)

verify() and verify_ucp_transaction() return DiagnosticResult. A schema violation is VERIFIED (as-invalid) with developer_fields.is_valid: false; BLOCKED (schema_verifier.parse_error / schema_verifier.validation_error) is reserved for schemas that cannot be parsed or validated.
proof_ref is computed deterministically from canonical JSON of the schema + instance evidence; unsupported values and cyclic structures fail closed to BLOCKED.
Recursive schema meta-validation: malformed keyword shapes (non-dict properties, invalid required entries, invalid or non-finite numeric constraints, negative size constraints) return BLOCKED instead of being silently treated as empty. Oversized integer bounds (e.g. 10**1000) no longer raise OverflowError.
UCP hardening: money arithmetic uses Decimal quantized to the currency precision (no more 0.01 float tolerance), tax is selected by key presence so tax: 0 is honored, verdict fields (transaction_type, currency, schema_verifier.ucp_* constraint ids) are complete on every path, and string/None amounts no longer raise.
agent_message is sanitized — no rule IDs, issue types, or schema internals leak into agent-facing output.

See the updated Schema verifier page.

`SQLVerifier` → `DiagnosticResult` (#295)

verify_sql() returns DiagnosticResult. A safe query is VERIFIED (sql_verifier.sql_valid, is_valid: true, proof_ref from the AST). A proven-malicious query is VERIFIED-as-malicious (sql_verifier.malicious, is_valid: false, malicious_classification: true) — proving malice is a successful proof, so it retains its proof_ref.
BLOCKED is reserved for incomplete or failed analysis: sql_verifier.parse_error, sql_verifier.schema_parse_error (takes precedence over malice detection — no authoritative proof from incomplete analysis), sql_verifier.complexity_limit_exceeded, and sql_verifier.execution_error.
POST /verify/sql attaches the admission field (ADMIT / BLOCKED) alongside the verdict.

See the updated SQL engine page.

`CodeVerifier` & `SecureCodeExecutor` → `DiagnosticResult` (#296)

verify_code(), verify_python_deep(), and verify_batch() return DiagnosticResult with the truth-vs-admission semantics described above.
verify_batch() returns per-item verdicts plus a summary (safe / unsafe / blocked counts, total_critical) and an overall is_valid that is true only when all snippets are safe — the batch is otherwise non-admissible.
SecureCodeExecutor.execute() no longer executes code wholesale on the verifier verdict: an unconditional OWASP LLM06 dangerous-pattern gate blocks execution with CONSTRAINT_DANGEROUS_PATTERN. The scan is AST-aware, so dangerous keywords appearing only in comments, docstrings, or string literals do not cause false denials.
ConsensusVerifier and StatsVerifier code stages now require is_verified and developer_fields.is_valid is True, so consensus results can no longer admit unsafe code.

See the updated Code engine page.

`StatsVerifier` → `DiagnosticResult` (#297)

verify_stats() returns DiagnosticResult with the execution-is-not-verification semantics described above.
The API boundary is a thin pass-through: POST /verify/stats forwards the engine’s DiagnosticResult through enforce_trust_decision() unchanged.
Logging is fail-closed and claim-aware: a non-authoritative result (BLOCKED / UNVERIFIABLE, proof_ref: null) can never be persisted as verified, even if mutable developer_fields.is_valid metadata is true.
A non-serializable sandbox result (e.g. a DataFrame) is coerced to a JSON-safe value before entering developer_fields, so a legitimate UNVERIFIABLE verdict is not silently downgraded to BLOCKED.

See the updated Stats engine page.

Fact & image batch verification fail closed (#297)

BatchFactVerifier.verify_batch() and ImageVerifier.verify_batch() — the last two public engine entry points returning ad-hoc dicts — now return a single DiagnosticResult with per-claim verdicts in developer_fields.results and a summary:

The batch is authoritative (VERIFIED + proof_ref) only when every claim is deterministically verified.
Any refuted or blocked claim fails the whole batch closed (fact_verifier.batch_blocked / image_verifier.batch_blocked).
An empty batch is BLOCKED (*.empty_batch).
The batch proof_ref binds full claim digests and the shared input (image digest for image batches, context digest for fact batches), never truncated display text.
Aggregation is shared via diagnostics.aggregate_batch_diagnostic() so the fail-closed logic cannot drift between engines.

Version bumps

Surface	From	To
`qwed` (PyPI)	`6.0.0`	`7.0.0`
`qwed_sdk` (Python)	`6.0.0`	`7.0.0`
`@qwed-ai/sdk` (npm)	`6.0.0`	`7.0.0`
`qwed` (crates.io)	`6.0.0`	`7.0.0`
API version marker	`6.0.0`	`7.0.0`

Included PRs

PR	Summary
#294	`SchemaVerifier` → `DiagnosticResult`
#295	`SQLVerifier` → `DiagnosticResult`
#296	`CodeVerifier` and `SecureCodeExecutor` → `DiagnosticResult`
#297	`StatsVerifier.verify_stats` → `DiagnosticResult` (+ fact/image batch)
#300	Release: v7.0.0 — Full DiagnosticResult Engine Conformance

Upgrading

If you only consume the high-level SDK clients and gate on admission / developer_fields.is_valid, bump the version and you are done. If your code branches on status from POST /verify/code for safety gating, or parses the legacy SUCCESS / ERROR shape from POST /verify/stats, apply the migrations shown above before upgrading.

v6.0.0 — Trust Boundary Completion

Released: August 2, 2026 · GitHub Release ↗ · GitHub PR #291

v6.0.0 closes the Trust Boundary Completion epic (Issue #263, 12/12 sub-issues). Every verification pathway now returns a unified DiagnosticResult and routes through enforce_trust_decision. The trust boundary is no longer advisory: the control plane requires and verifies attestation before admitting VERIFIED results, and VERIFIED is a protocol guarantee backed by a deterministic proof_ref — never by execution, agreement, confidence, or provenance.

Breaking change. /verify/* API responses now use the unified DiagnosticResult schema (status / agent_message / developer_fields / proof_ref). Consumers that parsed the previous ad-hoc dict responses must migrate to the unified 3-layer contract. The high-level SDK clients (qwed, qwed_sdk, @qwed-ai/sdk, and the qwed Rust crate) already surface DiagnosticResult at v6.0.0 and require no code change beyond the version bump.

Architecture: observation vs. admission

The API is the observation surface — an honest witness that returns what verification found. The control plane is the admission authority — the judge that decides what is admitted as VERIFIED. QWED_RULES.md now codifies this separation (#13 Separation of Responsibilities, #14 Verification Semantics, #15 Truth Before Policy; rules #7/#8 updated for admission-boundary and deterministic-proof semantics).

All /verify/* endpoints return DiagnosticResult — unified response contract across every verification surface (PR #276)
Control plane enforces mandatory attestation — require_attestation=True, attestation issued and verified at the admission boundary, and the enforced status drives the HTTP response status (PR #278). See Cryptographic attestations.
Batch math routes through the trust boundary — /verify/batch results carry DiagnosticResult, proof_ref, and attestation, and pass through enforce_trust_decision (PR #282). See POST /verify/batch.
Attestation scope alignment — attestations bind to the translated expression, not the natural-language query, so query_hash binds to what was actually verified (PR #285).

`VERIFIED` is a protocol guarantee

No engine emits VERIFIED without a deterministic proof_ref. Heuristic and advisory analysis now reports UNVERIFIABLE with structured advisory_checks instead of masquerading as verified.

ConsensusResult uses the DiagnosticStatus enum with proof_ref and verified_evidence (PR #280). See Consensus engine.
FactVerifier heuristic SUPPORTED verdict → UNVERIFIABLE with advisory_checks (PR #283).
Consensus code execution advisory-only — VERIFIED → UNVERIFIABLE (PR #281).
Consensus stats advisory-only — never VERIFIED (PR #277).
LogicVerifier migrated to DiagnosticResult (PR #262 — details in the entry below).
AgentStateGuard proof_ref is now a real sha256 of committed bytes, not a static sentence (PR #284). See Agent state guard.

