This project began as Absolute Self-Governance (ASG): a council of role-based agents voted on which roles a software task needed, team size was dimensioned from measured task complexity, and specialists executed against a repository’s own tests. It was built, benchmarked across five sweeps and a 589-round pre-registered study — and its own headline result argued against the council it was named for. A pre-registered analysis of the roster-voting mechanism found no significant benefit: +3.9 points, 95% CI [−2.7, +10.6], p = 0.247.
The project was renamed and reduced to what the evidence actually supported. What survived is Escalation-Gated Retry: try once, and only after two consecutive test failures, feed the real failure back and try again with a fresh perspective — the one mechanism the benchmarks found working. Consensus, when it happens at all, is earned by evidence of failure, never assumed by default.
The repository is now three things: a coding-assistant skill implementing Escalation-Gated Retry directly (no server, no API key, no orchestration runtime); a 14-page paper leading with the null result and walking through the three benchmark passes that produced it; and the underlying telemetry, pre-registrations, and task definitions every number in the paper traces back to, recomputable from raw data with the standard library alone.
The always-on, role-voting runtime this project started as still exists,
tagged v0.1.9-full-runtime under its original name, for anyone who wants
to reproduce the original sweeps. It is not in the current tree.