Non-autoregressive decision engine. Typed decisions over text in a single parallel forward pass, returned as JSON probabilities. No LLM, no text generation: nothing to parse, nothing to hallucinate.

Architecture

Full design, FLOPs budget and per-modality SLAs: docs/architecture.md.

Quickstart

uv venv && uv pip install -e ".[dev]"
python -m pytest -q

from opendecision import Decider

d = Decider()  # random weights until you train or load a checkpoint
r = d.predict({
    "state": "Hi, we were billed twice for March. Refund the duplicate or we cancel.",
    "questions": {
        "department": {"type": "choice", "instructions": "Which department?",
                       "criteria": {"billing": "invoices, refunds", "technical": "bugs", "other": "else"}},
        "urgency": {"type": "score", "instructions": "How urgent?", "criteria": ["low", "medium", "high"]},
        "churn_risk": {"type": "noul", "instructions": "Does the user threaten to leave?"},
    },
})
print(r["answers"]["department"]["probabilities"])

Response (same shape as Laya POST /v1/systemone):

{"model": "opendecision-0.1",
 "answers": {"department": {"type": "choice", "choice": "billing",
             "probabilities": {"billing": 0.91, "technical": 0.04, "other": 0.05},
             "unknown": 0.0, "confidence": 0.74, "answer_confidence": 0.91},
             "churn_risk": {"type": "noul", "noul": 0.89, "probabilities": {"no": 0.11, "yes": 0.89}}},
 "usage": {"input_tokens": 31, "output_tokens": 0}}

Training pipeline

Stage	Function	Purpose
0 Pretrain	train.mlm_step	masked-token CPT; CLM->MLM 25/75 schedule for text (arXiv 2507.00994)
1 Distill	train.distill_step	offline-cached teacher distributions (Apache teachers only)
2 Fine-tune	train.finetune_step	strictly proper loss (log / Brier / spherical, RPS for ordinal) + optional coherence loss
3 RL	train.rl_step	only with outcome feedback: reward r = c - p_a, leave-one-out baseline (unbiased half-Brier gradient, tested)
4 Calibrate	calibrate.fit_temperature, conformal_threshold	per-type temperature on a disjoint split; conformal as an abstain gate, not "calibrated probabilities"

Details: docs/training.md.

Evaluate

Pinned baseline version, same items and criteria, zero-shot and fine-tuned reported separately, raw and post-temperature rows, per-K, paired cluster bootstrap, frozen test split, contamination flags. docs/evaluation.md.

Docs

architecture · training · evaluation · serving · economics · long context · world knowledge · datasets & licences · research evidence

Credits

Builds on ideas from Laya (Apache-2.0), LAVOIR, eve-rlcd, ModernBERT, SigLIP 2, Perceiver IO. License: Apache-2.0.