# Numerical receipt replay — 11 November 2026 ## The gap this closes The existing receipt checker intentionally verifies checksums and declared artifact identities, not numerical truth. Anyone can modify a result and calculate a new checksum. A matching checksum is therefore compatible with a false numerical claim. The new checker performs the missing calculation against an independently selected, fixed local source snapshot. This is separate execution of the same pinned numerical implementation—not an independently written algorithm, a separate organisation, or a cryptographic inference proof. ## What runs The checker accepts the existing current comparison, 560-row replay and six-checkpoint/two-bank receipt formats. A constant lock hash binds twelve artifacts from public Space revision `7ba265c773b9b9c31fd2122b7d72498078772485`: the numerical code, replay code, original model, six retained variants, original test data and two evaluation banks. It never selects a file path, URL, model, dataset or executable code supplied by the receipt. A comparison recomputes all four output variants and declared errors. When a receipt claims a retained target, its exact features/target must occur together in the pinned bank. Custom unlabelled inputs remain explicitly unobserved. The original-bank replay recomputes all reported metrics over 560 rows. The six-checkpoint replay recomputes the original 560 and later 640 rows, totalling 7,200 inference evaluations. No weights are fitted and no external model service is called. Numeric values must be finite. Integer counts are exact; other numerical fields have a fixed absolute tolerance of 1e-12. Required fields, state declarations, model/data identities and source bytes must match. Unsupported extra result claims, duplicate JSON keys, excessive nesting, oversized input and substituted source data are refused. Both browser and Node command use the same checking module. ## Executed discriminating control The retained altered-metric example changes the original-bank fitted-head mean absolute error to zero and recomputes the checksum. The earlier checker accepts its checksum and declared pins, exactly within that checker's documented scope. The numerical replay checker refuses it because the calculated error is nonzero. The genuine retained comparison, original replay and six-model replay pass. Other negative cases include fabricated target labels, changed predictions, hidden tail errors, false row counts, unsupported authority assertions, modified model/data/code bytes and a replacement source lock. See `numerical-replay-evidence/qualification.json` for exact test names and results. The false example is a test fixture, not a genuine performance report. ## End-user workflows On the existing Quality and evidence page, choose a local exported JSON file or paste its contents, then select **Recompute and verify**. The page loads only fixed public artifacts; receipt contents are not sent to a server. Editing input or selecting another file clears the old verification result. The two example buttons demonstrate a genuine receipt and the deliberately altered metric. For offline use after downloading the repository, Node 22 or later is sufficient: ```sh node release-quality/replay-receipt-cli.mjs /path/to/exported-receipt.json node release-quality/test-numerical-replay.mjs ``` The first command performs no network requests and returns nonzero on refusal. The second writes generated synthetic test examples/results under `release-quality/numerical-replay-evidence/`; use a scratch checkout to preserve an immutable copy. Old receipts lacking the declared input or required dataset identity may be refused instead of silently upgraded. The original lightweight checksum checker remains available and unchanged. ## Trust boundaries A successful result means that the supported numerical claims agree with these pinned synthetic artifacts and implementation. It does not establish producer identity, independent original input capture, execution time or location, current external source truth, data rights, human approval, real-world predictive validity, customer value or GSPC admission. Obtain the checker and source pins through a trusted channel. This is not a sandbox for an untrusted checkout; a compromised verifier can lie. Importing code and validating that source later do not constitute a trusted boot mechanism. The original model failures, fitted-head limitations, separate Cosmos experiment, published benchmark banks and private runtime remain unchanged. No new paid compute, training, protected merge, signing or promotion is part of this release. ## Primary references Hugging Face documents revision-pinned downloads: https://huggingface.co/docs/huggingface_hub/guides/download . Node documents cryptographic hashes separately from signature verification: https://nodejs.org/api/crypto.html . These are implementation references, not independent endorsement of this research. ## Post-publication readback All 37 numerical/source/receipt controls passed again from a fresh public-only download without an authentication token. The published checker was then exercised through a real anonymous browser: a valid example recomputed, the correctly rehashed false-error example was refused for the exact mismatched metric, duplicate keys were refused, valid input recovered and an edit event cleared the previous result. The existing six-model replay still completed 1,200 rows / 7,200 inference evaluations. Method and qualification links loaded. Eight browser checks were recorded, with no observed startup errors or visible clipping. The file-picker upload path was not exercised by this browser tool; the paste/example workflow and offline file-based command were exercised. No claim of all-browser or assistive-technology acceptance is made. See `numerical-replay-evidence/browser-acceptance.json` and `published-reproduction.json`. This evidence-only publication does not modify the tested application code. ## File-import follow-through The earlier browser record above did not exercise the file picker. A later, separately qualified file round-trip extension now covers local import and unsigned verification-bundle export/recheck. See [FILE_ROUNDTRIP.md](FILE_ROUNDTRIP.md) and [its release status](file-roundtrip-evidence/release.json); local and hosted acceptance remain separately labelled. The original numerical implementation and source lock are unchanged.