Action Preserving Compression

Action-Preserving Compression

Compress experience only as far as action survives. The unit is not a principle; it is a handle that still chooses the right move.

Output contract

Return one block per candidate:

Handle: 
Target: 
Equivalence: 
Move: 
Probes: 
Counterexample: 
Refinement: 
Receipts: 
Disposition: candidate | strengthen | split | merge | retire | discard | catalog

Do not omit Counterexample. A handle without a searched-for failure is only a mnemonic.

Procedure

1. Decompress. Collect concrete cases, outcomes, attempted moves, and context until consequential distinctions are visible.

Done when: at least two cases share a possible move, and at least one boundary or counterexample has been searched.

1. Compress. Write the output contract fields. Prefer one handle, one target decision, one move.

Done when: a stranger can use the handle to choose an action without re-reading the source material.

1. Break. Test boundary cases. Ask where the same handle would choose the wrong move.

Done when: either a counterexample is recorded, or the search scope is named and Counterexample: not found.

1. Refine. When one handle contains situations requiring different moves, split on the minimum distinguishing feature and repeat.

Done when: the split changes the move and removes the counterexample.

1. Stop. Delete distinctions one at a time.

Done when: deleting any remaining distinction changes the move.

Rule: decompress until distinctions change action; compress until one more deletion would change it.

Catalog and review

For a one-off derivation, stop at the output contract.

When the request says catalog, review, strengthen, split, merge, retire, or update, read [HEURISTICS.md](HEURISTICS.md) and follow its admission rules.

Catalog work must:

• compare by target, equivalence, move, probe, and counterexample, not by handle name
• preserve old handles through the mutation ledger instead of silently rewriting them
• use git history as the rollback path; do not create a second versioning system
• keep unproven handles as candidate until receipts show repeated competent action

Integration points

• Daily audits: mine .notes/audit/*/learning-candidates.jsonl, audit-findings.md, and pattern-ledger.jsonl for cases and counterexamples. Do not promote daily-audit candidates without re-running this procedure.
• Git: use commits/diffs as receipts when a heuristic changed an action or when a split fixed a failure.
• Eval loops: when a heuristic can be tested repeatedly, freeze a small evaluation set before comparing revisions. Borrow the autoresearch pattern: fixed metric, append-only log, reviewable diffs.
• Distill-law: use $distill-law only when the candidate claims to be a falsifiable system law. Most action handles should remain heuristics.

References

• [HEURISTICS.md](HEURISTICS.md) - standing catalog, admission rules, and mutation ledger

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: TylerJNewman
• Source: TylerJNewman/skills
• License: MIT

Install and usage instructions live in the source repository linked above.