Engineering and security hardening

TOCTOU closure in enforce_trust_decision — developer_fields snapshotted via recursive rebuild (no deepcopy alias window); fail-closed snapshot (PR #290)
Attestation signature verified before claim decode — silent generic error for every failure mode (PR #287)
Tenant-isolated verification cache — VerificationCache keys namespaced by normalized tenant_id (PR #286)
Unicode normalization in AgentStateGuard canonicalization — NFC collisions rejected (PR #288)
Mandatory proof artifact for VERIFIED attestations at both issuance and consumption (PR #248)
Credential / JWT / dockerignore security alerts resolved (PR #249)
Math whitelist injection bypass removed (PR #251)
Engine classification docs — Proof / Policy Enforcement / Advisory (PR #247)

Version bumps

Every SDK, container, and manifest ships at 6.0.0:

Surface	From	To
`qwed` (PyPI)	`5.3.0`	`6.0.0`
`qwed_sdk` (Python)	`5.3.0`	`6.0.0`
`@qwed-ai/sdk` (npm)	`5.3.0`	`6.0.0`
`qwed` (crates.io)	`5.3.0`	`6.0.0`
Docker image `qwedai/qwed-verification`	`5.3.0`	`6.0.0`
Kubernetes deployment image tag	`5.3.0`	`6.0.0`
API version marker	`5.3.0`	`6.0.0`

Included PRs

The following 21 PRs shipped in the v6.0.0 release:

PR	Summary
#247	docs: engine classification — Proof / Policy Enforcement / Advisory
#248	Enforce mandatory proof artifact on `VERIFIED` attestations (issuance + consumption)
#249	Resolve credential / JWT / dockerignore security alerts
#251	Remove math whitelist injection bypass
#260	Hybrid engine advisory-only — never `VERIFIED` without proof
#261	`FactVerifier` advisory-only
#262	`LogicVerifier` migrated to `DiagnosticResult`
#276	Migrate all `/verify/*` endpoints to return `DiagnosticResult`
#277	Consensus stats advisory-only, never `VERIFIED`
#278	Control plane trust enforcement mandatory
#280	`ConsensusResult` `DiagnosticStatus` enum + `proof_ref` + `verified_evidence`
#281	Consensus code execution advisory-only
#282	Batch math `DiagnosticResult` → `proof_ref` + attestation + `enforce_trust_decision`
#283	`FactVerifier` `SUPPORTED` → `UNVERIFIABLE` with heuristic `advisory_checks`
#284	`AgentStateGuard` `proof_ref` real `sha256`
#285	Attest translated expression, not natural-language query
#286	`VerificationCache` tenant isolation
#287	Attestation verify-before-decode + silent generic error
#288	NFC-normalize `AgentStateGuard` canonicalization
#289	Mock network in secret redaction tests (CI)
#290	Close TOCTOU in `enforce_trust_decision`

Upgrading

If you are on v5.3.x and only consume the high-level SDK clients or attestations, bump the version and you are done — every engine already returns DiagnosticResult. If your code parses raw HTTP responses from /verify/*, migrate to the 3-layer DiagnosticResult shape before upgrading. Direct callers of LogicVerifier should also review the LogicVerifier migration notes below.

v6.0.0 — LogicVerifier conforms to `DiagnosticResult`

Released: August 2, 2026 · GitHub PR #262 · part of the v6.0.0 release

LogicVerifier is the second engine — after FactVerifier — to conform to the unified 3-layer DiagnosticResult contract introduced in v5.2.0. All nine public methods now return a DiagnosticResult with .status, .developer_fields, .is_verified, .agent_message, and .proof_ref. The legacy LogicResult dataclass has been removed. Two fail-closed input-strictness fixes ship with it.

Breaking change for direct LogicVerifier callers. The LogicResult dataclass is gone. Code that reads result.status == "SAT", result.model, result.error, result.proof_summary, or result.explanation must migrate to result.is_verified and result.developer_fields. The high-level SDK client (client.verify_logic()) and the /verify/logic API endpoint response shape are unchanged.

What changed

All nine methods return DiagnosticResult. verify_logic, verify_with_quantifiers, verify_bitvector, verify_array, prove_theorem, check_implication, check_equivalence, verify_optimization, and check_vacuity all now return a DiagnosticResult with the three diagnostic layers populated.
Per-method SAT/UNSAT disambiguation. Z3’s sat/unsat outcome is mapped to a DiagnosticStatus per method (for example, verify_logic sat → VERIFIED, but prove_theorem sat → BLOCKED with a counterexample; prove_theorem unsat → VERIFIED because the theorem was proved by contradiction). The raw solver verdict is preserved on developer_fields["deterministic_verdict"].
symbol_table on every VERIFIED result. developer_fields["symbol_table"] is a sorted list of {"name": ..., "type": ...} entries for every declared variable, so audit logs capture exactly what was proven.
proof_ref from the Z3 assertion stack. Every VERIFIED result carries a deterministic sha256:... proof_ref computed from the solver’s assertions — the authority bit that downstream gates use for admissibility.
Empty variables dict now fails closed. The _infer_variables type-guessing heuristic has been removed. Calls with an empty variables dict return BLOCKED with constraint_id = "logic_verifier.explicit_declarations_required". Declare every variable explicitly.
Malformed BitVec[N] declarations now fail closed. Type strings such as "BitVec", "BitVec[]", or "BitVec[abc]" no longer silently default to 32 bits. They return BLOCKED with constraint_id = "dsl_compiler.type_validation".
agent_message is sanitized. No raw Z3 output leaks into the agent-facing layer. Structured diagnostic details live under developer_fields.

Before and after

Reading a satisfiability result

# Before — LogicResult with a solver-level status string
result = verifier.verify_logic({"x": "Int"}, ["x > 0"])
if result.status == "SAT":
    solution = result.model

# After — DiagnosticResult
result = verifier.verify_logic({"x": "Int"}, ["x > 0"])
if result.is_verified:
    solution = result.developer_fields["model"]
    verdict  = result.developer_fields["deterministic_verdict"]  # "SAT"

Proving a theorem

# Before — SAT meant "theorem is valid"
result = verifier.prove_theorem(variables, premises, conclusion)
if result.status == "SAT":
    ...  # theorem valid
elif result.status == "UNSAT":
    counterexample = result.model

# After — VERIFIED means "theorem proved", BLOCKED means "counterexample found"
result = verifier.prove_theorem(variables, premises, conclusion)
if result.is_verified:
    ...  # theorem proved by contradiction
elif result.status.value == "BLOCKED":
    counterexample = result.developer_fields["model"]

Empty variables dict

# Before — types were guessed from constraint syntax
verifier.verify_logic({}, ["x > 5", "P and Q"])
# Inferred: x -> Int, P/Q -> Bool

# After — BLOCKED
result = verifier.verify_logic({}, ["x > 5", "P and Q"])
result.developer_fields["constraint_id"]
# "logic_verifier.explicit_declarations_required"

What this means for you

Callers that already treat non-VERIFIED results as unverified continue to work — they just need to migrate the field names (result.is_verified in place of result.status == "SAT", result.developer_fields["model"] in place of result.model). Callers that relied on implicit variable inference or on "BitVec" defaulting to 32-bit must now declare every variable explicitly. See LogicVerifier returns DiagnosticResult for the full status matrix, constraint_id list, and migration snippets, and the Verification Diagnostics guide for the 3-layer model.

QWED Control Plane â€” mandatory attestation admission on `/verify/math`, translated-expression attestation scope

Released: July 31, 2026 Â· GitHub PR #278 Â· GitHub PR #285

The control plane now treats attestation as an admission gate on the math verification path â€” the final response status is set by the enforcement step after signing, not by the raw verifier verdict. In the same release, the attestation qwed.query_hash now binds to the translated deterministic expression, not to the user’s natural-language query.

What changed

enforce_trust_decision(..., require_attestation=True) is now the single source of truth for the math response status and trust_boundary.overall_status. Both fields are set from the enforced decision after attestation is issued, never from the raw verifier verdict.
A VERIFIED math result is only surfaced once create_verification_attestation() returns ISSUED. Any attestation signing failure downgrades the response to BLOCKED â€” the error code is echoed under trust_boundary.attestation_error and no token is returned.
trust_boundary.attestation_policy is now always "mandatory". The previous advisory mode has been removed.
The attestation qwed.query_hash binds to the translated expression the engine actually evaluated, not to the natural-language query. The natural-language query is still returned in response.user_query for display, and trust_boundary.verification_scope = "translated_expression_only" continues to disclose the narrowed scope.

What this means for you

Downstream consumers verifying a QWED attestation must hash the translated expression when re-checking qwed.query_hash. Hashing user_query will not match â€” that is intentional, because QWED does not attest to the LLM translation step.
If your integration branched on trust_boundary.overall_status, it now reflects the post-attestation decision. Signing outages surface as BLOCKED with attestation_error, not as a raw VERIFIED verdict.
See Trust boundary and POST /verify/natural_language for the full response contract, and Cryptographic attestations for the AttestationResult fail-closed contract.

v5.3.0 — SymbolicVerifier: `DiagnosticResult` reference implementation

Released: July 25, 2026 · GitHub Release · Minor · PR #244

QWED-Verification v5.3.0 makes SymbolicVerifier the first fully DiagnosticResult-conformant engine. The unified 3-layer diagnostic model introduced in v5.2.0 is no longer aspirational — the Code engine is the reference implementation the remaining engines will migrate to.

Breaking change for the Code engine. All six SymbolicVerifier public methods now return a DiagnosticResult instead of a dict. Callers that read result["status"], result.verified, result["issues"], result["complexity"], result["loop_depth"], or result["recursive"] must migrate. See the Code engine migration guide for the field-by-field mapping.

