Skip to content

Learn verifiable inference

These four chapters explain the proofs for Hallmark's short TinyLlama response and classifier decisions: what they establish, how to run them, and which questions they leave unanswered. No cryptography background is assumed.

Step Question What you will be able to do
1. From inference to a claim What does a proof establish? Separate computation, model identity, input provenance and policy quality
2. Read a proof statement What exactly must be true? Identify public expectations, witnesses and checks in each Hallmark flow
3. Circle STARK proof paths Why SASH for LLMs and direct STWO for classifiers? Compare their roles without confusing unlike workloads
4. Verify one request How do the pieces meet? Follow the three classifier cases and the composed LLM request

Vocabulary

Statement
The precise claim being checked, including the expected model, input and output.
Witness
Values used to establish that claim, such as intermediate activations or an execution trace.
Proof
A cryptographic object checked by a verifier against public expectations. A proof file is not its own trust anchor.
Enrollment
The caller's decision about which model, program, parameters and keys it expects.
Integrity
Evidence that the stated computation holds under its assumptions.
Zero knowledge
A separate privacy property: the proof reveals no additional witness information beyond the statement. Hallmark's current LLM path makes no hiding claim.
Attestation
Evidence about a measured hardware/software execution environment. Hallmark's current integrity MVP does not claim hardware attestation.

Prefer running something first?

Choose a workflow. The software smoke test needs no GPU or account. The local classifier workflow produces real proofs; the smoke test does not.

Reading a benchmark

For each timing, check which computation and inputs were used, which verifier ran, and which stages the measurement includes. Hallmark's 53.94 s GPU figure covers summed proving sessions; preparation and verification take additional time. The performance guide shows how every timer nests.

Continue with from inference to a claim.