Level 1 fixes one claim about a reply for every row: the reply is the committed model's own greedy reply to a prompt the verifier is given (RFC-0018 (new tab)). This claim is called the level's statement. A system that proves another statement, such as VerInf, is listed under “a different statement” and has no row.
The model's weights and every activation it computes are together called the witness, and the verifier never holds the witness. What it is given differs by system, and each system's Statement block lists it with the publisher's words (FlashProve, SASH). FlashProve's cp_verify takes the commitment and its files from our commit record, and the prompt and the reply from the request. SASH's boundary-grid --verify takes the model roots and the grid's link counts from our commit record and the prompt from the request. It reads the reply, each segment's shape and the switches of that segment's checks from the proof itself, and that proof's Fixed statement is “produced at setup, trusted by the verifier (new tab)”. FlashProve claims zero knowledge, proved for its protocol only and not stated for its programs, and SASH states no hiding property.
Each system's statement
One FlashProve proof covers one forward pass over the request's 1,024 tokens. Under exact match, a YES “is meant to show that each checked token is the committed integer model's top choice at its position, up to the error bound (new tab)”.
One SASH proof covers one segment of a grid: a transformer layer, or the output head, on one block of 512 token positions. SASH's boundary-grid --verify exits 0 only when, among other conditions, every segment proof verifies, every implied link “ran and chained on its own axis (new tab)” and “every position's id is the prompt's or the previous position's proved argmax (new tab)”.
The committed weights
Neither statement says whose weights the commitment holds.
A yes does not say whose weights they are, that the shipped committer made the commitment, or which real model it is. A provider can commit to any model it likes, a small or degenerate one included, and a scale file within the admitted range can switch off model blocks, up to all of them (worked out by hand).
FlashProve, guide/known-limits.md at 9e2bdd8 (new tab)
FlashProve's README says to get the commitment ID from “a party your user trusts, never from the server being checked or from files sent with the proof (new tab)”. The harness gives each verify call the commitment it recorded at the commit step.
Integer arithmetic
Both systems prove a quantized copy. FlashProve gives its reason: “A proof needs exact arithmetic (new tab)”.
In SASH, “Every matmul consumes an int8 activation (new tab)”, values between components are bf16, and non-linear functions are lookup tables keyed by a bf16 value (new tab).
Served in floating point, a reply can change with batch size, GPU count or GPU version (Yuan and others, 2025 (new tab)). But with the full configuration recorded, unmodified vLLM and Hugging Face Transformers on NVIDIA GPUs rerun bit for bit, except where a kernel calls atomic functions (Cankaya, 2026 (new tab), a workshop paper).
Under exact match, a served reply is provable only if every token is the copy's own choice. FlashProve finds no such reply (new tab) among 96 stored replies from stock vLLM at 16 bits and 96 at FP8, and it scores them with a Python implementation of the copy, not with a proof.
Checked tokens
A proof covers only the positions it checks, and in FlashProve “a reply stopped short verifies as a shorter one (new tab)”. The level's request has all 512 reply tokens checked, and both systems run under exact match. A tolerance would let a dishonest prover pick among the tokens it admits (new tab).
The seams
SASH binds the seams between segments: “The residual stream leaving a segment and the keys and values a block writes are committed as their own trees, and boundary-grid compares each such root with the root the consuming segment committed for its input. (new tab)”
Fourteen of the sixteen edges (new tab) SASH lists can be turned off by an --unsound-skip- switch. Our adapter refuses to pass one outside a control job, and the Run check fails if one of our counted sessions used one, as its recorded commands show. But a proof can also come from a prover we did not run, made with a switch on. The proof's own Fixed statement carries the switches, so we expect the pinned verifier to accept such a proof and report the skip, and our verdict does not yet refuse it. We have not yet run the trial that would show whether SASH's verifier, called as the row calls it, refuses a proof made with one.
The error bound
FlashProve states “about q/2124 + 2-101 after q hash queries (new tab)” for its verifier written in Lean, at exact match and for one proof, among other conditions. It adds that “N proofs handed in at once under one commitment, or N members of a batch, cost N times that error, taken at the whole list's hash queries; for proofs checked one after another, or members checked in separate sessions, no bound is stated. (new tab)”
FlashProve's row proves 24 replies a batch, and the harness checks each alone, so the bound does not cover the run. The row also runs the C++ verifier, not the one written in Lean, and FlashProve states that “Neither figure, 124 bits per query or 128 bits per round, is claimed for the C++ programs. (new tab)”
SASH's security level is “100 conjectured bits (new tab)” per segment proof, and SASH states none for the whole grid.
Claim cards
No maintainer of either system has reviewed its claim card (FlashProve (new tab), SASH (new tab)). Under residual, SASH's card reads “not stated”, and FlashProve's lists six, among them “A proof names no client or time. (new tab)”