What changed

Six methods migrated to DiagnosticResult — verify_code, verify_function_contract, verify_safety_properties, verify_bounded, analyze_complexity, and get_verification_budget now return the unified 3-layer diagnostic type. Every result carries status (DiagnosticStatus), agent_message (Layer 1), developer_fields (Layer 2), and proof_ref (Layer 3, always None for this engine).
verification_mode on every result — developer_fields["verification_mode"] is "symbolic" for standard runs and "bounded_symbolic" for verify_bounded(), so callers can distinguish the two verification modes without inspecting the call site.
VERIFIED is intentionally never emitted — CrossHair’s search is timeout-bounded, not a completeness proof, so a clean run maps to UNVERIFIABLE with constraint_id = "symbolic_verifier.no_counterexample_found". The DiagnosticResult contract structurally requires a proof_ref for VERIFIED, and this engine has no proof artifact to bind. Downstream policy gates must reject symbolic-engine output for control flow.
verify_code() no longer accepts check_assertions — the parameter was never wired up. The current signature is verify_code(self, code: str) -> DiagnosticResult.
Advisory checks throughout — verify_safety_properties, analyze_complexity, and get_verification_budget attach structured AdvisoryCheck entries to developer_fields["advisory_checks"]. Advisory checks never influence the verdict.

Before and after

Before

result = verifier.verify_code(code)

if result["status"] == "verified":
    admit(payload)
elif result["status"] == "counterexamples_found":
    for issue in result["issues"]:
        log(issue["description"])

After

from qwed_new.core.diagnostics import DiagnosticStatus

result = verifier.verify_code(code)

# This engine never emits VERIFIED — treat every result as unverified for
# control flow, and use developer_fields for the specific reason.
reject(payload, reason=result.agent_message)

if result.status is DiagnosticStatus.UNVERIFIABLE:
    if result.developer_fields["constraint_id"] == "symbolic_verifier.counterexample_found":
        for issue in result.developer_fields["issues"]:
            log(issue["description"])

Field cheat sheet

Legacy field	Replacement
`result.verified` / `result["verified"]`	`result.is_verified` (always `False` for this engine)
`result["status"]` (string)	`result.status` (`DiagnosticStatus` enum) and `developer_fields["constraint_id"]`
`result["message"]`	`result.agent_message`
`result["issues"]`	`result.developer_fields["issues"]`
`result["complexity"]` / `"loop_depth"` / `"recursive"`	`developer_fields["complexity_score"]`, `"max_loop_depth"`, `"total_recursive_functions"`
`result["is_safe"]`	`result.developer_fields["is_safe"]`
`result["bounds_applied"]`	`result.developer_fields["bounds_applied"]`
n/a	`result.developer_fields["verification_mode"]` — new (`"symbolic"` or `"bounded_symbolic"`)
n/a	`result.proof_ref` — always `None` for this engine

Version propagation

Artifact	Previous	This release
`qwed` (PyPI)	5.2.0	5.3.0
`qwed_sdk` (Python)	5.2.0	5.3.0
`@qwed-ai/sdk` (npm)	5.2.0	5.3.0
`qwed` (Rust crate)	5.2.0	5.3.0
API version marker	5.2.0	5.3.0
K8s deployment image	5.2.0	5.3.0

Included PRs

#212 — feat: migrate SymbolicVerifier to DiagnosticResult (Phase 1)
#220 — fix: remove unused check_assertions parameter from verify_code
#239 — feat: add verification_mode to SymbolicVerifier DiagnosticResults
#240 — feat: migrate verify_bounded to return DiagnosticResult
#241 — feat: migrate get_verification_budget to return DiagnosticResult
#242 — feat: migrate analyze_complexity to return DiagnosticResult
#243 — feat: migrate verify_safety_properties to return DiagnosticResult
#244 — release: v5.3.0 — SymbolicVerifier: DiagnosticResult reference implementation

What this means for you

If your integration reads any dict-shape field from a SymbolicVerifier result, update it before upgrading — the return type has changed and the old keys are gone. The Code engine reference documents the new contract, and Symbolic execution limits has been rewritten around DiagnosticResult. Callers that gated control flow on is_verified for symbolic-engine output must move to result.proof_ref (the authority bit) instead — this engine intentionally never sets it, so symbolic results must not admit downstream execution.

QWED-A2A — `UNVERIFIABLE` verdicts carry no JWT, and five README behaviors realigned with the code

Released: July 27, 2026 · GitHub PR #40 · GitHub PR #57

The interceptor no longer produces a signed FORWARDED verdict for empty or malformed FINANCIAL_TRANSACTION and LOGIC_ASSERTION payloads, and no longer silently forwards GENERAL / DATA_QUERY messages. Both cases now return verdict.status = "unverifiable" with attestation_jwt = None — signing a token for content that was never verified would be a false cryptographic claim. The A2A docs are updated to match the shipping behavior on five points that previously overstated what the interceptor does.

What changed

Empty/malformed finance and logic payloads → UNVERIFIABLE, no JWT. A FINANCIAL_TRANSACTION missing data, line_items, or claimed_total, or with non-numeric amounts, now returns status=unverifiable from finance_guard. A LOGIC_ASSERTION with a missing, non-list, or empty assertions array, or malformed entries, returns status=unverifiable from logic_guard. Earlier releases collapsed these cases to verified=True + a signed FORWARDED verdict.
GENERAL and DATA_QUERY are no longer silently forwarded. The passthrough branch now returns an unverifiable verdict with engine_used="passthrough", attestation_jwt=None, and reason "No verification engine available for this payload type". Callers must decide their own downstream policy for unverified traffic.
No Bypass verification stage. The pipeline has four stages, not five. config.trusted_agents is pre-added to the TrustBoundary allowlist at interceptor construction time — trusted agents still flow through engine routing and receive a normal verdict. The Verification interceptor and Architecture pages have been corrected.
CodeGuard is AST-first, regex-second. Code payloads are first parsed into an AST and inspected for direct dangerous constructs (eval(, exec(, subprocess.run(, import subprocess, etc.). Only if the AST layer is clean does a regex heuristic scan run to catch obfuscation patterns (getattr(__builtins__, ...), base64-encoded exec, dynamic __import__). The verdict is BLOCKED if either layer triggers, or HEURISTIC_PASS if both are clean — not a proof of safety.
JWT attestation expiry is 300 seconds (5 minutes), not 24 hours. The default validity_seconds on A2ACryptoService is 300 — one A2A hop lifetime. The Crypto attestations page and its signing example were updated.

New verdict-status contract

Status	Meaning	JWT attestation
`forwarded`	Engine verified the payload.	Signed
`blocked`	Engine detected a violation.	Signed
`heuristic_pass`	Code guard found no known dangerous constructs. Not a proof of safety.	Signed
`unverifiable`	No engine could evaluate the payload (`GENERAL`/`DATA_QUERY`, empty or malformed finance/logic payload).	None

Behavior change. Downstream code that assumed every VerificationVerdict carries an attestation_jwt will see None on UNVERIFIABLE verdicts and must handle it. Callers that treated a signed FORWARDED verdict as proof that finance/logic content had been checked should also branch on verdict.status — an empty payload no longer produces a signed FORWARDED, and a GENERAL/DATA_QUERY message no longer produces one at all. Note that VerdictStatus.ERROR exists in the public enum but is not emitted by intercept() — internal exceptions surface as BLOCKED (or FORWARDED when block_on_error=False), never as an error verdict.

What this means for you

If you integrate QWED-A2A: audit any caller that reads attestation_jwt or trusts status == "forwarded" without also checking engine_used. Legitimate agent traffic on unsupported payload types is not blocked — it is marked unverifiable so your policy layer can decide whether to route it. See the updated Verification interceptor and Quick start pages for the full verdict contract and example payload shapes.

QWED verification — signature-first attestation validation, uniform rejection error, and TOCTOU-safe trust-boundary snapshot

Released: August 1, 2026 · PR #287 · PR #290

Two follow-ups harden the attestation trust boundary introduced last week. verify_attestation() now verifies the JWT signature before applying any issuer authorization or revocation checks, and every rejection path returns the same generic "Invalid token" error so callers cannot enumerate trusted issuers, probe token expiry, or confirm revocation state through response text. enforce_trust_decision() now returns a fully detached DiagnosticResult snapshot, closing a TOCTOU window where a concurrent caller could mutate developer_fields between the validation read and the admission decision.

What changed

Signature-first verification order. verify_attestation() cryptographically verifies the JWT (signature + expiration + required claims) before it applies the trusted-issuer authorization check or the revocation check. An unknown-issuer token that is not correctly signed is rejected on signature grounds and the trust-list check never runs.
Uniform "Invalid token" rejection. Every failure mode — expired, revoked, untrusted issuer, unsupported external issuer, malformed, oversized, malformed base64 — returns the same (is_valid=False, claims=None, error="Invalid token") triple. Detailed reasons are still recorded in the server-side attestation.rejected audit log; only the caller-facing error is generic. This closes the trusted-issuer enumeration side channel.
Detached-result snapshot in enforce_trust_decision(). Before validation, developer_fields is recursively rebuilt into an isolated snapshot that admits only immutable scalars, AdvisoryCheck, and JSON-safe containers. The value you validate is the value you return, so a concurrent mutation of the caller’s DiagnosticResult can neither skew the decision nor leak into the returned result.
Fail-closed snapshot failure. When developer_fields contains an unsupported value type (custom class instance, generator, open file handle) or an object that resists copying, enforce_trust_decision() returns a BLOCKED result with constraint_id="trust_gate.diagnostic_snapshot_failed". Only the exception type is logged — never the exception message, which could embed caller data.
Runtime dependency. cryptography is now a runtime dependency (previously an optional extra), so ES256 signature verification runs by default.

