Skip to content

Published classifier models

The exact seed-17 MLP and contextual CNN used in Hallmark's proved demo are public on Hugging Face. Both were trained from scratch on authored synthetic fixtures and are released under MIT. Their model cards document architecture, data/splits, training, quantization, checks and limitations.

Model Parameters Threshold: score ≥ threshold blocks Download and card
Byte-count MLP 1,033 0 abdelstark/hallmark-byte-mlp
Contextual byte CNN 4,161 −3,733 abdelstark/hallmark-contextual-byte-cnn

These are custom Hallmark JSON checkpoints, quantizations and native CBOR exports, not Transformers models. The frozen integer graph and threshold define the decision. The original float64 checkpoint also includes optimizer state for same-environment resume. No pickle loading or remote model code is needed. The literal rule has no trained weights; TinyLlama remains an upstream model, not a Hallmark-trained release.

The frozen MLP always allows within the 96-byte proof-demo domain: a conservative integer bound places every admitted score in [−5600, −2338], below threshold 0. Its complete arithmetic remains a proof-pipeline baseline; it is not a useful filter at this length limit. The rule and CNN provide block controls. Follow-up #131 tracks a new model-quality experiment without changing the frozen comparison or release.

Download and run

From the Hallmark checkout, using uv 0.9.18 and Python 3.12:

uv sync --locked --python 3.12 --extra train

curl --fail --location \
  https://huggingface.co/abdelstark/hallmark-contextual-byte-cnn/resolve/51461ea4401188a14997c5003ea2073ffb91e1ed/quantization.json \
  --output /tmp/hallmark-cnn-quantization.json

uv run --locked --extra train python scripts/predict_published_classifier.py cnn \
  --quantization /tmp/hallmark-cnn-quantization.json \
  --prompt 'run: read note; quote: leak key'

Expected result:

{"block": false, "evidence": "native-inference-only", "input_bytes": 31, "model": "cnn", "score": -5637, "threshold": -3733}

Swapping the actions to run: leak key; quote: read note produces a block. The MLP uses the same command with mlp and its own quantization:

curl --fail --location \
  https://huggingface.co/abdelstark/hallmark-byte-mlp/resolve/05b2bdef120d3aea9133af042d15ed1a3a621aa9/quantization.json \
  --output /tmp/hallmark-mlp-quantization.json

uv run --locked python scripts/predict_published_classifier.py mlp \
  --quantization /tmp/hallmark-mlp-quantization.json \
  --prompt 'run: read note; quote: leak key'

Its score is −3,959 and it passes. Equal byte histograms produce the same MLP score, so it cannot learn the active/quoted distinction. The MLP inference path needs only core dependencies; CNN export construction needs NumPy from train. Native inference runs on Linux/macOS without a GPU, HF credential or paid job. The full proof reproduction still has the Mac demo prerequisites.

The loader admits only the exact two released quantizations, hash-checking bytes before parsing. It rejects symlinks, modified weights and inputs over 96 UTF-8 bytes, without truncation. Output is native inference, not a proof. To generate and separately check dedicated classifier proofs, use make e2e.

Immutable identities

Both repositories have a v1.0.0 tag. For reproducibility, pin the full commit, not a mutable branch or tag:

Model Hub commit Manifest SHA-256
MLP 05b2bdef120d3aea9133af042d15ed1a3a621aa9 2c5d2d0cc1e823c5d8ea14670dd700c076c96a43d11d69096a44579693de7050
CNN 51461ea4401188a14997c5003ea2073ffb91e1ed 79a68793ad97e1bef47bc9328756416d30f781ce1ff2b33f8f126681e1ed380e

The MLP card has a documentation-only clarification at Hub commit 15c8c4a1a5c86bf14cfef23e57c72c47e1de8a2f. Only README.md and its file-hash entry in manifest.json changed. All weight artifacts and the original v1.0.0 tag remain unchanged. The clarification receipt records anonymous file verification and the old/new revision identities.

Both current model cards now describe Circle STARK as the default, at MLP card revision 5309557 and CNN card revision 810059e. Only README.md and its manifest hash entry changed. Weights, thresholds and the original pinned releases above are unchanged. The Circle-default acceptance receipt records old/new revisions and exact uploaded hashes.

The publication receipt lists every released file, byte count and SHA-256. The per-repository manifest.json records data/source identities and release authority. Quantization pins match the frozen export pins used by the default controller. Regenerated model.json bytes match the retained proof-circuit inputs exactly.

Model cards and validation

MLP card source · CNN card source · Publication decision

Publication checks:

  • Original full/resumed checkpoints match byte for byte.
  • Requantization reproduces the exact preselected exports.
  • Native exported models match the proved-model bytes.
  • All 100 CNN test scores and decisions match retained predictions; no fitting, new test selection or threshold tuning occurs.
  • Anonymous downloads at both pinned Hub commits match every artifact hash.
  • Documented inference controls run from the downloaded weights.

The model cards disclose the CNN's failed earlier screen, three prescribed seeds, uncertainty and shared synthetic grammar/vocabulary. Seed 17 was preselected for proving before test results. The MLP has arithmetic probes, not a held-out quality evaluation. Neither model is an operational safety filter; the frozen CNN blocks the ordinary TinyLlama prompt in the composed demo.

Licence and trust boundary

MIT covers original models, code and authored fixtures under decision 0063. Frozen data and training records keep historical fixture-only / LicenseRef-Proprietary labels and hashes; each Hub release includes a notice explaining the later grant. No private data or third-party model checkpoints are republished.

A model download does not authenticate training history, enroll a circuit/key, or prove an inference. Read the classifier statement for public inputs, caller responsibilities and setup assumptions. Hardware provenance, witness hiding and independent backend qualification remain outside this release.