Skip to content

From inference to a claim

Suppose someone returns:

The capital of France is Paris.

The sentence does not identify which model produced it. Adding a model name asserts an origin; adding a hash identifies bytes. To check the computation, you can rerun it or verify an execution proof against the model and inputs you expect.

Four questions, four boundaries

Question Evidence you need Hallmark today
Did the expected computation yield this result? Exact relation, expected inputs and a checked cryptographic proof Fixed TinyLlama response and dedicated classifier relations
Is this the model I intended? Independently enrolled model/program/parameter identities Caller-pinned expectations; optional signed enrollment
Where did the input or execution come from? An authenticated capture/provenance mechanism Not established by these proofs
Is the policy useful? Task-specific evaluation, suitable data and an adoption decision Synthetic engineering tasks only

Three columns: what is proved by the STARK, what the caller checks, and what is not claimed.

A proof can be valid while the policy is unhelpful. In the composed demo, the contextual CNN blocks the ordinary TinyLlama prompt. Its proof establishes the CNN computation, not a sensible real-world safety rule.

A useful abstraction

Let \(x\) be the expected input, \(\theta\) the enrolled model, and \(y\) the claimed result. A witness \(w\) carries the intermediate values needed by the selected proof relation.

\[ R(\theta, x, y, w) = 1. \]

The verifier checks a proof against public expectations. It does not infer the correct model or input from a prover's file names. In Hallmark, root/key enrollment and request binding are explicit parts of the surrounding protocol.

The LLM and classifier statements define Hallmark's implemented relations in detail.

Inference becomes an exact arithmetic problem

Floating-point library calls leave many choices implicit: rounding, clipping, accumulation, quantization and padding. The proof relation needs a fixed answer to those questions.

Hallmark's LLM case names a disclosed mixed-quantization profile, y-attnout-w12. The classifier exports specify integer graphs and thresholds. The proofs bind those exact arithmetic choices. They do not prove that every input would receive the same answer from the original floating-point network.

Integrity and hiding are separate

A proof of a correct computation does not automatically hide the weights, prompt or witness. Hallmark's current research path focuses on integrity. Its published classifier weights and synthetic examples are public, and the LLM backend has no witness-hiding claim.

An execution proof also leaves hardware provenance and other activity on the host unverified. The MVP scope lists these remaining gaps.

Check your understanding

A valid classifier proof returns block. Did verification fail?

No. The proof can establish that the classifier correctly computed a block verdict. Proof validity and classifier approval are separate outcomes.

The prover supplies a model hash. Is that sufficient enrollment?

No. The caller needs an expected identity from an independently trusted source. Accepting whatever identity the prover supplies lets it choose the claim.

Continue with read a proof statement.