Before and after

Rejection error — before

# Distinct messages leaked expiration, revocation, and trusted-issuer identity.
is_valid, claims, error = client.verify_attestation(bad_jwt)
# error = "Untrusted issuer: did:qwed:node:staging"
# error = "Attestation has expired"
# error = "Attestation has been revoked"
# error = "External issuer key resolution not implemented"

Rejection error — after

# Every rejection path returns the same string.
is_valid, claims, error = client.verify_attestation(bad_jwt)
# is_valid = False, claims = None, error = "Invalid token"

Trust-boundary snapshot — before

result = engine.verify(query)  # DiagnosticResult(developer_fields={"score": 0.9})
decision = enforce_trust_decision(result, attestation_token=token, ...)

# Concurrent code holding `result` could mutate developer_fields AFTER
# validation but BEFORE the caller admitted the decision — the returned
# reference aliased the input.
result.developer_fields["score"] = -1.0
assert decision.developer_fields["score"] == -1.0  # TOCTOU: mutation leaked in

Trust-boundary snapshot — after

result = engine.verify(query)
decision = enforce_trust_decision(result, attestation_token=token, ...)

# `decision.developer_fields` is a detached snapshot. Mutations to the
# original never reach the enforced result.
result.developer_fields["score"] = -1.0
assert decision.developer_fields["score"] == 0.9  # unchanged

What this means for you

Callers of verify_attestation() must not branch on the error string any more — treat any is_valid=False result as a hard block and rely on the server-side audit log for the actual reason. Callers of enforce_trust_decision() should keep developer_fields values to JSON-safe types (strings, numbers, booleans, None, lists, dicts) plus AdvisoryCheck; anything else will be rejected by the snapshotter and blocked at the trust boundary with trust_gate.diagnostic_snapshot_failed. See Uniform rejection error and Detached result snapshot for the full behavior.

QWED verification — mandatory proof artifact for VERIFIED attestations and a single trust-boundary entry point

Released: July 28, 2026 · PR #248

Two related enforcement-boundary changes ship together. Issuance side, the crypto layer now refuses to sign a VERIFIED verdict that has no proof artifact. Consumption side, a new enforce_trust_decision() function is the single trust-boundary entry point that release gates route through — it verifies the attestation JWT, binds its claims to the result (status, query hash, proof hash), and fails closed on any mismatch.

What changed

create_verification_attestation() blocks VERIFIED without proof. Calling with verified=True (or status="VERIFIED") and an empty or missing proof_data now returns AttestationResult(status=BLOCKED, token=None, error_code="VERIFIED_WITHOUT_PROOF", error=...). The crypto layer will not sign a VERIFIED result that has no evidence to hash into the token’s proof_hash claim. UNVERIFIABLE and BLOCKED verdicts are unaffected.
New enforce_trust_decision() trust-boundary gate. Exported from qwed_new.core, this is the single consumption-side entry point every release gate must route through. It takes the engine’s DiagnosticResult plus the attestation token, verifies the JWT (signature, expiry, trusted issuer), and blocks the decision when the token’s qwed.result.status, qwed.query_hash, or qwed.proof_hash claims do not match the result. UNVERIFIABLE/BLOCKED results pass through unchanged.
require_attestation policy toggle. enforce_trust_decision(result, require_attestation=True, ...) (default) is the mandatory policy — VERIFIED without a valid token is downgraded to BLOCKED. require_attestation=False is the advisory policy for staged rollouts: a missing token still passes as VERIFIED, but a present-and-invalid token still blocks. Both modes emit structured trust_gate.blocked audit logs on rejection.
Control plane integration. The math control plane now routes every verification through enforce_trust_decision() and records the outcome and policy on the trust boundary (trust_enforced, attestation_policy) so operators can watch the mandatory rollout from telemetry before flipping the switch.

Before and after

Issuance — VERIFIED without a proof artifact

# Before — the crypto layer signed a VERIFIED verdict even with no proof_data,
# emitting an attestation whose proof_hash pointed at an empty artifact.
create_verification_attestation(
    status="VERIFIED",
    verified=True,
    engine="math",
    query="2+2=4",
)
# AttestationResult(status=ISSUED, token="eyJ...", error_code=None)

# After — issuance is refused; there is no way to obtain a signed
# attestation for a VERIFIED claim without evidence.
create_verification_attestation(
    status="VERIFIED",
    verified=True,
    engine="math",
    query="2+2=4",
)
# AttestationResult(
#     status=BLOCKED,
#     token=None,
#     error_code="VERIFIED_WITHOUT_PROOF",
#     error="VERIFIED status requires proof_data — cannot sign attestation without proof artifact",
# )

Consumption — release gate on a VERIFIED result

# Before — release gates inspected result.status directly and could admit
# a VERIFIED result whose attestation was missing, invalid, or bound to a
# different query. There was no single enforcement point.
if result.status == "VERIFIED":
    proceed()

# After — one call gates the boundary. Missing/invalid tokens or
# mismatched claims downgrade the decision to BLOCKED with a
# structured audit event.
from qwed_new.core import enforce_trust_decision

decision = enforce_trust_decision(
    result,
    attestation_token=token,
    require_attestation=True,
    trusted_issuers=["did:qwed:node:production"],
    query="2+2=4",
)

if decision.status.name == "VERIFIED":
    proceed()
else:
    reject(decision)

What this means for you

If you call create_verification_attestation() directly, always pass proof_data (typically json.dumps(evidence, sort_keys=True)) when signing a VERIFIED verdict, and check result.is_issued before treating the token as valid. If you consume verification results in a release gate, route them through enforce_trust_decision() — start in advisory mode (require_attestation=False) while your engines are being migrated to emit proof artifacts, watch the trust_boundary.trust_enforced telemetry, then flip to mandatory once every VERIFIED path is signing with proof. See Fail-closed contract and Enforce the trust boundary for the full parameter reference and fail-closed matrix.

QWED-A2A — `verify_attestation()` now requires an `AttestationContext` (breaking)

Released: July 26, 2026 · GitHub PR #41

A2ACryptoService.verify_attestation() used to accept a token on its own — a cryptographically valid signature was treated as sufficient proof. A detached attestation could therefore be lifted from one exchange and replayed against a different sender, receiver, or payload without detection. The method now requires a second argument, an AttestationContext describing the current request, and rejects the token if the sender, receiver, payload hash, or session claim does not match.

Breaking change. The single-argument form verify_attestation(token) has been removed — there is no context-free overload. Every call site must construct an AttestationContext and pass it as the second positional argument, or the call raises TypeError. See the migration example under “What this means for you” below for the exact before/after.

What changed

New AttestationContext dataclass in qwed_a2a.security.crypto with fields sender_agent_id: str, receiver_agent_id: str, payload: Any, and optional session_id: str | None. The payload is hashed internally with the same deterministic method as sign_verdict(), so callers never handle raw hashes.
New context-binding verification step — sender, receiver, payload hash, and (when supplied) session_id are compared against the token’s qwed_a2a claims and sub claim. Any mismatch returns a structured rejection: "Attestation sender mismatch: expected=..., got=...", "Attestation receiver mismatch: ...", "Attestation payload hash mismatch — detached attestation rejected", or "Attestation session mismatch: ...".
Verification order updated. Context binding now runs after the deployment-context check but before the jti replay check. This is deliberate: an out-of-context token no longer pollutes the jti registry, so a legitimate future presentation of the same token in its correct context is not locked out.
Signing is unchanged. sign_verdict() still takes the same arguments — only the verify side gained a parameter.

What this means for you

Every caller of verify_attestation() must be updated. Build an AttestationContext from the same identifiers and payload the receiving side is about to act on, then pass it as the second argument:

from qwed_a2a.security.crypto import AttestationContext

context = AttestationContext(
    sender_agent_id="procurement-agent",
    receiver_agent_id="treasury-agent",
    payload=request_payload,
    session_id=request_session_id,  # optional
)
is_valid, claims, error = crypto.verify_attestation(token, context)

See Verifying an attestation for the full field-by-field reference, the updated ordered verification steps, and the complete list of rejection messages.

