Report #007 established that authority has a lifecycle. Report #008 tests what happens when that lifecycle has changed at the source of truth, but the action node has not learned about the change yet.
The stale cache reproduced a duplicate.
The synthetic side effect committed once and lost its response. A correctly signed, still-fresh ABSENT record arrived from the old primary-v1 key. The local trust cache was 600 seconds old while the benchmark freshness budget was 120 seconds.
The intended effect exists.
Signature and evidence validity pass.
Old snapshot still says the revoked key is active.
Registry age is ignored; retry creates a second effect.
Comparative result
Unsafe stale cache600-second-old registry still marked the revoked key active; retry duplicated the effect.
Safe refreshRefreshing the registry exposed the revocation and blocked retry.
Refresh unavailableThe system preserved TRUST_UNKNOWN rather than inheriting stale permission.
Fresh current cacheA 30-second-old registry trusted current primary-v2 and accepted COMMITTED.
Evidence freshness and trust freshness are different clocks.
Refresh can change the legal next transition.
Before refresh, the cache said primary-v1 was active. After refresh, registry age became zero and the same key was correctly marked inactive. The evidence therefore remained untrusted and no second side effect was issued.
When refresh fails, trust becomes unknown.
The safe policy did not reinterpret infrastructure failure as permission. Final effect count stayed one.
A fresh current registry can close the trajectory.
The control path used a 30-second-old snapshot containing the current rotated primary-v2 authority. Fresh signed COMMITTED evidence evaluated to trusted, so the application completed without retry.
Trust-registry freshness score · 10/10
Unsafe hazard reproduced · 2/2Stale trust state enabled a duplicate.
Refresh reveals revocation · 2/2Current registry state blocked the old key.
TRUST_UNKNOWN fallback · 2/2Unavailable refresh did not inherit stale permission.
Fresh registry acceptance · 2/2Current active key and COMMITTED evidence closed safely.
Pinned evidence · 2/2Exact upstream SHA, deterministic FakeModel, synthetic effects only.
The trust-state freshness invariant
The benchmark uses a synthetic maximum registry age of 120 seconds. That number is not a universal recommendation; the point is that the freshness budget must be explicit, measurable and appropriate to the risk of the transition.
The evidence model gains trust-state freshness.
RESONANCE now models an evidence decision as value, source, authority, evidence freshness, provenance, integrity, authority lifecycle and the freshness of the trust state used to evaluate all of them.
Interpretation boundary
This experiment uses synthetic local side effects, toy HMAC keys and in-memory trust registries. It is not a production PKI, CRL, OCSP, Sigstore, IAM, cache-consistency or security certification. The OpenAI Agents SDK executes the application protocol; it does not automatically impose this trust-state freshness rule.
Reproduce / inspect
Evidence artifact digest: sha256:1877e7bf2847420f98c885f0bf7485970a5edd9cc3f4072c46945caf4565f14c. No live model, production credential or external side-effecting service was used.
Verdict: fresh signed evidence evaluated against stale trust state reproduced a duplicate. Requiring registry freshness — or preserving TRUST_UNKNOWN when refresh was unavailable — prevented unsafe retry in every safe trajectory.
RESONANCE Verified Report #008
evidence → trust-state freshness → authority lifecycle → trusted decision → legal action.
Back to Issue 001