rlcd-lite
Laptop-scale reimplementation of RL for Calibrated Decisions: parallel constrained decoding trained with GRPO and a Brier reward.
Overview
rlcd-lite is a small reproducible reconstruction, from public descriptions, of the training idea behind Jev that TypeSafe calls RLCD. It is not TypeSafe's unpublished algorithm. The engine prefills a context once and scores every decision field in one fused forward pass with a block-diagonal mask, masking logits to single-token option labels (verified per tokenizer so 77- and 151-way intents work), and assembles always-valid JSON. Each decision has a stated-confidence field.
Training uses GRPO (reducing to REINFORCE with a group baseline for one-step decisions) with an RLCR-style reward of correctness plus a Brier term, and ablations for Brier-only, binary-only and decoupled advantages. Evaluation covers softmax, stated and temperature-scaled confidence, Brier decomposition, ECE with bootstrap CIs, risk-coverage and option-order checks. The README reports findings: Brier on stated confidence alone is a trap (accuracy fell from 0.47 to 0.23), RLCR gets accuracy and calibrated stated confidence while binary reward gets only accuracy, calibrated is not the same as informative, and RL collapses the raw softmax.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | What is the severity of this item (A-D) and how confident are you (1-5)? | A · B · C · D |
| Choice | Which intent label fits (77- or 151-way)? | — |
How it works
- Prefill context and score all fields in one masked forward
- Sample (choice, confidence) rollouts per field
- Reward correctness plus a Brier term and update with GRPO
- Evaluate accuracy and multiple calibration signals
Strengths
- Tests the calibration claim on a laptop
- Clear ablations and reported negative findings
- Always-valid JSON by construction
Caveats
- Not TypeSafe's algorithm
- Small-scale experiments
- No licence stated
At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- PyTorch
- Has code
- yes
Use cases and integrations
How curated lists describe it
Simplified RL for Calibrated Decisions: parallel constrained JSON decoding +…
wh000wh000/awesome-jev-live
Frequently asked questions
Is this TypeSafe's RLCD?
No; TypeSafe has not published RLCD, and this is a reconstruction from public descriptions.
What is the main finding?
A correctness-plus-Brier reward gives accuracy and calibrated stated confidence, while binary reward gives only accuracy.
Can I run it without big hardware?
Yes, it is designed to be testable on a laptop.
Related projects
Exact-target audit finding Jev's Choice probabilities are a step function on stated base rates while per-option Noul tracks them closely.
ProjectresearchPythonResearch replicating Jev-style typed decisions with open LLMs via logprobs; Qwen3.5-9B reached 94% with good calibration in ~1 s.
ProjectresearchPythonLocal reproduction of Jev-style inference that reads option-token logits from a small LLM in one pass, with an honest failure log.
ProjectresearchPythonOpen decision models that derive calibrated confidence from recurrent looped-layer dynamics instead of RL, with early exit and a Jev API.
ProjectresearchPython- AnyJev★ 1,255
Nokia's training-free library that turns open LLMs into Jev-style deciders robust to option order, plus self-distilled Tacit models.
ModelresearchPython Jev-style replica: frozen Qwen3-1.7B plus a ~0.8M-parameter trained head returns calibrated Choice, Boolean and Score distributions.
ServiceresearchPython