ICL-Native R5: bounded semantic opcode kernel
Private custom-Python model staging. Actual 2,774 trained parameters, no base model. This model selects finite opcodes/errors from structured input; it is NOT general source-text generation, an LLM or a learned arithmetic system.
python -I -S inference.py --request examples/demo.json
python -I -S validate.py
Only Python 3.11+ standard library is needed. model/weights.json holds the selected seed17 fitted weights. The model graph and all twelve core runtime files are pinned by the original manifest. examples/demo.json returns Ok(-4) after independent VM verification. Context/memory errors and denied journal examples are included. No external tools are installed. Confidence is uncalibrated.
Read SOURCE_MODEL_CARD.md for the original 18-run training record, 16 intents, 22 classes, 1–32 supplied steps, split design and negative budget-comparison result. Historical benchmark scores are not new replications here. Exact arithmetic, code emission, state, memory and permissions are deterministic. The 30,000-example data and complete checkpoints are preserved in the sibling github/icl-native-r5 research package.
This repository is not Transformers AutoModel/Inference API compatible. Public licensing, rights and author approval are pending; no public HF upload is performed. Source code remains unmodified in this repackaging. No license is granted by this preparation.
Verified packaging revision — 2026-10-20
Preserved model weights reloaded; saved or synthetic example checks passed. No training, held-out evaluation, clinical or hosted verification.
Runtime state: PASS. Rights and release approval are separate. See RELEASE_VERIFICATION.json. No paper or conference submission is included as a new software result.