Engineering-loop evidence

Status: experimental Last updated: 2026-08-05 Sources: local-engineering-loop-measurement, ecc-loop-controls

Current understanding

The engineering loop defines useful completion evidence: baseline or reproduction, a coherent change, focused checks, required checks, final diff review, and a handoff that records uncertainty. The repository does not treat the existence of that workflow as proof that its skill is more efficient.

Evidence

The measurement protocol compares three variants with the model, reasoning effort, repository state, and done conditions held constant:

1. no skill;
2. the earlier full skill; and
3. the current lean skill.

It records quality gates, retries, context noise, elapsed time, and tokens when available.

Implications

Examples and demos establish that the workflow is runnable, not that it generalizes. Team recommendations require repeated measurements on representative work.

Promotion status

The workflow is published, while its efficiency claim remains experimental. Future run data should update this page before changing the recommendation.

For repeated autonomous tasks, the loop now defines an optional judge, iteration and retry budgets, stop-loss rules, and a resumable checkpoint. Measurements should include stopped runs, first-judge pass rate, and iterations to green so one eventual success is not mistaken for reliable control.