Skip to content

Choose a workflow

make e2e trains and proves classifiers locally. Use make e2e-llm to prove the TinyLlama response, or choose another workflow below.

I want to… Command Where / resources Success signal
Check software contracts make check-conformance Linux/macOS; core environment; no account 156 matched cases, 32 unavailable full-evidence obligations; no proof
Run a published classifier scripts/predict_published_classifier.py Linux/macOS; core for MLP, train extra for CNN Integer score after checking model hashes; native inference only
Train, prove and verify classifiers make e2e Apple Silicon Mac; 12 GiB free; uv 0.9.18 Nine Circle proofs, 63 public-input and 18 substitution rejections
Prove a complete TinyLlama response make e2e-llm Mac preparation; 32 GB RAM, 16 GiB free; HF CPU jobs 23 new segments, two honest verifications, 14 rejected mutations
Verify retained LLM proofs locally make verify-llm LLM_RUN=… Mac; native Rust toolchain; 8 GiB free; no HF credentials Two honest checks, 36 rejected mutations
Measure SIMD/CUDA proving make benchmark-llm-gpu … Mac controller; HF CPU/H200 jobs; completed caller-owned LLM run Matched full runs, 69 identical artifacts, fresh CPU/Mac checks
Bind LLM and classifier proofs make e2e-composed Both sets of prerequisites; 32 GiB free Same-request verified relations and classifier verdicts

How Hallmark works: the caller pins roots, programs and the request; SASH proves the TinyLlama answer as 23 linked segments and STWO proves three classifier verdicts; the verifier checks every proof and that both name the same request.

Prefer an interactive interface? Run make demo-tui or make demo-web from a checkout. The workbench uses these same drivers and retains their results; opening an interface does not start a proof or paid job.

Get the checkout

git clone https://github.com/AbdelStark/hallmark.git
cd hallmark

Python packages are not published. Trained classifiers are available on Hugging Face.

The smallest software check

uv sync --locked --python 3.12
uv run --locked hallmark validate-data fixtures/smoke.jsonl
make check-conformance

This validates data and software contracts. It does not generate a cryptographic proof or run hardware attestation.

A local proof workflow

With Git, Make, uv 0.9.18, Rust nightly-2025-07-14 and an Apple Silicon Mac:

make e2e

Allow several minutes plus the first Rust build; the sampled worker cap is 30 GiB. No GPU, token, cloud account or pretrained checkpoint is needed. Read the outputs and troubleshooting guide.

A complete LLM response

Set your HF token in the environment or ignored .env, check prepaid credits, then use the LLM workflow. It starts three sequential paid CPU jobs, under a $2.50 startup-inclusive cost ceiling.

The GPU benchmark is separate: it reuses a completed reviewed public witness archive and starts CPU/H200 jobs under a $5 ceiling. It does not train a new model. Read the run plan and balance-input requirements before launching.