Every headline on this site is supposed to survive being checked. These are the receipts that let you check it.
Each one is a static page: plain HTML, no JavaScript, no login, nothing fetched at page load. Paste any of these links into an AI assistant and it will read the whole thing, because there is nothing to run. The bodies are the working receipts as they were filed, with their file fingerprints and their stated limits. The one change made for the website: internal file locations, internal job and message identifiers and staff notes were removed or replaced with neutral labels. Every number, table, caveat and sha256 fingerprint is unchanged.
Detailed run records are available to reviewers on request.
Read the limits with the results
Where you see 300 of 300, the full claim is this: those stored cards were built for those exact 300 real MMLU-Pro questions, and 16 of the 300 were repaired with the answer key. It is exact recall of a stored method, replayed by code — including on a second machine that never received the key — and it is not blind solving of new questions. The receipts say so themselves, in their own words, and they say it before they say the number.
The next 700 questions are the test of the harder claim. Our prediction of that result is
already sealed and hash-stamped: sealed prediction sha256
4551cb9a377385b15c867f419cb20a0a88691da301cb580a7476458321694c1d, hashed
2026-09-16T20:39:38Z, published at
/sealed-prediction-2026-09-16/, OpenTimestamps-stamped
and archived at web.archive.org the same night. The predicted numbers stay sealed until
the runs finish.
The four receipts
The first 300 MMLU-Pro questions, 300 of 300 by code, and the same 300 of 300 on a second machine
Stored Hollerith Cards, run by code with no AI model, picked the keyed answer on 300 of 300 real MMLU-Pro questions, and again on a second machine that never received the answer key: same answers, same sha256, 0 model calls, 0 network attempts. The cards were built for those exact 300 questions and 16 were repaired with the key, so this is exact recall, not blind solving of new questions.
Original filed receipt: 24194 bytes, sha256 df7c936995b6ada0…, available to reviewers on request. Reproduced here with internal locations and identifiers removed.
Six flagship AIs on the same first 300 MMLU-Pro questions (paid run)
What six paid frontier models scored on the identical 300 questions, what each vendor billed, and every file fingerprinted with sha256.
Original filed receipt: 16549 bytes, sha256 2f5d35c605563dc2…, available to reviewers on request. Reproduced here with internal locations and identifiers removed.
Free open models on the first 300, and the sealed prediction for the next 700
Free, locally runnable models on the same 300 questions, plus the hash-stamped sealed prediction for the next 700 MMLU-Pro questions: registered in advance so anyone can check we called it before we ran it.
Original filed receipt: 8160 bytes, sha256 a52c837ee2b73574…, available to reviewers on request. Reproduced here with internal locations and identifiers removed.
Hex Connectors reached a 5.32-billion-token store at 50 tokens per answer
The measurement behind the headline: 5.32 billion tokens in reach from 50-token Hex Connectors, the same 50 tokens whether the store holds a thousand files or 660,000. 0 model calls, $0, rebuildable on demand.
Original filed receipt: 10131 bytes, sha256 d1eaf7fbbbec392d…, available to reviewers on request. Reproduced here with internal locations and identifiers removed.
A note on the evidence tables inside the receipts. Each row names a run record by a short label, with its size in bytes and its sha256 fingerprint. The records themselves are not published here; they are available to reviewers on request, and the sha256 beside each one is what lets you confirm that the file you are given is the file that was measured.
Patent pending. Where there is a fee, it comes out of what the Substrate saves them.