Skip to content

Releases: goun7/veridict

v0.5.0 — machine-checked ladder + SPDX 3.0.1 export

Choose a tag to compare

@goun7 goun7 released this 15 Sep 15:21

Ledger format, payload schemas and tier rules: byte-identical to 0.4.0. New: the adjudication ladder is machine-checked in Lean (proofs/ladder/: 7 arbitrary-length invariants + a 28,080-case build-time oracle whose first run caught a model/code divergence and made the model move); veridict export --format spdx emits SPDX 3.0.1 AI-profile JSON-LD validated against the pinned official schema; the Marketplace action gained anchor/anchor-required inputs with e2e proof through uses: ./; veridict --version reports the truth (the installed-wheel version probe had been querying a distribution name that never existed). Assets: the ledger, certificate and Rekor sidecar this release was dogfooded on — check every layer offline: veridict verify --ledger dogfood_ledger.jsonl --cert dogfood_cert.json --anchor receipts-anchor-v050-dogfood.json (exit 0 = signature, hash chain, AND transparency-log binding against a pinned key — no trust in us required).

veridict 0.4.0 — research-sweep release (jury calibration, Rekor anchoring, SLSA export, ladder proofs)

Choose a tag to compare

@goun7 goun7 released this 15 Sep 10:34

What changed

The research-sweep release: seven grounded gaps from the 2026-09-14 literature/regulatory review, closed in dependency order — additive minor; the ledger format and digest preimages are untouched (§14).

  • Jury calibration layer — structural mitigations for the three documented LLM-as-judge failure modes (position/self-enhancement bias, verbosity bias, self-preference): multi-sample majority voting with honest split-abstention, evidence-first extraction with validation, and author-family exclusion (new actor-family action input). Cited in-module: arXiv:2306.05685, 2305.17926, 2410.21819.
  • External anchoring to Sigstore Rekor — the design doc's checkpoint-hardening promise ships: audit --anchor rekor, anchor publish/verify, verify --anchor. Offline verification against the TUF-pinned Rekor key; existence-at-time proof without transferring authority (anchor key ephemeral by design).
  • SLSA v1.2 VSA exportexport --format vsa projects any certificate onto an in-toto verification_summary/v1 statement for supply-chain tooling; honest lossy mapping + spec-sanctioned extension binding; optional DSSE envelope.
  • Adjudication ladder bounded-exhaustive verification — 28,080 cases covering the complete finite input space; 10 safety invariants PASS (no-silent-pass, W1a-decisiveness, doctrine-cannot-topple, split-visibility, determinism, …); CI locks the truth-table digest; a meta-test proves the checks see an injected silent-pass bug. Issue #8 annotated; general proofs stay open.
  • EU AI Act transparency watcher (11th example) — deterministic Art. 50 / GPAI Code-of-Practice surface checks, vacuous-SUPPORTS for non-AI repos.
  • AuditWeave positioning (arXiv:2607.09682) in design-doc §10.5 + jury-literature citations; IP & prior-art watch section in GOVERNANCE.md (US App 19/561,229 assessed against RFC 6962 + intervening art; not legal advice).

Receipts

  • Test suite: 297 tests, full CI green on 51632d3 (3.12/3.13/3.14 + dogfood self-audit + badge regen + real-jury receipt with 3 samples)
  • Rekor: v1's dogfood certificate anchored live before this release — docs/receipts-anchor-v1-dogfood.json (logIndex 2843194981); this release's dogfood cert is anchored too (sidecar attached here)
  • Ladder proof receipt: docs/receipts-ladder-verification.json (truth table sha256 d0ab9978…)
  • PyPI: veridict-standard==0.4.0 via Trusted Publishing (tag v0.4.0)

Notes for Action users

@v1 will be re-pointed at this release (the actor-family input landed after the previous v1 pin). New optional inputs work immediately; everything else is compatible.

