GitHub (new tab)

Learn · For policy readersGlossary

Deployments and treaties

Proving a reply costs a large multiple of serving it, but checking the proof needs no weights and no trusted chip. Under a treaty, an operator could prove replies sampled from a trusted record.

3 of 3

A cloud customer and its operator can trust the GPU's maker, and a company and its auditor can trust a locked server rack. But two states that expect each other to cheat may trust no hardware they share.

Rerun, enclave or proof

WayNeedsTrustsCostToday
Rerun the modelThe weights and GPUsThe machine running it againOne more run per reply checkedAnthropic's prototype (new tab), which “recomputes inference on a trusted server”
Trust an enclaveThe chip maker's key, the expected measurementThe chip maker, and the server's physical securityFirst-token latency 9 to 16 % longer, throughput 2.88 % lower at 64 concurrent requests and 15.38 % lower at one (Cai and others, 2025 (new tab): Llama 3 8B, one H100)Intel, AMD and Nvidia server chips
Check a proofThe commitment and the verifierCryptographic assumptions, the verifier's code, whoever supplies the commitment, sometimes a third partyA large multiple of serving to prove, CPU time to verifyResearch prototypes, such as SASH and Anthropic's FlashProve

The authors of TEE.fail (new tab) forged attestations of Intel's CPU enclaves (SGX, TDX) with a memory-bus interposer costing “under $1000” to build. Intel and AMD leave interposer attacks outside their threat models.

The same authors write that “the fact that Nvidia's attestation reports are not bound to the identity of a specific CVM means we can “borrow” Nvidia attestations from devices that aren't ours (new tab)” (a CVM is a confidential virtual machine). So a GPU attestation does not show which machine ran the workload.

TEEs and activation checks are cheap, and for many settings they are the right tool; TEEs are a sensible starting point for a first agreement. What they cost is trust: in a hardware vendor, in a third party, or in a tolerance. A proof moves that trust into cryptographic assumptions and into code that both sides can read.

SASH, Towards Proving Frontier-Scale LLM Inference, September 2026 (new tab)

What a relying party supplies

The verifier checks a proof against inputs it takes as given. For each system on the ladder, we print the claim card's words on each input, linked to the publisher's line at the row's commit:

From the board, release v1.0, runs to 2026-11-16

FlashProveSASH
The verifier, pinnedcp_verify at 9e2bdd8boundary-grid --verify at 4f41667, pull request #36 (new tab)
The commitment, from a register

the commitment and the files commit wrote: policy.txt, KL.bin, layout.txt, key.bin, 0.bin

The commitment ID comes from "a party your user trusts, never from the server being checked or from files sent with the proof". source (new tab)

"The commitment ID is a hash, in 64 hex digits, of the configuration, the Merkle root and shape of the first weight table, a masked encoding of the weights, and the out-of-domain (OOD) seeds and answers. [...] the ID covers them by design, but the soundness theorem on `verify` is stated at given public files and does not itself say that one ID cannot verify under two sets of them. The ID does not cover the leaf files or the key file's proofs, which `cp_verify` checks on reading." source (new tab)

the commitment and the files commit wrote: model-roots.json, grid.json

The model root is not derived by the verifier: it is "the publisher's commitment to weights the verifier does not hold", compared with the published model manifest (boundary-grid --expect-model-roots). source (new tab)

The verifier derives each segment's shape root from the statement's own parameters and the per-layer calibration the model manifest publishes beside the model commitment; given no calibration, the shape root is asserted by the prover and checked by nothing, and boundary-grid --verify refuses to run without one. source (new tab)

The prompt, and the reply where it is given

the prompt, 512 token ids, and the 512-token reply from our record

"Only the named output positions are covered, and a reply stopped short verifies as a shorter one. [...] A proof names no client or time." source (new tab)

the prompt, 512 token ids

The verifier is given the prompt's ids (--prompt), against which every prompt position's id is checked; boundary-grid --verify refuses to run without them. source (new tab)

The decoding rule

greedy, at every reply position

"the policy under which every checked token must be the integer model's top choice. It is `cp_verify`'s default." The published commitment's acceptance policy is exact match. source (new tab)

greedy, at every reply position

the argmax binds the token a head segment emits to the lowest-indexed maximum of the logits it computed, "which is the rule the model itself decodes by" source (new tab)

What no call checks

"The client must trust what no call checks: the commitment ID, the expected policy and its numbers, the tokenizer and chat template, and `cp_verify` itself." source (new tab)

not stated

In an agreement

Take a hypothetical bilateral AI agreement between the United States (US) and China, under which declared data centres in both countries serve only models whose commitments are in a shared register. An inspectorate that both countries agree on holds the register and checks an operator in either country, and the operator keeps its weights. This is an illustrative scenario, not a description of an existing deal.

With accepted proofs and a trusted traffic record, the inspectorate could sign one finding: each sampled reply is what the registered quantized copy gives for its recorded prompt.

Run on this site

This reply is what the registered model gives

The inspectorate verifies a reply's proof against the registered commitment. But that does not show that the reply came from live traffic, or that the weights stayed hidden. We run this step on open-weights models, with 512-token prompts and replies.

Shown in research

The model that was evaluated is the model that serves

When evaluation replies and served replies are proved against one commitment, both are what the same weights give. South and others (new tab) proved evaluations of small models this way in 2024. But no proof says whether the evaluation asked the right questions.

Proposed, parts built

A sample stands for the traffic

A certifier at the boundary of the inspected data centre, in either country, records a fingerprint of every prompt and reply, and it must be trusted. SASH's publisher (new tab) writes: “Nothing yet ties a proof to what the certifier recorded.”

Later, the inspectorate draws replies for the operator to prove. SASH's publisher adds that “the draw must be unpredictable to the provider, and the provider must commit to its answers before the draw.” By our arithmetic, when both hold, 300 draws catch an unregistered model that serves 1 reply in 100 about 95 % of the time, but one that serves 1 in 10,000 about 3 % of the time.

Any reply may be drawn, so the quantized copy must generate every reply. Where a system's page prints a generation part, that part is paid on every reply, drawn or not. Of the “commit and proving” part, only proving shrinks with the sample.

Outside a proof's reach

The data centre did nothing else

The AI 2040 verification plan (new tab) would keep data centres to inference, and currently recommends aiming for “a network-tap and recomputation based approach”. Amodo Design (new tab) grades the plan's measures: 4 items with an active effort, 6 not started, 7 “Not on track”.

Outside a proof's reach

Nothing was left undeclared

To find undeclared data centres, a RAND working paper (Baker and others, 2025 (new tab)) names chip tracking, satellite images, whistleblowers and national intelligence.