Contribution scope: performance (rewarded via gittensor) · real-HA tool-dispatch fixes (else closed)

[ai-hpc]

What we accept (and what we don't)

To keep review quality high while M1 is active, a PR is accepted only if it lands in one of these two buckets, with reproducible on-device proof. Anything else is closed.

💎 Performance PRs are rewarded. Land a performance-improvement PR that meets the rules below — measurable Jetson win, reproducible before→after proof — and you're eligible for a reward through gittensor, the Bittensor subnet that pays out for merged open-source contributions.

1. Performance improvement

Measurable latency / throughput / memory wins on Jetson Orin Nano 8 GB, with before→after numbers.

• Example: genie-ai-runtime#85 — in-memory KV prefix cache, ~13× faster prefill (16s → ~1s per command); cut the BFCL eval from ~62 min to ~20 min.

2. Tool-dispatch / real-Home-Assistant error fix

Correctness fixes in tool routing or home actuation, reproduced against a real Home Assistant. A runnable sample HA config is provided at deploy/homeassistant/ — bring it up, reproduce the failure, and prove the fix by showing the entity state change confirmed via the HA API.

• Examples: #400 — canonicalize home_control action synonyms (model emits "turn off", runtime required "turn_off"), proven by a live off → on actuation; #380 — stop leaking unparsed tool-call JSON to the user.

---

Required: every such PR includes a Real Behavior Proof — what you ran, on what hardware, what changed. For HA fixes: live-HA before/after, confirmed via the API.

Closed on sight: anything outside these two buckets, prompt growth without a measured accuracy gain, generated/demo-only code, duplicate or churn PRs, or changes without reproducible proof.

This mirrors the README's Accepted contribution scope (PR #401). Questions / better-collaboration ideas → Discord.