pi-autonoxis-model

A Pi extension that gets conductor decisions for autonomous coding lanes from a local model. The recommended model is Polaris 3 (polaris-3, HF damianborek/polaris-3), the autonoxis decision model (a LoRA adapter on Bespoke-Nimble-9B, itself built on Qwen3.5-9B). It runs on your own GPU and costs nothing per call. Polaris 2 (polaris-2) still works with this version (same questions). Polaris 1 (polaris-1) still works with older versions of this extension (0.3.x).

The model does not generate text. It picks one label from a fixed set and returns a probability for each label.

Links

• Polaris 3 model (recommended): https://huggingface.co/damianborek/polaris-3
• Polaris 2 model (previous): https://huggingface.co/damianborek/polaris-2
• Polaris 1 model (older): https://huggingface.co/damianborek/polaris-1
• Vega 1 model (earlier, generative): https://huggingface.co/damianborek/vega-1
• Claude Code plugin: https://github.com/damian87x/autonoxis-model

How it works

1. Pi (or the gate CLI) sends a lane packet, the text describing the lane's current state, to the local server.
2. The server asks the model one question for the chosen track and scores every allowed label:
   • decision: STOP | ASK | DISPATCH
   • manager: ACCEPT | VERIFY | REJECT | REOPEN | ESCALATE
3. The extension returns the top label, its confidence and all probabilities.
4. It acts only when confidence >= 0.8 (act: true). Below that, act is "escalate" and a human or a stronger model decides.

Confidence is (n * max_prob - 1) / (n - 1) for n labels.

The request shape is the same as TypeSafe Jev's /v1/systemone, so the server also answers general choice / noul / score questions.

Install

pi install git:github.com/damian87x/pi-autonoxis-model        # for your user
pi install git:github.com/damian87x/pi-autonoxis-model -l     # for this project only (.pi/settings.json)

The extension is a client only. It needs the model server running locally; see server/README.md for requirements, the adapter download and the run command.

Usage

• Tool autonoxis_conductor: {track: "decision" | "manager", packet: string} returns label, confidence, probabilities and act.
• Tool autonoxis_evaluate: {state, questions} in Jev's shape returns choice / noul / score answers with probabilities and confidence.
• Command /autonoxis-model status | test | url <endpoint>: health check, session counters, endpoint override.
• CLI bin/autonoxis-gate.js -c "<criterion>" [--diff | -f file | stdin] [-p 0.7] [--json]: exits 0 on pass, 1 on fail, 2 on error. Works as a pi-subagents gate. The state is cut to 6000 characters; if the server still rejects the prompt as too long, the gate shrinks it and retries up to 3 times and reports truncated: true.

The model was trained on one call shape only, and autonoxis_conductor sends exactly that: one question with id label, state {"packet": <packet>}, and the question text from src/conductor-questions.json. Polaris 3 (like Polaris 2) was trained with these exact questions for both tracks. Use the question file that ships with the extension version matching your model.

Configuration

There is no API key. The server has no authentication and binds to loopback only.

Accuracy

Polaris 3 (recommended): about the same accuracy as Polaris 2 on real packets (~77%), slightly better confidence gate (v9 seed averages 77.1% vs 76.3%, within seed-to-seed variation; Polaris 1: 64.9%); v5–v8 all correct.

From the Polaris 3 model card:

• v9 is 96 real orchestrator packets. Its labels are model labels, not human labels: the majority of three frontier models (Opus 5.5, Grok 4.7, Astra gpt-6-astra) labelling blind. The 2 packets with no majority are left out, so scores are over 94.
• Polaris 3 scores 75/94, Polaris 2 73/94, Polaris 1 61/94. The untrained Jev contract prompt scores about 74.5%.
• The 0.8 gate is only marginally better than Polaris 2: it keeps 91 of 94 v9 packets at 81.3% (Polaris 2: 91 at 80.2%).
• Confidence is not a guarantee: three v9 misses are unsafe (two VERIFY packets answered ACCEPT, one ASK answered DISPATCH), all at confidence >= 0.96, so the 0.8 gate does not catch them. Treat ACCEPT as "verify first".
• Prompts are limited to 2048 tokens, English only, and the labels follow one decision contract.

Tests

npm test    # offline, against a fake server

License

MIT for this code. The model adapter is Apache-2.0; see its model card.