Lucid Computing (new tab) runs this project and builds the site. It sells inference run in hardware enclaves: “All inference runs inside hardware-encrypted Trusted Execution Environments.” It builds none of the listed provers, but two systems listed as another kind of check (Private ML SDK and Tinfoil) are enclave systems too.
We drafted Level 1's rules with an AI model made by FlashProve's publisher. RFC-0018 names the rules this touches and the readings we took that favour FlashProve.
SASH runs commit 4f41667 of pull request #36 to SASH. This project's owner wrote that pull request, upstream has not merged it, and a SASH maintainer has since added commits to it. At upstream's own commit, eb95bfa, SASH refused one of the level's prompts.
Four rules added to RFC-0018 after the level was frozen bear on rows. Under CH-1a, FlashProve's Model check passed on its prover answering “failed”. Under SE-3a, 2 SASH slots went to a later session. Under SE-3b, one SASH session ran on a host whose self-test failed and no longer counts. Under SI-4a, a passed row takes a position only on its publisher's recorded statement that an accept is the committed model's own greedy reply, or on our reading of one. But such a rule is reserved for a new level (LV-1).
We run each system at one pinned public commit, and our adapter passes no setting the system's documentation does not name. FlashProve runs its upstream commit unchanged. SASH runs a commit of a pull request to SASH that this project's owner wrote and upstream has not merged (see Ties).
Every system gets the same request. Its prompts come from a public randomness beacon: drand's first round after the commit that froze the level, by that commit's own time. Our own clock sets that time, and GitHub recorded each freeze 2 s after it. We froze the level three times on 2026-11-11, and each freeze drew new prompts. Every counted run uses the third draw.
When a verifier accepts a changed model, prompt or reply in our adapter, the check reads not decided (not reproduced), and we tell the publisher. The check reads failed only once a record shows the system's own command, run outside our adapter, accepting the same input.
What we do not do
Payment
We take no payment for a listing, a position or a run.
The rules, the adapters, the measuring code and the run records are in one public repository. Proofs, weights and other bulk files stay outside it, named by digest. The proofs are not public yet.
Anyone can rebuild the board from the run records, recompute every figure and validate the repository on a laptop, with no GPU. Verifying a proof again needs its bytes.
Frozen 2026-11-11. The models, the request, the prompts' corpus and beacon, and the constants of the cost, fidelity and order rules are one file. We froze it three times that day, before the first counted run, each time with a fresh beacon round, and it has not changed since. A change of rule is meant to make a new level. But RFC-0018 has changed since the last freeze, and some rules added to it decided rows.
A statement petal moves on a sentence the publisher wrote at the pinned commit, quoted with its link, or on a failed check. A state we read rather than copied says so and is graded Mostly met at most. Fidelity moves on a fidelity run.
A check moves on a stored run record only. A changed input accepted in our adapter fails a check once the system's own command accepts it too.
Corrections
2026-11-16: SASH's slot 2 counts a new session, run on a host whose self-test passed and verified again on a host with no GPU. SASH's row reads Passed again.
2026-11-16: The petals are renamed Integrity, Confidentiality, Fidelity, Formally verified and Trust assumptions, and the state words are plain. No state, step, position or figure moved.
2026-11-15: SASH's Trust assumptions petal is read by the same test as FlashProve's: from Not stated to Mostly met, read as. Both rest on our reading of the publisher's sentences, printed with the petal.
2026-11-15: The mandate said two SASH slots went to later sessions. One did. Slot 2's replacement ran on a host whose self-test failed, and it no longer counts.
2026-11-15: A SASH session that failed the machine self-test no longer counts (RFC-0014), so SASH's row reads Not passed yet until a replacement session runs.
2026-11-15: A state that rests on our reading of a publisher's documents, not on a sentence that says it, is marked “read as” and graded Mostly met at most, for every system. FlashProve's Trust assumptions petal moves from Met to Mostly met.
2026-11-15: SASH's pin is disclosed: it runs a pull request this project's owner wrote. The mandate said our adapter changes no source.
2026-11-15: The beacon line said nobody chose the prompts. The freeze time is ours, and each of the three freezes drew new prompts.
2026-11-15: The reasons a system is not run follow one written rule: DeepProve, JOLT Atlas and zkLLM are listed under no public verifier, zkGPT under a different statement.