Checkpoint Surgery — artifact

Evidence archive for Checkpoint Surgery: Eval-Gated, Agent-Proposed Tensor Repair of Trained Models

Every number in the paper has a lane report behind it. This archive is those lane reports, plus the scorers that turn them back into the paper's tables. No weights are redistributed — what ships is manifests, hashes, diffs, eval outputs and scripts.

Download archive ↓

size: 2.3 MB
sha256: 1d2871160e078a207ef472efe484af8bdc911587b73fc6206f90ab267be7c4ec
doi: minted at camera-ready

Reproducing each result

CPU-only. No GPU, no checkpoint download. Each scorer prints the table's numbers rather than a summary, so you can disagree with a scorer's choices without re-renting anything.

Paper table	Lane report	Scorer
tab:c1a-arms, ledger P-C1	sprint1-c1a/LANE.md	gpu-c1a-score.py
tab:pb2	sprint2-pb2/LANE.md	pb2_score.py
tab:c2a	sprint2-c2/LANE.md	gpu-c2a-score.py
tab:sentinel	sprint1-detection/redesign-eval.md	sentinel_redesign.py
tab:amber	sprint1-amber-mining.md	amber_mine.py
tab:armA	sprint2-armA/LANE.md	armA_score.py

tar xzf checkpoint-surgery-artifact-1d2871160e07.tar.gz
cd checkpoint-surgery-artifact
python3 scripts/gpu-c1a-score.py measurements/sprint1-c1a

What is inside

measurements/<lane>/ — the lane directories verbatim: raw --json diffs, derivation records, per-item eval JSONs, cost TSVs and the lane report.
measurements/PROTOCOL.md — the pre-registration protocol. A prediction is only scoreable if its receipt commit predates the run.
measurements/RESEARCH-LOG.md — the campaign ledger.
scripts/ — the six scorers above.
runbooks/ — the operational runbooks for the lanes this paper reports.

On the falsifications

Six of the twenty-three pre-registered predictions were falsified, and the paper reports them at equal weight with the eleven that held. The lane reports here record each verdict as it fell, including the ones that went against us. Where a lane report and the paper disagree, the lane report wins.