QWED Math Engine — fail-closed on mode ambiguity, eigenvalue cardinality, and IRR convergence

Released: July 24, 2026 · Jump to details · PR #217 · PR #218 · PR #219

Three core-verifier fixes tighten the math engine’s fail-closed contract. verify_statistics(statistic="mode"), verify_matrix_operation(operation="eigenvalues"), and verify_irr() no longer produce VERIFIED when the underlying claim is ambiguous, under-specified, or numerically unproven. Callers see BLOCKED or CORRECTION_NEEDED with structured diagnostics instead of a best-effort answer.

Behavior change. Inputs that previously received VERIFIED may now receive BLOCKED or CORRECTION_NEEDED — treat both as unverified. If your code branched on result.verified or result.status == "VERIFIED" alone, it will continue to work; if you consumed calculated/calculated_irr/calculated_eigenvalues from a BLOCKED result as a fallback, those fields are not present on BLOCKED responses. They remain available on CORRECTION_NEEDED responses for diagnostic use. Consume the new structured fields (ambiguous_modes, calculated_count/claimed_count, converged, iterations_used) for richer diagnostics.

What changed

verify_statistics(statistic="mode") requires a unique mode. When two or more values tie for the maximum frequency, the engine returns BLOCKED with an ambiguous_modes list instead of heuristically picking one. Only a unique mode can produce VERIFIED.
verify_matrix_operation(operation="eigenvalues") requires cardinality match. The claimed eigenvalue list length must equal the calculated count (counting algebraic multiplicity). Mismatched lengths return CORRECTION_NEEDED with calculated_count/claimed_count. Previously the value comparison used zip, which silently truncated to the shorter list.
verify_irr() requires proof of Newton-Raphson convergence. BLOCKED is now returned when cash flows have more than one sign change (multi-root ambiguity per Descartes’ rule), zero sign changes (no real IRR), all-zero cash flows (IRR undefined), the Newton derivative stalls at zero, or the method fails to converge within 100 iterations. Successful results include converged: true and iterations_used.

Before and after

Ambiguous mode

# Before — heuristically picked one of the tied values and could return VERIFIED
client.verify_statistics(statistic="mode", data=[1, 1, 2, 2, 3], expected=1)
# status: "VERIFIED"

# After — BLOCKED with the full list of tied values
client.verify_statistics(statistic="mode", data=[1, 1, 2, 2, 3], expected=1)
# status: "BLOCKED"
# ambiguous_modes: [1, 2]

Incomplete eigenvalue claim

# Before — zip truncated the comparison to length 1 and returned VERIFIED
client.verify_matrix_operation(operation="eigenvalues", matrix=[[2, 0], [0, 3]], expected=[2])
# status: "VERIFIED"

# After — CORRECTION_NEEDED with explicit cardinality diagnostics
client.verify_matrix_operation(operation="eigenvalues", matrix=[[2, 0], [0, 3]], expected=[2])
# status: "CORRECTION_NEEDED"
# calculated_count: 2, claimed_count: 1
# calculated_eigenvalues: [2.0, 3.0]

IRR with multiple sign changes

# Before — Newton-Raphson returned a best-effort iterate that could be VERIFIED
client.verify_irr(cash_flows=[-100, 230, -132], expected=0.10)
# status: "VERIFIED"

# After — BLOCKED because two sign changes admit multiple real IRRs
client.verify_irr(cash_flows=[-100, 230, -132], expected=0.10)
# status: "BLOCKED"
# sign_changes: 2

What this means for you

Existing code that already treated non-VERIFIED statuses as unverified continues to work. New policy code should read the structured diagnostic fields on BLOCKED/CORRECTION_NEEDED results — ambiguous_modes, calculated_count/claimed_count, and converged/iterations_used — to explain why a claim was rejected and to satisfy audit requirements. See Math engine — Fail-closed semantics for the full state tables and response schemas.

QWED-Finance — v2.1.0 released

Released: July 22, 2026 · GitHub PR #40

qwed-finance v2.1.0 is now the current release. This is a version-sync release: the Python package, npm wrapper, GitHub Action, and quickstart workflow reference are all realigned so QWED Finance Guard reports a single consistent version across PyPI, npm, and SARIF output in GitHub Advanced Security.

What changed

qwed-finance Python package is now v2.1.0 on PyPI (qwed_finance.__version__ == "2.1.0").
@qwed-ai/finance npm package version bumped to 2.1.0 to match.
GitHub Action auto-syncs its reported version from the installed package — the QWED Finance Guard name in workflow logs and the version field on SARIF uploads to the GitHub Security tab now both read 2.1.0 instead of a hardcoded v2.0.
Quickstart qwed-verify.yml workflow is re-pinned from the stale v1.1.4 SHA to the v2.1.0 SHA (QWED-AI/qwed-finance@19ce969f21d1fc2019da4d89fff23bc108e15a98 # v2.1.0).

What this means for you

Upgrade your dependency to pick up the current release:

pip install --upgrade qwed-finance

If you run QWED Finance in CI, update the action reference so SARIF findings and workflow-run names line up with the current package:

- name: Verify banking calculations
  uses: QWED-AI/qwed-finance@v2.1.0

This is a version-sync patch for the existing v2.1.0 release (May 2026). It contains no new API or guard behavior changes beyond those already shipped in v2.1.0 (Decimal migration, fail-closed enforcement, rate parsing fix). For the original breaking changes, see the v2.1.0 release notes. See GitHub Action (CI/CD) for the updated workflow example.

QWED-UCP — v0.3.0 released

Released: July 20, 2026 · GitHub PR #38

qwed-ucp v0.3.0 is now the current release. This version ships the typed TrustStatus enum on every verification result, alongside the fail-closed middleware and internal-error handling delivered over the v0.2.x series.

What changed

qwed-ucp Python package is now v0.3.0 on PyPI.
Express middleware qwed-ucp-middleware package version bumped to match.
TrustStatus enum is confirmed as available from v0.3.0 onward — see Trust status for the full state table and usage.
GitHub Action should now be pinned to QWED-AI/qwed-ucp@v0.3.0.

What this means for you

Upgrade your dependency to pick up the current release:

pip install --upgrade qwed-ucp

If you audit checkouts in CI, update the action reference:

- name: Audit Commerce Transactions
  uses: QWED-AI/qwed-ucp@v0.3.0

Existing code that branches on result.verified continues to work unchanged. New code should branch on result.status (a TrustStatus) to distinguish FAILED from ENGINE_ERROR, UNVERIFIABLE, and other non-VERIFIED verdicts.

QWED-UCP — Express middleware fails closed on internal verification errors

Released: July 18, 2026 · GitHub PR #36

When a guard raised an unexpected exception, the Express middleware previously logged the error and called next() — letting an unverified checkout through to the downstream handler. That defeated the trust boundary. The catch block now short-circuits with HTTP 500, X-QWED-Verified: false, and code: "INTERNAL_VERIFICATION_ERROR" so an internal crash can no longer be mistaken for a passing verification.

What changed

Express middleware catch block is now fail-closed. On any exception raised during verifyCheckoutLocally(), the middleware returns:
{ "error": "QWED-UCP Verification Failed", "message": "Internal verification error: verification could not be completed", "code": "INTERNAL_VERIFICATION_ERROR" }
The underlying exception is logged server-side via console.error but is not included in the response body — raw stack traces, file paths, and internal messages stay out of client-visible output.
X-QWED-Verified: false is set on the 500 response so upstream policy layers can treat it the same as a 422 failure.
npm package fix. qwed-ucp-middleware.js is now included in the published files array. Previous versions installed via npm were missing the entrypoint that index.js required at runtime.

Behavior change. Any client code that treated an Express middleware error as a soft pass (e.g. retrying without checking the status, or relying on next() being called) will now see a 500. Treat 500 INTERNAL_VERIFICATION_ERROR the same as 422 VERIFICATION_FAILED — the request has not been verified and must not be settled.

What this means for you

Existing integrations that already branched on X-QWED-Verified or on the response status keep working — they just start seeing a new 500 path that was previously invisible. New integrations should treat both 4xx and 5xx responses from the middleware as an unverified request. See Express.js middleware — Fail-closed on internal verification errors for the full response contract.

QWED-UCP — typed `TrustStatus` enum on every verification result

Released: July 15, 2026 · GitHub PR #33

Verification results previously exposed a single verified: bool that collapsed “proof disproved,” “proof could not be established,” “input outside supported semantics,” and “verifier engine crashed” into the same False bucket. Every result dataclass now also carries a typed status: TrustStatus field so downstream policy code can make trust-aware decisions without parsing error strings.

What changed

New TrustStatus enum exported from qwed_ucp with seven states: VERIFIED, FAILED, UNVERIFIABLE, UNSUPPORTED, PARTIAL, ENGINE_ERROR, and QUARANTINED (reserved).
status field on every result — GuardResult, UCPVerificationResult, and each per-guard result type (MoneyGuardResult, StateGuardResult, SchemaGuardResult, LineItemsGuardResult, DiscountGuardResult, CurrencyGuardResult, RefundGuardResult, TipGuardResult, FeeGuardResult, AttestationResult).
UCPVerifier.verify_checkout() now surfaces ENGINE_ERROR when any guard raises an exception, instead of silently collapsing to verified=False.
verified: bool is preserved as a backward-compatible derived field: result.verified is True only when result.status == TrustStatus.VERIFIED. Existing constructors that pass verified=True/verified=False continue to produce the corresponding VERIFIED/FAILED status.

