Loop-de-loop

Loop-de-loop is a loop hardener.

Give it an approved loop, and it tries to make that loop better by testing the original against challenger versions. A challenger only wins if it preserves the task, keeps the proof bar, respects side-effect limits, and improves something real: quality, reliability, cost, speed, verification, or safety.

It is not a prompt-shortener. A shorter loop can lose. A longer loop can win. The point is better real execution, not prettier wording.

The full loop is in LOOP.md.

What It Checks

• Is the loop approved to be hardened?
• What can the loop safely touch?
• What counts as a real win?
• What benchmark, scorecard, and proof are needed?
• Did the challenger actually improve the workflow?
• Did the improvement survive held-out or fresh checks?
• Should the challenger be promoted, rejected, measured again, or left as no promotion?

What To Watch For

Loop-de-loop may run for a long time if the target loop needs real measurement. It may also stop quickly if the target is not approved, cannot be measured safely, has no meaningful improvement path, or would require fake evidence.

If you are new to loops, watch the run closely. The safest result is not always promotion. NO PROMOTION, INCONCLUSIVE, and NEEDS_MEASUREMENT are valid outcomes.

Provenance

• Full loop: LOOP.md
• SHA-256: see CHECKSUMS.txt
• Public timestamp: this repository's commit history

Rights

All rights reserved unless a separate license is added later.