v0.3.0

Release gate benchmark

Generated by npm run gate from the floors in evals/gate.json. Deterministic: no timestamps, no
model calls. CI regenerates this file and fails if the committed copy differs, so it cannot drift
from the code it describes. npm run gate:drill proves the gate blocks a deliberate regression.

Check	Result	Measured	Floor
eval-suite	PASS	9 checks passed (exit 0)	>= 9 checks, exit 0
guard-coverage	PASS	9/9 caught	>= 9 mutations, all caught
refusal-correctness	PASS	incomplete PO refused, complete PO passes, injection minted nothing, performed nothing, and its denial names the rule that refused	verifier exits as specified; no authorization or effect under injection
replay-determinism	PASS	2 runs, byte-identical ledgers	2 identical ledgers
audit-replay	PASS	audit exit 0, tamper-check exit 0	both exit 0

Release allowed.

Assay v0.2.0 — deterministic control layer reference

Assay is now a release-ready reference implementation of a deterministic control layer around probabilistic AI-agent reasoning.

Highlights:

• Binds write authorization to the recomputed verdict outcome, verdict hash, and complete evidence summary.
• Fails closed on malformed or substituted verdicts.
• Demonstrates typed decisions, schema validation, policy enforcement, approval tokens, tamper-evident receipts, replay protection, and deterministic hooks.
• Includes a reproducible synthetic corpus, 71 unit tests, 12 end-to-end verification steps, and 8 mutation tests proving the controls can fail.
• Adds Apache-2.0 licensing, governance, security, contribution, issue, and pull-request policies.
• Adds Node 20/22 CI, CodeQL v4, pinned GitHub Actions, Dependabot, and private vulnerability reporting.
• Updates Zod to 4.6.4.

Boundaries:

• This is a reference implementation using synthetic data, not a production claim.
• It does not claim external adoption, customers, vendor endorsement, or assurance beyond the documented threat model.
• A successful refusal is an expected outcome in the denial demonstrations.