pandu-jev
Pandu: tiny local policy networks (2.9K-5K parameters) with calibrated uncertainty that fall back to a teacher when unsure.
Overview
Pandu (pandu-jev) explores minimalist local neural policies of about 2,900 to 5,000 parameters as zero-cost reflexes in closed-loop environments such as GridWorld, 2D racing and arena games including Snake. Policies are trained from an A* expert oracle, run in 17-24 microseconds on CPU or Apple Silicon, and use calibrated uncertainty to hand off to a slower teacher when unsure.
The README reports an ECE of 0.86%, error-detection AUROC of 0.952, and that hybrid fallback matches the teacher's success while cutting average latency by 70.3% and cost by 69.1%. A Canonical Intent Protocol grounds semantic intent from frozen language models into a 16-dimensional vector, separating language understanding from real-time control. It is inspired by Jev but is a control-policy research project rather than a typed-question API.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which action should the agent take in this environment state?(inferred) | — |
| Yes / no | Is the policy confident enough, or should it defer to the teacher?(inferred) | act · defer |
How it works
- Train a tiny policy from expert demonstrations.
- Evaluate actions locally in microseconds.
- Measure uncertainty; defer to the teacher when unsure.
- Ground language intents via the Canonical Intent Protocol.
Strengths
- Microsecond latency and zero API cost.
- Calibrated uncertainty for fallback.
- Separates language understanding from control.
Caveats
- Toy environments; not a general decision API.
- Figures are self-reported.
Getting started
Use the bin/pandu-jev CLI runner from the repository.At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- CLI
- Has code
- yes
Use cases and integrations
How curated lists describe it
Pandu, is mini Jev model, inspired by Jev.
wh000wh000/awesome-jev-live
Frequently asked questions
How small are the policies?
About 2,900 to 5,000 parameters.
How is uncertainty used?
Low-confidence decisions fall back to a teacher, keeping success while cutting cost.
Is it a Jev clone?
No, it is inspired by Jev but focuses on tiny control policies.
Related projects
Instrumented 2048 web lab where every move is a Jev Choice over four directions, with probabilities, latency and cost shown live.
ProjectbetaTypeScriptReal-time bot battles where you write a fighter's strategy in plain English and a System One model pilots it several times a second.
ProjectbetaTypeScriptBot that plays Red Alert 2 web skirmishes, with code handling mechanics and a decision model making stance, threat and timing calls.
ServicebetaPythonMultiplayer 3D science-fantasy version of Houston's River Oaks District with Jev-powered NPC companion, chauffeur and smart drive.
ProjectbetaTypeScriptBrowser 3D planetary rover sandbox where a TypeSafe controller chooses the next action from complete candidates, versus a baseline.
ProjectbetaTypeScriptBilingual turtle-soup (lateral-thinking) puzzle site where Jev acts as host, judging player questions and grading solutions.
AppbetaTypeScript