What this means for you

Existing code that reads result.verified or constructs results with verified=... keeps working unchanged. New policy code should branch on result.status to distinguish a disproved proof from a crashed verifier — see Trust status for the full state table and an example.

QWED-UCP — middleware fails closed on empty and non-JSON request bodies

Released: July 14, 2026 · GitHub PR #32

The FastAPI and Express middleware previously forwarded requests with an empty body, malformed JSON, or a non-object JSON payload to the downstream handler without running any guards. On a /checkout-sessions route that defeated the trust boundary — an attacker could send {}’s worth of nothing and skip verification. Both middlewares now reject those requests with HTTP 422, X-QWED-Verified: false, and code: "UNPARSEABLE_REQUEST" before the handler runs.

What changed

FastAPI: returns a distinct message per case:
- Empty body → "Empty request body: cannot verify empty payload"
- Malformed or non-UTF-8 body → "Malformed request body: expected JSON"
- Top-level non-object JSON → "Invalid request body: expected JSON object"
Express: all three cases return "Empty or non-JSON request body: cannot verify unparseable payload".
Non-protected methods and paths (for example, GET /health or a POST to a route outside verify_paths) still pass through untouched.

This is a fail-closed behavior change. Any client that was previously reaching a /checkout-sessions, /checkout, /cart, or /payment handler with a missing, malformed, or non-object JSON body will now receive 422 UNPARSEABLE_REQUEST instead. Send checkout payloads as application/json with a JSON object at the top level, or narrow verify_paths if a route should not be treated as a checkout endpoint.

What this means for you

If you deploy QWED-UCP as middleware in front of a UCP merchant server, unparseable requests can no longer bypass the guards. See FastAPI middleware, Express.js middleware, and the troubleshooting entry for the exact response shape and how to configure protected paths.

QWED-A2A — persistent signing key and JWKS discovery endpoint

Released: July 10, 2026 · GitHub PR #29

A2ACryptoService no longer generates an ephemeral ECDSA P-256 key per process. The signing key is now loaded from the QWED_A2A_SIGNING_KEY_PEM environment variable (or the pem_key constructor argument) so attestation JWTs issued before a restart remain verifiable afterwards. A new /.well-known/jwks.json endpoint publishes the current public key for external consumers.

What changed

Persistent signing key — A2ACryptoService reads an unencrypted PKCS#8 P-256 PEM from QWED_A2A_SIGNING_KEY_PEM on first use. The derived key_id is a SHA-256 fingerprint of the public key, so it stays stable as long as the PEM does.
Fail-closed on missing key — sign_verdict, verify_attestation, get_public_key_jwk, and A2AVerificationInterceptor.intercept() all raise RuntimeError when the PEM is missing, malformed, or uses a curve other than SECP256R1. The FastAPI gateway surfaces this as HTTP 503 Signing key unavailable.
JWKS endpoint — A new wellknown_router exposes GET /.well-known/jwks.json for downstream services and auditors to fetch the current public key without an out-of-band exchange.
get_public_key_jwk() — Returns the current public key as a JWK (kty, crv, x, y, kid, use, alg).

Breaking changes

A2ACryptoService() and A2AVerificationInterceptor() no longer produce a working signer without configuration. Every deployment must now set QWED_A2A_SIGNING_KEY_PEM (and continue to set QWED_A2A_DEPLOYMENT_ID) before the service starts. Generate a key with:

openssl ecparam -name prime256v1 -genkey -noout \
  | openssl pkcs8 -topk8 -nocrypt \
  > qwed_a2a_signing_key.pem

Load the PEM into QWED_A2A_SIGNING_KEY_PEM from your secrets manager, and make sure every replica in a logical deployment uses the same PEM so their kids match and tokens cross-verify.

What this means for you

If you deploy the QWED-A2A gateway, you now get audit continuity across restarts and rolling deployments — a JWT signed by an earlier process is still verifiable by the next one, provided the PEM is unchanged. External services can also verify attestations by fetching /.well-known/jwks.json directly. See Crypto attestations and Deployment for the full setup, JWKS shape, and key rotation guidance.

QWED-Tax — structured 3-layer diagnostics for TDS, ITC, and GST-RCM

Released: June 21, 2026 · GitHub PR #45

QWED-Tax adopts the same 3-layer DiagnosticResult contract introduced in QWED-Verification v5.2.0 — as an independent model, no cross-package dependency. The first three guards (TDS, Input Tax Credit, GST reverse-charge) now expose a structured diagnostic in addition to their existing dict response.

What’s new

Tri-state status — Every diagnostic resolves to VERIFIED, UNVERIFIABLE, or BLOCKED. No HEURISTIC or AMBIGUOUS middle ground.
Three disclosure layers — A short agent-safe summary (no statute IDs or detection logic), a structured developer evidence block (constraint ID, statute, jurisdiction, audit trace, deduction, net payable), and an optional proof reference.
Proof reference is the authority bit — A deterministic sha256 hash of the audit trace, present only when a result is VERIFIED. Absent on UNVERIFIABLE and BLOCKED.
First migrations — TDSGuard, InputCreditGuard, and GSTGuard (reverse-charge) each expose the new diagnostic format alongside their existing return shape.

Compatibility

Additive release. The legacy dict API on every guard is unchanged — the diagnostic format is opt-in. Existing integrations continue to work without modification. The remaining nine QWED-Tax guards (CapitalGains, Classification, Speculation, Setoff, Crypto, Valuation, Remittance, PoEM, Withholding) will migrate in follow-up releases.

What this means for you

If you’ve already adopted the QWED-Verification diagnostic contract, the same response shape now applies to TDS, ITC, and GST-RCM checks — including the proof-reference authority bit you can gate execution on. See the Verification Diagnostics guide for the response shape and the tax guards reference for guard-by-guard coverage.

QWED-Tax — exact paise comparison, edge-case input rejection, and strict payload schemas

Released: June 21, 2026 · GitHub PR #44

Three input-strictness fixes ship together. CryptoTaxGuard.verify_flat_tax_rate now compares quantized paise values exactly instead of within a 0.1 tolerance, ValuationGuard and RemittanceGuard reject the edge-case numeric inputs that previously slipped through, and every QWED-Tax Pydantic input model now forbids unexpected fields.

What changed

CryptoTaxGuard.verify_flat_tax_rate — The Decimal("0.1") tolerance was removed. Both the computed expected_tax and the caller’s claimed_tax are now quantized to two decimal places using ROUND_HALF_UP and compared with an exact ==. A 1-paise (0.01) deviation correctly returns verified=False.
ValuationGuard.verify_conversion — Added explicit range checks: discount must be in [0, 1), and cap, next_round_price, and investment must all be strictly positive. DivisionByZero is now caught alongside InvalidOperation so a degenerate cap = 0 or discount = 1 no longer crashes — it fails closed with a structured {"verified": False, "error": "..."} response. The previous behavior allowed a discount > 1 to produce a negative share count.
RemittanceGuard.verify_lrs_limit — After numeric parsing, amount_usd and financial_year_usage are now checked for negativity. Negative inputs return {"verified": False, "error": "BLOCKED: ..."} instead of being summed into the limit check, where a negative usage could mask a transaction that exceeds the $250,000 LRS cap.
Input models — Address, WorkArrangement, WorkerClassificationParams, ContractorPayment, TaxEntry, DeductionEntry, PayrollEntry, and VerificationResult are now configured with model_config = ConfigDict(extra="forbid"). Any payload with an unexpected key raises a Pydantic ValidationError at the boundary. QWEDTaxMiddleware already surfaces this as status: "BLOCKED" with risk: "INVALID_PAYLOAD".

Breaking changes

Callers that previously relied on the 0.1-rupee tolerance in CryptoTaxGuard.verify_flat_tax_rate will now receive verified=False for claims that differ from vda_income * 0.30 by 1 paise or more. Round your claimed tax to two decimal places with ROUND_HALF_UP before calling the guard.

from decimal import Decimal, ROUND_HALF_UP

claimed = (vda_income * Decimal("0.30")).quantize(Decimal("0.01"), rounding=ROUND_HALF_UP)

Payloads sent through QWEDTaxMiddleware (or constructed directly with the QWED-Tax input models) that include keys outside the declared schema now raise ValidationError. Strip unknown fields from AI-generated output — including typos and speculative overrides like "override_verification": true — before invoking the middleware.

ValuationGuard.verify_conversion no longer returns a result for cap <= 0, next_round_price <= 0, investment <= 0, discount < 0, or discount >= 1. These inputs now return {"verified": False, "error": "..."} instead of crashing or producing nonsensical share counts.