0.3.3 — Marketplace-ready action + CI on Node-24

Choose a tag to compare

@goun7 goun7 released this 14 Sep 07:39

What changed

  • action.yml moved to the repository root (audit//). GitHub Marketplace only lists actions whose action.yml lives at the repo root — the 0.3.2 layout worked via uses: but could never be listed. Invocation is now uses: goun7/veridict@v1; the v1 major-tracking tag points at this commit. Closes L0-2 of roadmap issue #10 for real (0.3.2 claimed it prematurely — the listing URL 404'd).
  • All actions/* CI steps upgraded to Node-24-native majors (checkout v7, setup-python v7, upload-artifact v7, download-artifact v8, upload-pages-artifact v4, deploy-pages v5) — clears the Node-20 deprecation warnings every 0.3.2 run carried. Verified: this release's CI run carries zero deprecation annotations.

Receipts

  • CI run 34817690404: success (260 tests across 3.12/3.13/3.14, dogfood self-audit, badge regeneration) — zero Node-20 warnings
  • PyPI: veridict-standard==0.3.3 published via Trusted Publishing (run 34818699653, success)
  • v106fb990

Maintainer step (manual, ~2 min)

Publish the Marketplace listing: repo → right sidebar Actions (or https://github.com/goun7/veridict/actions) → find the "veridict audit" action draft → Publish to GitHub Marketplace. The draft exists only now that action.yml is at the root; it cannot be done via API.

0.3.2 — launch follow-through

Choose a tag to compare

@goun7 goun7 released this 13 Sep 22:16

0.3.2 — launch follow-through

  • Composite audit action (audit/action.yml) — the audit as a Marketplace-listable GitHub Action: uses: goun7/veridict/audit@v1 (reusable workflows can't be Marketplace-listed; this can)
  • v1 major-tracking tag — the README's @v1 Action reference now resolves (was a broken ref before)
  • publish-pypi trigger hygiene — semver-tags-only (v*.*.*); the moving v1 ref can never fire a publish
  • Brand: V-mark logo (SVG), full favicon set, 1200×640 social card; Pages carries the favicon, READMEs carry the mark
  • Adoption: GitHub Discussions inbox (open, seeded with the adoption-support thread); repo topics expanded
  • Docs: plain-language tier bullet; PyPI badge + pip install veridict-standard lead in all launch texts
  • PyPI: 0.3.2 published via Trusted Publishing (OIDC) — no tokens stored anywhere

Receipts: 260 tests × 3 Python versions, dogfood certificate valid + offline verify rc 0, badge regenerated by CI.

v0.3.1 — adoption surface + hardening night

Choose a tag to compare

@goun7 goun7 released this 11 Sep 02:09

Contracts for external implementers (JSON Schemas validated against every real artifact, watcher+ladder conformance vectors), the complete registry CLI (init/register/revoke/list/index with key files), audit --registry and --watcher wiring, signature-format interop proof, certificate mutation differential (caught + fixed spec-verifier crash, errata D8), real-jury provider hardening (keyless crash fixed), canary corpus 15/11, nightly soak + Scorecard + public docs site, and errata D9: the static analyzer no longer walks hidden directories — when it did, the §6 fail-closed gate correctly blocked a release on third-party findings, proving the gate fails on the safe side. 220 tests, CI green on 3.12–3.14.

v0.3.0 — Phase 3 open-source substrate

Choose a tag to compare

@goun7 goun7 released this 10 Sep 15:40

See CHANGELOG.md for the full list.

Highlights:

  • Chain format binds every stored field (incl. ISO-8601 string timestamps) — a fuzz-caught retroactive-timestamp-edit gap is closed
  • Standard v1.0.0-draft with normative ladder, divergence formula, replay-scope guard, and a public errata ledger (§14.2)
  • Conformance test vectors — deterministic, language-neutral, byte-pinned; regenerate with scripts/build_test_vectors.py
  • Spec-only verifier (examples/spec_verifier.py) — implemented from the standard alone, zero imports of the reference code; parity with the reference verifier pinned by cross-implementation fuzz
  • Canary corpus: 6 defect classes, 9 cases — 5 catches, 2 honest misses, 0 false positives (a quality sheet that catches everything is a rigged one)
  • CI runs the full receipt set on every push (Python 3.12–3.14)