Test the journey.
Keep every claim bounded.
The existing public lab, corrected interactions and reusable receipt checks. No new trained model or production approval.
Published test records describe their exact revision and scope. They do not prove every input, browser, private tool, source claim or real-world outcome correct.
Start with a real workflow
Load a retained row, compare, inspect the worst error, export the receipt, then reopen and check it. Or replay the six frozen variants and inspect the original and later synthetic banks separately.
Full quality guide and limitations · Claim-to-evidence register · Reproduced gaps before the fix · Receipt and input controls · Numerical replay checks
Hosted acceptance records are revision-specific and listed in the release readback.
Verify an exported receipt
node release-quality/verify-receipt.mjs /path/to/exported-receipt.json
Inspect the verifier. A matching checksum and declared pins establish neither a signature nor the producer's identity, factual correctness or action permission. Recompute the model separately.
What remains outside this release
Authenticated source context, human approval-to-effect, model capability expansion, real serving-cost proof, distributed BFT, general-world prediction and GSPC admission remain separate. Neither the 34-axis roster nor the 100-entry inventory means those areas have all been implemented or tested.
Existing dependency-ordered next steps · Original methods paper · Public versus private MCP guide
Recompute an exported research receipt
This check uses twelve fixed, hash-checked code/model/data artifacts. It recomputes supported comparison and replay results on this browser CPU. Your pasted receipt or file is not uploaded. A matching result is not a signature, human approval, current source authentication or real-world validation.
No receipt submitted. No inference has run.
No current verification result.
Method and limits · Executed checks · Pinned source and artifacts
The exported bundle contains the exact receipt you supplied and its recomputed result. It is unsigned and can be rechecked in this same form or offline. Export only information you intend to share. Files are read locally; no selected file is uploaded by this checker.
File round-trip method and limits
The altered example is a deliberate negative control: it claims zero error and has a correctly recomputed checksum. Numerical replay should refuse it.
Executed across platforms
Inspect the completed Kaggle CPU replay and Cloudflare static delivery. Same pinned inputs; different runtime and delivery paths. No production promotion.