The sealed exam.

Sealed exam

A benchmark you can prove is fresh

BitGraph A binds a request naming a public question bank. Twenty questions are derived from A’s nonce, unseen until A is bound, so no training set frozen before A can contain them. The model answers, and the answer sheet is bound as BitGraph B, after A in the same sequence. Anyone checks the whole sitting offline, with nothing but Node.

Download the packageSource on GitHub

The results

These exact questions could not have existed before BitGraph #14 was bound

Claude Haiku 4.5 scored 14/20 on 24 October 2026. BitGraph #14 and BitGraph #18 are counts in one sequence. The AWS Nitro attestation’s clock, a hardware layer signed by AWS, puts them at 20:14:24 UTC and 20:14:43 UTC, so the answers were bound by B’s attested time. The bank is small and easy on purpose: this is about freshness, not difficulty.

BitGraph A #14, the requestBitGraph B #18, the answers
BitGraph A
the request
20 questions
From A’s nonce
The model
Answers
BitGraph B
the answers
The questions could not exist before A.
The answers are bound at B.

The claim

What it proves, and what it does not

  • Provesevery question is derived from BitGraph A’s nonce, which the enclave drew and bound in one atomic action and nobody could see before A existed, so the exact questions did not exist before A, and no training set frozen before A was bound contains them; A’s request named the bank, so the bank was fixed before that nonce; BitGraph B bound the answer sheet, which names the paper and A, at a later count in the same epoch; both proofs carry an AWS Nitro attestation of a published BitGraph enclave image, and B’s AWS-signed attestation time is when the answers were bound.
  • Does not provethat the template families are unfamiliar to the model; that the model worked alone, without tools, or quickly; that the answers came from the model the sheet names.

Check it

Three commands, no network

Unzip the package (11.2 MB; Node 22.13 or later). Everything it needs is inside it, and the verifier runs in a process where fetch and sockets throw, so turning the wifi off first is a fair test.

node verifier/demo.mjs
node verifier/exam.mjs verify demo/claude-haiku-4-5.exam
node verifier/exam.mjs selftest

The answers are ACCEPT, REJECT (naming what disagrees) or NO-EVIDENCE (something is not in hand). The package carries nine fixtures that should fail, among them a proof re-signed with a laptop key that a signature check alone passes. The zip’s SHA-256:

e99d692e33c3dc2171f07b365be89cc57fa844de67e29115f654d3be964ac57a

To run a sitting of your own, with your own API key:

npx @mikeargento/exam-cli run --provider anthropic --model <id> --name mine