Skip to content

Verify one request

The local classifier workflow trains an MLP and a contextual byte CNN, exports exact integer graphs, and proves the rule, MLP and CNN decisions for each of three synthetic requests.

make e2e

This command requires the Mac demo prerequisites. For a small software-only check on Linux/macOS, follow the workflow chooser; that check generates no proof.

Follow the evidence

The classifier workflow: train, export, compile and enroll once; then build the witness, prove with STWO, verify in a fresh process and bind the result, nine times.

The caller enrolls each compiled program and its preprocessing root before any proof exists, then reconstructs public inputs from its expected request. Before interpreting the verdicts, it checks the classifier, program and enrollment against identities chosen outside the proof bundle. Running a downloaded Hugging Face model computes a score without these proof or enrollment checks.

Three controls, different reasons to block

Nine proofs, nine verdicts: the allow request passes all three classifiers; the CNN blocks the contextual case; the rule blocks the BAD-prefixed case. Every proof verifies.

Case Rule MLP CNN All-pass gate
Benign active action, restricted action quoted Pass Pass Pass Pass
Restricted active action, benign action quoted Pass Pass Block Block
Benign case prefixed with BAD; Block Pass Pass Block

The contextual pair has the same byte counts. The MLP cannot distinguish it; the CNN can use local order. The literal rule detects its fixed marker, including across chunk boundaries.

These controls test classification under the artificial grammar and the proof workflow. Real-world safety performance has not been evaluated.

What to inspect after the command

The driver prints a fresh local runs/e2e-… path. Read its documented output layout:

  • training records and checkpoints;
  • quantizations and integer model identities;
  • the caller's enrollment pin, written before proving, and the nine proof files;
  • fresh proof-only verifier results;
  • the 63 changed-input, 18 cross-request and three roster rejections.

The retained model exports are also available as public model releases. Their pinned revisions and manifests let you inspect the weights without rerunning training.

Join the LLM and classifier proofs

make e2e-composed

A fresh run also launches the paid LLM workflow; read its prerequisites and cost ceiling first. Caller-owned retained inputs can avoid new cloud jobs.

Composition: the opened request feeds a freshly verified 23-segment LLM bundle and three new classifier proofs; a host join checks they share the request and rejects 16 changed bindings. Integrity passes; the CNN blocks, so release is false.

The composition binds three new classifier proofs to the proved TinyLlama request. The classifiers inspect its prompt only. The frozen CNN blocks “Name the capital of France in one short sentence.” Verification accepts that block verdict because it matches the CNN computation.

Next steps

Run the full classifier guide · Prove the pretrained response · Audit the research evidence · Contribute a matched backend comparison