Pre-Action Governance Conformance Registry

A neutral, continuously-updated check of which agent-governance verifiers actually satisfy the WYRIWE pre-action invariants — against their live endpoints, recomputed from public bytes, no trust in the verifier (or in us) required. The board reports which invariant each mechanism actually satisfies; it does not bless a mechanism. Anyone can reproduce every result with the open suite: github.com/babyblueviper1/preaction-governance-conformance.
VerifierRecompute depthCanonical envelopeIndependent admissionAnchor existencePre-outcome precedenceAction chainCertified
agentoracle · live
ed25519-jcs
● terminal✓ pass✓ pass
2 signers verified
✓ pass
via on-chain (Arbitrum)
✓ pass
via on-chain (Arbitrum)
— n/aCERTIFIED
invinoveritas · live
nostr-event
● terminal✓ pass✓ pass
embedded
✓ pass
via Bitcoin OTS
✓ pass
via Bitcoin OTS
— n/aCERTIFIED
moyan-pmi · live
ed25519-jcs
◐ bounded · anchor✓ pass✓ pass
embedded
⏳ pending⏳ pending— n/anot yet
safeagent · live
bip340-schnorr
◐ bounded · anchor✓ pass✓ pass
embedded
⏳ pending⏳ pending— n/anot yet

Graded benchmarks

Submitted artifacts recomputed from their own bytes at a pinned commit — a different axis from the live-endpoint verifiers above (no WYRIWE invariant cells, because neither kind is a live /governance endpoint). For a graded trace dump: CERTIFIED = the verdict survives the row’s own recorded jitter (|score−threshold| > jitter_margin); PARTIAL = inside the band; DEPTH-0 = fault / non-recomputable. For a deterministic vector: every linked identifier either recomputes byte-for-byte from its own preimage or it doesn’t — no jitter concept applies. A grade against a pinned commit stays true forever, but the project can move past it — so every run (1) checks how far the default branch has drifted from the pin and (2) where we have a registered reverify recipe, re-pulls evidence and recomputes it fresh — a full re-derivation for a small artifact (ibex), a cheap byte-range sample for a large one (Correctover’s 20k-row dump is canary-sampled, not fully re-graded every cycle — a full re-grade is heavier and could silently flip a published % without a human look, so the sample answers "is this still alive and consistent," not "re-publish the percentage"). Drift is a signal, not a verdict; neither check ever silently changes a published grade. Recomputed by us, reproducible by anyone at the pinned commit.
BenchmarkRecompute-depth gradePinned commitFreshness
Correctover v1.1.0
graded trace dump · 20071 rows · recipe examples/benchmark-grade/grade_traces.py
26.5% certified · 52.8% partial · 20.7% depth-0a27999b333drift unknown
reverify: unreachable
checked 2026-07-01
ibex-agent-verification verifiable-action-chain v1
deterministic vector · 1 record(s) linked · recipe conformance/verifiable-action-chain-v1.json + tests/test_full_chain_vector.py
RECOMPUTES 4/4
action_id → decision_id → execution_outcome_id → audit_record_id
814d670c88drift unknown
reverify: pass
checked 2026-07-01
Recompute depth = how far a verifier's evidence recomputes from public bytes before a step taken on trust. terminal recomputes through an external anchor on an independent clock; bounded · anchor recomputes independent admission but the anchor is not yet confirmed; bounded · admission recomputes the envelope but admission rests on the producer; attested means the envelope does not recompute from its own bytes. Rows are ordered by this tier, then alphabetically — both re-derivable from the suite states, so the order is not our opinion (the alphabetical tiebreak is why we are not first within a tier).
Invariants. Canonical envelope: SHA-256 of the verifier's declared canonical bytes recomputes to its committed hash. Independent admission: an identity that is not the acting agent signed that hash; the sub-label is key provenance — embedded when the verification key is in the fixture, or k/N signers verified for a multi-signer claim, so a row never implies it is fully recomputable from the fixture when only some signers are. Anchor existence: the commitment is externally confirmed by the declared mechanism (it provably exists). Pre-outcome precedence: the accepted anchor point provably precedes the terminal outcome — a confirmed-but-backfilled stamp proves existence, not ordering, so existence and precedence are reported separately and existence can never pass as ordering. Action chain: the terminal record joins back to the same proposed action (n/a where an endpoint is pre-action only). pending is honest, not a failure.
Certified = the three core invariants — canonical envelope, independent admission, and pre-outcome precedence — currently passing, with nothing in a hard fail. We report our own pending/gaps the same way we report everyone's; the neutrality is the point.
Want your verifier listed, or to confirm a vendor is conformant? See the reference /ledger or open an issue on the suite repo. Last run: 2026-07-01T02:20:48.049052+00:00.