What this means for you

If your agent forwards Indian VDA tax claims, startup conversion math, or LRS remittance requests through QWED-Tax, audit the calling code for: (1) tax claims that aren’t pre-quantized to two decimal places, (2) discount/cap/investment inputs that can legitimately be zero or out of range, (3) negative remittance amounts being passed defensively, and (4) AI payloads that include fields not declared on the QWED-Tax input models. See the CryptoTaxGuard, ValuationGuard, RemittanceGuard, and middleware integration guide for the updated contracts.

Audit reference

Issue	Area	Fix
#20	`CryptoTaxGuard` accepted claims within a 0.1-rupee tolerance	Quantize to paise with `ROUND_HALF_UP`, compare with exact `==`
#21	`ValuationGuard` and `RemittanceGuard` accepted edge-case numeric inputs	Range checks on discount/cap/investment; negativity checks on LRS amount and usage; `DivisionByZero` caught
#22	Input models silently accepted unexpected fields	`model_config = ConfigDict(extra="forbid")` on all eight input models

QWED-Tax — middleware never returns full “verified” and ReciprocityGuard no longer always-passes

Released: June 21, 2026 · GitHub PR #43

Two fail-closed fixes ship together. The Gusto interceptor middleware no longer overstates a gross-to-net arithmetic pass as full tax verification, and ReciprocityGuard no longer returns verified=True for arrangements with no reciprocity agreement.

ARITHMETIC_VERIFIED is a pre-conformance middleware-layer status, not part of the DiagnosticResult tri-state vocabulary (VERIFIED / UNVERIFIABLE / BLOCKED) introduced in v5.2.0. The middleware will migrate to DiagnosticResult when engine-conformance work lands. Until then, treat ARITHMETIC_VERIFIED as a distinct middleware signal — it is not equivalent to VERIFIED.

What changed

QWEDTaxMiddleware.process_ai_payroll_request — The success response is now status: "ARITHMETIC_VERIFIED" with execution_permitted: false. The middleware only verifies gross-to-net math; classification, withholding legality, reciprocity, and filing checks are still required before execution. The response includes checks_run (what was verified) and checks_not_run (what is still required) so callers can decide what to run next.
TaxPreFlight report — Every report now includes a checks_not_run list covering both unselected guards and known gaps (checks not yet implemented for the action). For example, action="hire" reports payroll_arithmetic, withholding_legality, reciprocity, and filing_obligations as known gaps.
ReciprocityGuard — The Z3 solver was removed (the prior expression was tautologically satisfiable and ignored the same_state parameter). The guard is now a deterministic lookup against the reciprocity-pair table with explicit fail-closed paths: same state → verified, known pair → verified, no agreement → verified=False, unknown state → verified=False. A new verify_reciprocity(residence_state, work_state, same_state=None) method takes string inputs; determine_withholding_state(arrangement) is kept for backwards compatibility.

The reciprocity-pair table covers only the pairs modeled within the State enum’s 8-state scope (NJ, PA, MD, VA). Pennsylvania has additional reciprocity agreements (with IN, MI, OH, VA, WV, WI) that are not modeled here because those states are not in the enum. VA-PA in particular is a known gap — callers will receive verified=False for that pair until the enum and table are extended.

Breaking changes

The middleware no longer returns status: "VERIFIED" or execution_permitted: true. Any caller that gated execution on decision["status"] == "VERIFIED" or decision["execution_permitted"] being truthy will now always block. Update your integration to handle ARITHMETIC_VERIFIED and run the checks listed in checks_not_run (worker classification, withholding legality, reciprocity, filing) before forwarding to Gusto/Avalara.

ReciprocityGuard no longer returns verified=True for state pairs without a reciprocity agreement. Callers that previously relied on a Z3-backed “always sat” result for, e.g., NJ → NY will now correctly receive verified=False. Route these to your withholding logic for the work state or to human review.

What this means for you

If your agent forwards payroll payloads to Gusto/Avalara based on a VERIFIED status from the middleware, those calls will start blocking until you run the remaining guards (ClassificationGuard, WithholdingGuard, ReciprocityGuard.verify_reciprocity, Form1099Guard) yourself. See the tax integration guide for the updated response shape and the ReciprocityGuard reference for the new lookup contract.

Audit reference

Issue	Area	Fix
#19	Middleware overstated partial verification as full verification	Status narrowed to `ARITHMETIC_VERIFIED`, `execution_permitted` forced to `false`, `checks_run`/`checks_not_run` surfaced
#40	`ReciprocityGuard` Z3 solver always returned sat	Z3 removed; deterministic lookup with explicit fail-closed branches

QWED-Tax — fail-closed on ambiguous classification and unverified claims

Released: June 21, 2026 · GitHub PR #42

Six more QWED-Tax guards now refuse to sign off when they can’t independently prove a result. Ambiguous worker classifications, unparseable trade dates, unknown set-off heads, and reverse-charge inputs the guard doesn’t recognize all return verified=False with a structured error instead of a quiet pass.

Bug fixes

CapitalGainsGuard — Unparseable acquisition or disposal dates and unknown asset types now block instead of being coerced into a sentinel value that flowed through to verified=True. SLAB-rate verification can no longer succeed without an income bracket — slab rates can’t be proven from the claim alone.
ClassificationGuard — Mixed employee/contractor signals now return verified=False with an “ambiguous classification” error. The guard only returns CONTRACTOR when no employee indicators are present.
SpeculationGuard — Set-off verification now requires a known income source (intraday, f&o, futures, options, delivery, business, capital_gains). Unrecognized sources are blocked instead of silently treated as non-speculative.
InterHeadAdjustmentGuard — Set-off eligibility now runs against an explicit allowlist of heads. Salary loss set-off is added to the prohibition matrix, and unknown heads are blocked rather than defaulting to allow.
GSTGuard (RCM) — Unrecognized service or entity types in reverse-charge checks now surface as a verified=False error naming the unknown value, instead of being coerced to OTHER or INDIVIDUAL and potentially suppressing a statutory RCM obligation. The verifier also gained an optional claim parameter — when you pass a claimed_is_rcm value, the guard compares it to the computed result and only returns verified=True on an exact match. Calls without the claim get a computed_only=True flag so calculation and verification are no longer conflated.
CryptoTaxGuard — Zero VDA income now verifies a claimed tax of zero, instead of returning verified=True regardless of claim. Negative VDA income (a loss) returns verified=False with a message directing the caller to use verify_set_off for loss treatment — the method does not internally invoke verify_set_off, so callers must handle the verified=False branch explicitly.

What this means for you

If your agent relied on any of these guards returning verified=True for inputs they don’t model — ambiguous worker status, unknown set-off heads, unrecognized RCM service types, or capital-gains transactions with malformed dates — those calls now block. Surface the error to a human reviewer or extend the guard’s configured rules before re-running. See the QWED-Tax guards reference for the updated contracts on each guard.

QWED-Tax — fail-closed on unknown tax rules

Released: June 19, 2026

Tax guards no longer silently pass when they encounter a service, asset, jurisdiction, or payment type they don’t model. Six guards now return verified=False with a structured error instead of an unsafe default.

What changed

TDSGuard — unrecognized payment categories no longer return “verified, zero deduction.” This closes a path where an agent could classify a payment into an unknown bucket and have it execute with no withholding. TaxPreFlight now blocks these payments.
CapitalGainsGuard — unknown asset class or holding term fails closed instead of returning “no hard constraint.”
NexusGuard — states not in the configured risk list now require manual review instead of being treated as low-risk.
AddressGuard — unknown state codes fail closed with “manual review required” instead of “assumed valid.”
Form1099Guard (US) — unmodeled payment types return filing_required=None with a manual-determination flag, instead of defaulting to “no filing required.”
InputCreditGuard (GST) — verified=True remains the legal default (ITC is allowed unless specifically blocked), but unknown categories now carry an explicit unverified_category=True flag in the result and an audit_trace entry of category_match: "default_allow" so downstream consumers can distinguish known-eligible from default-allowed.

What this means for you

If your agent currently relies on a verified=True response for inputs the guards don’t model, those calls will start blocking. Add explicit rules for the categories you care about, or route unverified results to human review. See the tax guards reference and tax integration guide for the updated contracts.

QWED-Infra — ecosystem policy framework adopted

Released: June 17, 2026

qwed-infra now ships with the shared QWED governance baseline: QWED_RULES.md, contributor guidance, PR template, CodeRabbit config, and a boundary-check workflow that is consistent with the other QWED repositories.

No runtime behavior changes. Affects contributors and anyone consuming the repository’s CI.

QWED-MCP — boundary-check parity with QWED-Infra

Released: June 18, 2026

The qwed-mcp boundary-check tool now catches the same import-alias, module-alias, wildcard-import, eval/exec alias, and shell=True patterns that qwed-infra does, and fails closed when the scan root is missing.

What this means for you

If you run qwed-mcp boundary checks in CI, expect to catch additional bypass patterns that previously slipped through. Existing passing scans should continue to pass; previously hidden findings may now surface as failures. See the MCP tools reference for the current rule set.

