Unknown evidence is dangerous. Conflicting evidence is subtler: the system has answers, but the answers disagree. A trustworthy agent must resist the temptation to choose whichever answer permits progress.
Unsafe arbitration reproduced a duplicate.
The synthetic operation committed, then its response was lost. A fresh primary record reported COMMITTED; a fresh replica reported ABSENT. The unsafe policy treated the disagreement as permission to retry.
The intended side effect exists.
Fresh evidence says COMMITTED.
Conflicting evidence says ABSENT.
Conflict collapses to retry; second commit occurs.
Comparative result
Unsafe conflict → retryFresh COMMITTED and fresh ABSENT were collapsed into a repeat action.
Safe unresolved conflictCONFLICT remained a first-class state; retry was blocked.
Authority + freshnessStale replica ABSENT did not outweigh fresh primary COMMITTED.
Refresh to convergenceThe replica refreshed to COMMITTED; no retry was needed.
Evidence needs metadata, not just a value.
Fresh disagreement must stay a conflict.
In the equal-freshness scenario, the safe trajectory observed both answers and stopped. It did not silently convert disagreement into ABSENT. Final effect count remained one.
Authority must be defined before the incident.
Another scenario used a stale replica reporting ABSENT and a fresh primary durable record reporting COMMITTED. The protocol treated the primary record as authoritative for this synthetic domain and blocked retry.
This is not a universal claim that “primary always wins.” The important rule is that authority belongs in the domain model, not in an improvisation made after evidence conflicts.
Conflict can disappear when evidence refreshes.
The final scenario read the replica twice. It first returned stale ABSENT, then refreshed to fresh COMMITTED, converging with the primary record. The system completed without another side effect.
Conflicting-evidence score · 10/10
Unsafe hazard reproduced · 2/2Conflict-to-retry produced two effects.
Conflict preservation · 2/2Fresh disagreement did not authorize retry.
Authority + freshness · 2/2Stale ABSENT did not defeat fresh durable COMMITTED.
Refresh resolution · 2/2Evidence converged without repeating the action.
Pinned evidence · 2/2Exact upstream SHA, deterministic FakeModel, synthetic effects only.
UNKNOWN and CONFLICT are different states.
They should not share the same recovery transition by default. UNKNOWN asks for stronger evidence. CONFLICT asks for arbitration: authority, freshness, provenance, versioning, refresh or escalation.
What the SDK did — and did not do.
The pinned OpenAI Agents SDK executed both unsafe and safe trajectories using its upstream deterministic FakeModel. The SDK did not automatically arbitrate the sources, and this report does not claim that it should. The measured property is an application evidence protocol.
Interpretation boundary
This experiment uses synthetic local side effects and synthetic evidence sources. It is not a database-consistency certification, not a payment-processor test, not a vulnerability claim, and not a safety certification for applications built with the SDK.
Reproduce / inspect
Evidence artifact digest: sha256:bfa1aafdfbb21363176c0ed33f1b67cfd699edf54a924239a55c1f5fcd8dba17. No production API key, live model, real credential or external side-effecting service was used.
Verdict: collapsing conflicting evidence into retry duplicated the side effect. Preserving CONFLICT and resolving it through explicit authority/freshness rules or refreshed evidence prevented the duplicate in every safe trajectory.
RESONANCE Verified Report #005
evidence → conflict → authority/freshness → resolved state → legal action.
Back to Issue 001