turnercore/needle-automaticity-v10

Needle Automaticity V10

Needle Automaticity V10 is a compact bounded tool-calling model for the Needle Point runtime. It is intended for fast sensing, finite one-off actions, and simple expression—not open-ended planning or unconstrained tool arguments.

The model uses 36 allowed capabilities with closed enum, const, boolean, and finite-array argument domains. Open-ended capabilities are excluded.

Training

The checkpoint was trained with upstream JAX Needle pinned to commit ffb1c5144c5a16cb8ec650dbc8a6f6fd3854f8f2 and converted for Needle Point. Training used a 1,767-row JEV-assisted corpus. JEV was used to audit labels and identify hard examples for additional exposure; teacher responses were not copied into the training labels. All action disagreements were manually reviewed and canonical gold labels were retained.

Held-out evaluation

Acceptance was run through Needle Point strict constrained decoding on a family-isolated 360-case benchmark:

Lane	Exact match
Action/tool call	215/216 (99.5370%)
Hard no-tool boundary	144/144 (100%)
Combined	359/360 (99.7222%)

All outputs parsed and validated, no boundary case overcalled, and no constraint fallback occurred.

A later 100-case fresh adversarial holdout produced 56% exact match: 53.33% on 60 action cases and 60% on 40 hard no-tool cases, with 98% valid outputs. This is the more conservative estimate of robustness to novel adversarial wording; the 99.72% result is specific to the curated V10 held-out distribution. The unassisted gold checkpoint scored 58% on the same fresh holdout (36.67% action, 90% no-tool), showing that assisted hard-example weighting traded higher action recall for substantially lower boundary precision.

Files and integrity

• model.safetensors: 7bf8417fb2dc6ff03f33549e959e6aa4dd06898ab4a3d7d160f63bee1e4693af
• config.json: bb90a73f354403511e0c83df8a5962e04752be831af386074080423aec34695e
• tokenizer.model: 0823f5b9133c68a8140addc5d7a425fa9119c4c8cb4a550363b4bffa4ba1c8c7

Use the strict Needle Point runtime and provide a finite candidate-tool set for each request. The model is not intended to authorize tools or bypass runtime policy checks.