idkmesh

Issue #97 Selection and Reference-Evidence Completion

Date: 2026-08-29 Scope: Select one currently unclaimed issue, complete it through a bounded pull request, and integrate only after current exact-head checks pass.

Project-owner requirement

Select an issue that no other contributor is currently working on, solve it professionally, open a pull request, and merge the verified result into main.

Selection evidence

Issue #97 was selected after checking current GitHub and repository state:

• it had no assignee and no issue comments;
• neither open PR (#159 or #242 at selection time) overlapped R4;
• PR #100 had already merged the executable R4 harness;
• PR #101 was closed because it contained only a failing one-shot workflow and none of the result files claimed by its description;
• the old experiment/r4-reference and biology-aco-stigmergy-v0 branches were stale historical refs, not active review surfaces.

The issue was publicly assigned before implementation to prevent duplicate work. The replacement branch started from current main@5123f8c1f47c4ec5b1982863d0e90c8bf8b00232; no stale branch was merged.

Work completed

The documented default and lock-in scenarios were regenerated from the merged harness with full per-step events. The change publishes:

• both raw machine-readable traces;
• a readable comparison with exact commands and SHA-256 provenance;
• fixed artifact-digest and cross-runtime replay tests, with byte-for-byte regeneration on the recorded Python 3.12 runtime family;
• status/documentation updates that preserve the synthetic-only boundary.

Findings

• evaporation plus exploration outperformed permanent pheromone in both frozen scenarios;
• permanent pheromone never sampled the strong late expert in the lock-in trap;
• Thompson sampling slightly led realized success in the default trace and decisively beat every stigmergic policy in the lock-in trap;
• every stigmergic run recorded zero positive pheromone change from unverified activity.

The conventional-bandit win is retained as a negative result. The evidence does not authorize production routing, contributor scoring, governance power, or automatic acceptance.

Community impact

The issue is no longer a misleading open implementation candidate, and future contributors can reproduce or challenge the evidence without recovering a failed workflow or stale branch. Follow-up parameter sweeps remain separable research rather than an implicit requirement for closing this bounded issue.

AI/tool provenance

Codex selected the bounded issue, generated the deterministic artifacts with the repository CLI, prepared the analysis and tests, and ran the documented verification. GitHub exact-head checks remain the integration gate.

This site is open source. Improve this page.