Intent Model (Preview)
In development. This entry describes direction and engineering progress, not a shipped, generally available feature. No detection-accuracy claims are made here.
What this is building toward
- On-device, privacy-first. The model runs at the endpoint. Conversation content stays on the endpoint. Only structured, de-identified geometry is ever used for learning, never raw text.
- The model proposes, the scope disposes. The Intent Model never makes an enforcement decision on its own. Its judgment shapes a scope that is then enforced deterministically, alongside fixed sensitivity floors, policy ceilings, and a kernel truth channel the agent cannot author. Fooling the model still leaves an attacker inside those deterministic limits.
- Calibrated per policy key. Operating thresholds are calibrated per action class and sensitivity, not by a single global setting, so the decision boundary reflects the real risk of each capability.
Engineering progress
- Established a leak-free evaluation methodology for intent-to-action licensing. The instrument is constructed so that a model cannot appear accurate by reading the action shape alone; it must judge the actual relationship between intent and action. Rigorous, gameable-metric-resistant evaluation is a foundation we are building before making capability claims.
- Confirmed an on-device, no-GPU training path for the everyday model update, keeping the learning loop cheap and the endpoint footprint bounded.
- Upgraded the on-device encoder to a larger, more capable model, chosen and evaluated on the same leak-free instrument. It continues to run at the endpoint, per turn and asynchronously (off the blocking path), and the everyday model update remains a cheap, GPU-free retrain. No conversation text leaves the endpoint.