Skip to content

0064: Qualify a real LLM response proof

2026-10-23. Research protocol under the owner's standing implementation and compute approval. Relates to #19. No backend adoption or change to the existing classifier proof statements.

Target

The owner requested a short generated response from a real pretrained LLM, accepted a disclosed quantized model, and allowed minutes to a few hours of proving with fast verification. Witness privacy is optional. The acceptance gate remains a complete response proof, including the prompt computation, all layers, output selection and token continuity; a synthetic layer or a one-token milestone does not complete it.

Assess SASH at eb95bfa887299d1637550228d42086c27cbb2b3b first. Its root and vendored Stwo licences are Apache-2.0. The source review records the exact relation and alternatives. TinyLlama-1.1B-Chat-v1.0 at fe8a4ea1ffedaf415f4da2f062534de366a451e6 is the first model candidate: public, ungated, Apache-2.0, and an ordinary Llama architecture. Compatibility and quantized response usefulness are unmeasured at protocol freeze.

Ordered gates

  1. Build the unmodified pinned source with its pinned Rust toolchain on the Mac. Reproduce the upstream synthetic digest before changing any build code. Also require the aligned synthetic shape, whose layer bindings are complete. The smaller README example explicitly leaves two edges unbound.
  2. Install the upstream locked witness environment separately from Hallmark. Run its model-free primitive and extraction checks. Fetch only the exact reviewed model/config/tokenizer files; retain their hashes and never execute remote model code. Freeze calibration and demo prompts before inference.
  3. Generate a short answer using the declared quantized arithmetic. Compare against the original model, retaining every mismatch and failure. This is an engineering screen, not a language-model benchmark or safety evaluation.
  4. Extract and prove the complete prompt and 8–16 generated tokens. Enroll model roots, calibration, tokenizer, arithmetic and expected coverage outside the prover's output. Verify in a fresh process without weights or witness traces.
  5. Reject changed prompt, answer, model/calibration, omitted/reordered segments, KV substitutions, argmax ties that violate the decoding rule, and proof-byte changes. Record preparation, witness, proof and verification costs separately.

Keep this work sequential under #19. Revisit dependent classifier and verifier issues after the whole-graph gate. Existing hardware/release dependencies remain; they do not block this integrity-only experiment.

First build envelope

Mac arm64, four build/prover threads, one compute worker, 1,800 seconds per job, 24 GiB sampled process-group RSS stop, 16 GiB additional local disk, zero cloud spend. Preserve the compiler output, command, lockfile/source hashes, measured costs and failed directory. A timeout or compiler failure is a failed build, not a reason to weaken arithmetic or disable a binding.

Upstream documents macOS type checking but no linking. If the actual build confirms that limitation, use the already approved HF/Modal/Vast compute under a separately frozen job manifest with a fresh balance/price check, finite timeout, no automatic retries and an initial total campaign ceiling of $5. The source currently needs CUDA 12.9 libraries even for the Linux CPU backend. No new top-up, scheduler, private upload, collaborator or hardware purchase is part of this protocol.

The prior classifier results and evaluation contracts remain unchanged. This experiment cannot establish historical inference provenance, zero knowledge, hardware attestation, production soundness or classifier quality.