QWED Open Responses v0.3.0 — version sync and dependency fix

Released: June 12, 2026

Aligns qwed-open-responses with the rest of the QWED package versions and patches a transitive qs CVE flagged by Dependabot.

No API changes. Upgrade to pick up the dependency fix.

v5.2.0 — Structured Verification Diagnostics

Released: June 19, 2026 · GitHub Release · Minor

Introduces the unified 3-layer DiagnosticResult model — the diagnostic contract that all QWED verification engines will conform to. This is an additive release: no existing engine return types are changed. Engine conformance is tracked in blocked issues (#129, #130, #131, #133, #134, #162, #163, #164, #190, #205).

New: `DiagnosticResult` model

Three disclosure layers:

Layer 1 — Agent-Safe: agent_message: str — agent/model-facing summary, no internals leaked
Layer 2 — Developer: developer_fields: dict — structured evidence (constraint_id, advisory_checks, methods_used, evidence)
Layer 3 — Proof: proof_ref: Optional[str] — sha256 hash of retained proof artifact; the authority bit

Key design

Tri-state status — VERIFIED / UNVERIFIABLE / BLOCKED only. No HEURISTIC or AMBIGUOUS proliferation; richer distinctions live in developer_fields.constraint_id
proof_ref is the authority bit — present = admissible for control flow, None = reject. No separate authoritative boolean needed (resolves #190 design debate)
VERIFIED requires proof — structurally enforced in __post_init__; “VERIFIED without proof” is impossible to construct
Frozen dataclasses — DiagnosticResult and AdvisoryCheck are frozen=True; post-construction mutation blocked
Advisory checks never influence verdicts — AdvisoryCheck.advisory_only=True enforced via __post_init__
compute_proof_ref() — deterministic sha256 hashing of JSON-serialized evidence
from_legacy_dict() — migration helper for ad-hoc engine dicts (fail-closed states only; raises for legacy VERIFIED)

Version propagation

Artifact	Previous	This release
`qwed` (PyPI)	5.1.2	5.2.0
`qwed_sdk` (Python)	5.1.1	5.2.0
`@qwed-ai/sdk` (npm)	5.1.2	5.2.0
`qwed` (Rust crate)	5.1.2	5.2.0
API version marker	5.1.2	5.2.0
K8s deployment image	5.1.2	5.2.0

Tests

83 new tests covering: status taxonomy, all 3 layers, authority contract, fail-closed enforcement, advisory checks, proof hashing determinism, serialization round-trip, legacy migration, frozen dataclass immutability, and realistic scenarios drawn from the 10 blocked issues.

Compatibility

Additive release. No breaking changes. Existing VerificationResult dataclasses and ad-hoc engine dicts continue to work. DiagnosticResult is opt-in — engines migrate incrementally.

Included PRs

#206 — feat(diagnostics): unified 3-layer DiagnosticResult model (#204)
#207 — release: v5.2.0 version propagation

See the Verification Diagnostics guide for full API documentation.

Older entries: Changelog Archive

​v9.1.0 — Allowlist admission for Python code

​Security fixes

​What changed

​v9.0.0 — Eleven advisories and models you configure

​Security fixes

​What changed

​QWED Verification — list agents, and owners read agent activity

​What changed

​QWED Verification — session sign-in for key rotation and threats, richer key list, and user list

​Added

​Fixed

​QWED Verification — audit CSV exports carry verdict and admission; natural-language audit domain

​What changed

​QWED Verification — qwed init can create an owner who can sign in

​What changed

​QWED Verification — whole-chain trail scope and verifiable Verification Context audit rows

​Fixed

​QWED Verification — startup hooks can be allowed by content digest

​What changed

​QWED Verification — SQL and code verifier false-positive fixes, and Python 3.14

​What changed

​QWED Verification — math identities read diff(), and batch math matches /verify/math

​What changed

​QWED Verification — Claude providers stop sending temperature

​What changed

​QWED Verification — AgentStateGuard gives one verdict on a JSON number

​What changed

​QWED-MCP v0.2.3 — qwedai.com no longer trusted and MCP SDK pinned below 2.0

​What changed

​v8.0.0 — Attested verdicts and upgrade guide

​What’s new

​Also in 8.0.0

​QWED Verification — batch and agent verify routes return enforced, attested results

​What changed

​QWED Verification — audit log paging, caller and admission fields, and security events

​Added

​QWED Verification — Python SDK, integrations, and GitHub Action fail closed

​Security

​QWED Verification — pinned sandbox image and hardened container

​Security

​QWED Verification — CrossHair runs only in the Docker sandbox

​What changed

​QWED Verification — API audit rows written through the signed hash chain

​What changed

​QWED Verification — logic DSL compiles IFF and quantifiers and rejects Boolean coercion

​What changed

​QWED Verification — proven-false math claims are VERIFIED with admission DENY

​What changed

​QWED Verification — GitHub Action scans use mode defaults and deny empty scans

​What changed

​QWED Verification — session sign-in and whole-chain check for compliance endpoints

​What changed

​QWED Verification — security activity summary replaces the SOC 2 report

​What changed

​QWED Verification — role checks on API-key and audit routes, hashed agent tokens

​Security

​QWED Verification — Docker image runs the API server and the Compose stack builds

​What changed

​QWED Verification — Go, TypeScript, and Rust SDKs read the DiagnosticResult envelope

​What changed

​QWED Verification — signed Verification Context documents and attestation checks

​What changed

​Action required

​QWED Verification — contacts and links moved to qwedai.tech

​What changed

​v7.3.0 — Five advisories, tenant scoping, and credential hardening

​Security fixes (advisories)

​Soundness and access control

​Hardening and resource limits

​Credential handling

​SDKs

​Upgrading

​QWED Verification — tenant-scoped admin endpoints and working RBAC

​What changed

​QWED Verification — verify_fact fails closed offline and Money rejects non-finite amounts

​What changed

​QWED Verification — request body caps, timeouts, and sandbox concurrency limits

​What changed

​v7.2.3 — Code, secret, and SQL verification security release

​Security fixes

v9.1.0 — Allowlist admission for Python code

Security fixes

What changed

v9.0.0 — Eleven advisories and models you configure

Security fixes

What changed

QWED Verification — list agents, and owners read agent activity

What changed

QWED Verification — session sign-in for key rotation and threats, richer key list, and user list

Added

Fixed

QWED Verification — audit CSV exports carry verdict and admission; natural-language audit domain

What changed

QWED Verification — `qwed init` can create an owner who can sign in

What changed

QWED Verification — whole-chain trail scope and verifiable Verification Context audit rows

Fixed

QWED Verification — startup hooks can be allowed by content digest

What changed

QWED Verification — SQL and code verifier false-positive fixes, and Python 3.14

What changed

QWED Verification — math identities read diff(), and batch math matches /verify/math

What changed

QWED Verification — Claude providers stop sending `temperature`

What changed

QWED Verification — AgentStateGuard gives one verdict on a JSON number

What changed

QWED-MCP v0.2.3 — `qwedai.com` no longer trusted and MCP SDK pinned below 2.0

What changed

v8.0.0 — Attested verdicts and upgrade guide

What’s new

Also in 8.0.0

QWED Verification — batch and agent verify routes return enforced, attested results

What changed

QWED Verification — audit log paging, caller and admission fields, and security events

Added

QWED Verification — Python SDK, integrations, and GitHub Action fail closed

Security

QWED Verification — pinned sandbox image and hardened container

Security

QWED Verification — CrossHair runs only in the Docker sandbox

What changed

QWED Verification — API audit rows written through the signed hash chain

What changed

QWED Verification — logic DSL compiles IFF and quantifiers and rejects Boolean coercion

What changed

QWED Verification — proven-false math claims are `VERIFIED` with admission `DENY`

What changed

QWED Verification — GitHub Action scans use mode defaults and deny empty scans

What changed

QWED Verification — session sign-in and whole-chain check for compliance endpoints

What changed

QWED Verification — security activity summary replaces the SOC 2 report

What changed

QWED Verification — role checks on API-key and audit routes, hashed agent tokens

Security

QWED Verification — Docker image runs the API server and the Compose stack builds

What changed

QWED Verification — Go, TypeScript, and Rust SDKs read the DiagnosticResult envelope

What changed

QWED Verification — signed Verification Context documents and attestation checks

What changed

Action required

QWED Verification — contacts and links moved to `qwedai.tech`

What changed

v7.3.0 — Five advisories, tenant scoping, and credential hardening

Security fixes (advisories)

Soundness and access control

Hardening and resource limits

Credential handling

SDKs

Upgrading

QWED Verification — tenant-scoped admin endpoints and working RBAC

What changed

QWED Verification — `verify_fact` fails closed offline and `Money` rejects non-finite amounts

What changed

QWED Verification — request body caps, timeouts, and sandbox concurrency limits

What changed

v7.2.3 — Code, secret, and SQL verification security release

Security fixes