Independent reproduction

Reproduce the frozen result.

The repository contains the FIC/5 implementation, exact public evidence, H2 holdout report, frozen candidate identity, and benchmark scripts. GitHub Actions are not required.

Environment

Node 22.x
npm install
npm run verify

Focused FIC/5 checks

npm run test:codec-v5
npm run bench:codec-candidate-fic5

The full verification command also runs the memory-machine tests, security checks, prior codec generations, public corpus gates, FIC/5 sweep, and cold-start accounting.

Evidence to inspect

codec/fic5-color-grid.mjs — frozen FIC/5 implementation.
codec/corpus/FIC5_BALANCED_FREEZE.json — frozen candidate identity.
codec/results/FIC5_PUBLIC_SWEEP.md — public result.
codec/results/H2_HOLDOUT_REPORT.md — consumed H2 result.
CITATION.cff and research/PAPER.md — citation and canonical paper.

What must remain fixed

Do not alter the candidate source, transform set, tile size, baseline matrix, normalization, or metrics and still call the result a reproduction. H2 is consumed and cannot be reused as unseen optimization data.

Independent replication target

A useful replication should report exact-pixel agreement, total-package ratio, incremental crop-payload ratio, cold-object crop ratio, and decode-time ratio against the same declared complete exact-deduplicated tile-pyramid baseline matrix.