jeba-en (System One decision engine)

Status: pre-release. This card describes the intended release. Weights and measured metrics are pending the RTX 3060 training run (uv sync --extra train); numbers below are placeholders to be replaced by benchmarks/report.md.

Model details

• Developed by: The jeba Authors.
• Model type: non-autoregressive encoder with three task distributions (noul, choice, score), answering typed questions about a state in one forward pass.
• Trunk: ModernBERT-large (English) and mmBERT-base (100+ languages); see ADR-0007.
• Adapters: munod/jeba-en, munod/jeba-multi (LoRA; load base + adapter).
• Licence: Apache-2.0.
• Repository: https://github.com/munod/jeba

Uses

jeba answers atomic choice / score / noul questions about a state and returns typed values with probabilities and confidence. It speaks the TypeSafe Jev /v1/systemone wire protocol as a drop-in and runs locally/offline with no API key. Compose several atomic answers in code rather than asking one broad question.

Out of scope: free-form text generation, multi-step reasoning, and any decision requiring extended deliberation — decompose those into atomic questions and combine results in code.

Bias, risks, and limitations

• Probabilities are only meaningful after calibration; the shipped temperature must be applied (see docs/training.md).
• Synthetic training data can inherit generator biases; public probes are evaluation-only.
• Confidence is a property of the distribution, not a guarantee of correctness.

Training

Deterministic synthetic JSONL (training/generate_data.py) supervised with an RLCD proper-scoring objective (training/finetune_rlcd.py), then temperature-calibrated on a held-out split (training/fit_calibration.py). Configs and seed live under training/configs/.

Evaluation

Reported by training/evaluate.py and rendered by benchmarks/report.py (accuracy, ECE, p50/p95 latency per primitive and language).

Full-scale run (single RTX 3060 12GB): 9,000 English / 18,000 multilingual train / 1,500 eval deterministic synthetic records (fully localized per language, a learnable other team with rich descriptions, per-record RNG, one-in-six distractor clauses), LoRA (r=16) plus a dedicated low-rank choice head (r=32, near-identity init), 4 epochs, batch 16, bf16 + gradient checkpointing.

Checkpoint	Accuracy	ECE (calibrated)	p50 (ms)
English (ModernBERT-large + LoRA + choice head)	0.781	0.077	22.9
Multilingual (mmBERT-base + LoRA + choice head)	0.711	0.073	12.4

Per primitive (English): choice 0.708, noul 0.744, score 0.892; (multilingual): choice 0.734, noul 0.716, score 0.682. The localized, per-record-RNG data (B-1) lifted multilingual choice from 0.40 to 0.73 and English overall from 0.72 to 0.78. Per-language calibration for a few multilingual languages (es, nl, de) and multilingual score remain the next targets. The CUDA-graph fast path (JEBA_FAST=1) gives a 2.7× p50 speedup with 0 top-label flips. Full tables and environment are in benchmarks/report.md.

Citation

@misc{jeba2026,
  title        = {jeba: a local-first System One decision engine},
  author       = {The jeba Authors},
  year         = {2026},
  howpublished = {\url{https://github.com/munod/jeba}}
}