Scherbehuufe (proximity-prize-frontier-archaeology)
Use this skill after an accepted better.codes frontier move or when comparing a competitor's promoted proof with its immediate accepted predecessor. This is incumbent-history research. Never place its source, receipts, conclusions, route hints, or terminology in a clean-room re-descent packet.
Bind the accepted pair
- Refresh the live benchmark and record the accepted score, accepted source commit, benchmark ID, track, and observation time.
- Resolve the immediately preceding accepted source from the benchmark or repository lineage. Do not substitute a convenient older checkpoint.
- Work in a read-only clone containing both commits. Confirm the benchmark contract tree and editable submission path.
- Save the public PR or submission note locally when one exists. Preserve its URL and retrieval time in the surrounding campaign receipt; the helper binds the exact note bytes.
- Run:
python3 scripts/frontier_archaeology.py capture \
--repo /absolute/path/to/repo \
--older PREDECESSOR_COMMIT \
--newer ACCEPTED_COMMIT \
--submission-dir ProximityPrize/SubmissionLower \
--note-file /absolute/path/to/public-note.md \
--output /absolute/path/to/frontier-delta.json
The command fails if the pair is non-ancestral, if another score-changing commit lies between the two, or if the newer ref is not itself the next score promotion. Use a separately reviewed explicit lineage receipt for histories that cannot satisfy those checks; do not weaken the default.
Inspect the mechanism
Read the receipt and the exact diff, then answer in this order:
- What score and radius changed?
- Did
Solution.leanswitch proof roots or only retune the same root? - Which imported modules, theorem providers, and constants changed?
- What does the public note claim, and which changed lines actually support it?
- What is inherited infrastructure, what is the new lever, and what is packaging or cleanup?
- Can the lever transfer to the current accepted tree without importing a dead route or falsified assumption?
Treat generated transfer_review fields as an audit form, not an automatic conclusion. Fill each claim with file-and-line evidence. A module name, commit subject, or PR prose alone does not establish the mechanism.
RCI loop
Run one bounded loop per promoted pair:
- Inspect: direct-predecessor receipt, exact diff, proof root, public note, and live campaign obstruction ledger.
- Critique: name the smallest unsupported mechanism claim, lineage ambiguity, or hidden inherited dependency.
- Improve: narrow the claim or isolate one executable falsifier for the alleged lever.
- Verify: replay the receipt and run the falsifier against both predecessor and accepted source.
python3 scripts/frontier_archaeology.py check \
--receipt /absolute/path/to/frontier-delta.json
Only route a lever onward when the evidence distinguishes it from accompanying churn. Record TRANSFERABLE, SOURCE-BOUND, FALSIFIED, or UNRESOLVED; never silently convert UNASSESSED into a strategy recommendation.
Read references/receipt-and-review.md before publishing a lineage conclusion or dispatching a lever to another lane.