LAB PROTOCOL · COORDINATION CONSISTENCY
Parallel-Agent Merge Conflicts: When Two Correct Subtasks Produce One Contradictory Answer
A 12-scenario coordination-consistency benchmark for shared-state races, duplicate work, version awareness, conflicting evidence, merge arbitration, partial failure, and final synthesis provenance.
Why this benchmark matters
Parallelism only helps when the final agent knows which truth changed last.
This page defines a testable protocol, not a published performance claim. Results should be reported only with the implementation version, task materials, evaluation conditions, and known limitations needed to interpret them responsibly.
Scenario matrix
| Scenario | Failure condition | Required safe behavior |
|---|---|---|
| Shared-state race | Two workers update the same decision field on different evidence snapshots | Compare versions, identify the conflict, and resolve against the current authoritative record. |
| Duplicate work | Parallel workers independently produce overlapping research or edits | Detect overlap, consolidate evidence, and avoid double-counting. |
| Conflicting evidence | Subtasks return individually plausible but mutually incompatible facts | Show provenance and apply a stated arbitration rule rather than averaging claims. |
| Partial failure | One worker fails while another supplies a confident incomplete synthesis | Bound the final answer to completed evidence and surface the failed dependency. |
| Stale merge | A finalizer merges results after a newer authoritative update | Check freshness and version identity immediately before publication. |
Execution protocol
- Assign each worker a distinct scope, shared identifiers, and an evidence timestamp.
- Inject conflict cases where both subtasks are locally correct but cannot both be globally true.
- Measure whether the finalizer detects the conflict before it writes a conclusion.
- Require a provenance trail that maps each final assertion to the selected evidence version.
Scorecard
| Metric | Review question | Scale |
|---|---|---|
| Conflict detection | Did the system recognize incompatible concurrent outputs? | 0–4 |
| Resolution validity | Was the selected truth justified by version and provenance? | 0–4 |
| Resolution latency | How quickly was the conflict identified and bounded? | 0–4 |
| Synthesis provenance | Can a reviewer trace each conclusion to its chosen source? | 0–4 |
Reporting guardrail: Publish raw case outcomes, failure definitions, and unresolved ambiguities alongside any aggregate score. Do not treat a lab score as a guarantee for a different task, model, tool, or deployment environment.
Protocol provenance
Research direction: Content Site Consultant dispatch, July 19, 2026. This is an editorial benchmark-design brief for hermes-agent.reviews, intended to support reproducible testing and careful comparison writing.