Hallmark documentation¶
Read the documentation site · Learn verifiable inference · Choose a workflow · Inspect the evidence
Start with the architecture, commands and proof flows and the LLM integrity MVP and acceptance record, then run the TinyLlama workflow. Read the exact proof statement before relying on a result. The composed demo adds classifier proofs over the same opened request; results preserve measurements and limitations.
Use and reproduce¶
| Document | Purpose |
|---|---|
| Performance | Every measurement with its run, machine and timer boundary; how the timers nest; what was never measured |
| Architecture and proof flows | Make target matrix, classifier/LLM/composition diagrams, exact proof boundaries, SASH source/patch ownership and backend trade-offs |
| MVP scope and acceptance | Completed fixed-case integrity scope, evidence matrix and deferred v1 roadmap |
| Composed LLM/classifier demo | make e2e-composed; fresh verification and classifier proofs over one exact request |
| What is proved | Classifier claim, public inputs, caller checks and exclusions |
| TinyLlama response proof | make e2e-llm, local verification, signed parameter enrollment and exact inference relation |
| GPU prover measurements | H200 matched pair: 53.94 s CUDA proving, 10.15× over SIMD; generated constraints, zero CPU fallback, cost/timing boundaries and reproduction |
| FlashProve proofs | Pinned upstream FlashProve: make verify-flashprove on the Mac, H200 proving of the synthetic small kit (4.1 MB proof, 48 tokens) |
| SASH and FlashProve compared | What each prover proves, and proof size, verify time, prover cost on Hallmark's runs; make compare-provers |
| Slurm + Apptainer | CUDA 12.8.1 and 12.9.1 container profiles, sbatch templates and launcher for generic GPU clusters (decision 0092, proposed) |
| End-to-end demo | Train, export, prove and separately verify three classifiers on one Mac |
| Published classifier models | Hugging Face cards/weights, immutable revisions, hash-checked inference and publication receipt |
| Results and limitations | Model/data notes, measured results, failed experiments and evidence links |
| Operations | Focused experiment recipes, enrollment, retained artifacts and recovery |
| Packaging | Pinned build environment, reproducible archives and installed-package checks |
| Contributing | Development setup, test selection and change review |
| Security | Trust assumptions, qualification gaps and private reporting |
Understand the design¶
The canonical specification is the entry point to the future private v1 system, beyond the owner-selected integrity MVP. A specified interface is not necessarily an implemented one. Check the supported API table and current state before using a command or relying on a capability.
| Document | Purpose |
|---|---|
| Baseline contract | Original software arithmetic and commitment contract; dated scope |
| ML plan | Classifier candidates, training and export strategy |
| Evaluation | Data lineage, split isolation, calibration, metrics and claim boundaries |
| Implementation tracker | Work-item dependencies and links to live issues |
| Reference proofs and timings | Download actual LLM/classifier proofs, verify them and audit every stage, machine and job |
| Circle STARK classifiers | Rule/MLP/CNN proofs, circuit/caller boundary and STWO source ownership |
| Specification checks | Local links, requirement coverage and explicit live tracker audit |
| Dependency notices | Reviewed software versions, licences and redistribution limits |
For the current demo, the shortest design trail is:
- 0058: specialized trained classifier path
- 0059: contextual byte CNN
- 0060: revised synthetic screen
- 0085: Circle STARK default
- 0090: Circle STARK is the only classifier backend
Inspect the research record¶
- Real LLM proof plan: pretrained response target, SASH source review, measured preparation and remaining proof gates.
- STATE: completed MVP and future qualification gates. Older entries below the opening summary are historical.
- Ledger: chronological commands, failures and results.
- Program: hypotheses, controls and stopping rules.
- Sources: primary sources and inherited research limits.
- Source packet: dated design material, preserved separately from later normative decisions.
Raw run directories, proofs and unreviewed checkpoints stay local. The explicitly approved model releases publish an enumerated set of original synthetic checkpoints/exports on Hugging Face. Public evidence files contain numeric summaries and hashes. A recorded hash is an identity check; it does not make an untrusted source an authenticated enrollment authority.