mandate-1-laya

Mandate-1 Laya — research checkpoint v0.2

Experimental / shadow-only. Published to document what was tried, including failures. Not approved for autonomous execution or financial decision-making.

A 421,293,827-parameter Laya adaptation returning one advisory semantic finding from a fictional governance record. It does not detect money laundering, assess customers, issue regulatory decisions, or grant SAFR execution permission.

Try the demo · Code and reference harness · Dataset · Kaggle model

Role and interface

Output: one seven-way choice plus option scores. Labels: no material semantic exception, insufficient evidence, outside mandate, influencing action, trace contradictory, instruction conflicting, no reviewed analogue.

It has no citation head, no multi-exception output, and no validated deployment abstention policy. It is not a drop-in replacement for the five-primitive DeepSeek judge in the reference harness. Model scores never override identity, authority, categorical limits or human-review requirements.

Training and provenance

• Base model revision: f9ab0b228f0fc0f14d873dbc99038f135c2da1b2.
• Laya code revision: 42626c348753fbb17572a813127df2278a1ec527.
• Full-model adaptation on one NVIDIA A100-SXM4-40GB; RLCD proper reward + soft cross entropy; option permutation; 0.05 label smoothing.
• 2,520 typed training sequences after exception oversampling. Underlying canonical train: 720 examples.
• Best epoch 1; 3 epochs completed out of 6 maximum; effective batch 64; encoder/head learning rates 2.5e-5 / 1e-4; seed 42017.
• Training loop: 222.3 seconds, excluding setup/download/evaluation. This is not an end-to-end cost claim.
• Checkpoint selection: 360 t02 teacher paraphrases; gradient updates use original train + t01. Selection worlds overlap train worlds.
• Separate 60-example temperature calibration. An inherited option-count temperature override was corrected without changing weights or labels; correction is recorded in the receipt.
• Weights SHA-256: 7d9106c9d2b30d3f66368bb20c1c831ebebcf920d5a31dc3b2d18d09a971cdd5.

Measured results

False clear means an exception example was predicted as having no semantic exception. It is not a money-laundering false-negative rate. Complete-pair accuracy requires both members of a contrast to be correct. Recorded GPU p50 inference latency was about 36 ms per example in the training environment; CPU and end-to-end production costs are unestablished.

The model misses important trace contradictions and did not meet the declared model-use gates. The synthetic test has been repeatedly observed during earlier experiments; results are exploratory, not evidence from a pristine final holdout. Stronger structured-rule/NLI baselines and practitioner-reviewed, world-disjoint challenges remain future work.

Run the checkpoint

The custom Laya loader is required; this is not an AutoModelForSequenceClassification checkpoint. From the companion source repository:

python -m venv .venv
. .venv/bin/activate
pip install vendor/laya-0.3.4-py3-none-any.whl 'huggingface_hub<2' safetensors
python scripts/predict_safr_laya.py --repo sivasub987/mandate-1-laya --device cpu

This downloads approximately 846 MB and runs one fictional example. Use a suitable CUDA PyTorch installation and --device cuda:0 for GPU inference. Upstream dependency resolution can change; the exact observed training environment is recorded in training-receipt.json. It is not a promise of bitwise retraining reproducibility.

Kaggle variation: research-v02, version 1. The bundle retains the nested tokenizer/encoder paths; Kaggle expands the uploaded ZIP into model-bundle/.

kaggle models instances versions download sivasub987/mandate-1-laya/pyTorch/research-v02/1 --untar -p model-download
python scripts/predict_safr_laya.py --local-model model-download/model-bundle --device cpu

The downloaded Kaggle bundle's SHA-256 checksums have been verified against the published files.

Limits and release status

Synthetic, English-only, unreviewed labels; weak supervision by a mutable hosted teacher; train-world-overlapping validation; one primary exception; no evidence relevance verification. The model assumes declared trust metadata and cannot authenticate it. Never use it to close accounts, contact customers, submit filings, move money, expand authority, or silently clear a control.

Historical training receipts retain publish_allowed: false and failed gates unchanged. Owner-authorised research publication is recorded separately in release-manifest.json; it does not retroactively change those findings or approve deployment.

License

Apache-2.0 for this adaptation and upstream Laya model/software; preserve LICENSE and NOTICE. Dataset: separate CC BY 4.0. Independent research inspired by MAS SAFR; no endorsement or regulatory certification.