Decision models (optional)

Small classification models that pick among listed options: cheap and fast enough to run on every event, without spending LLM tokens. Optional, but they power intent labels, quality checks and anomaly flags. The owner picks the vendor, or a custom model.

Contract

• Input: {state, question, options}; state is the event or session excerpt.
• Output: one probability per option.
• Runs in the owner's infrastructure when the model is local; only the label is stored.

Vendors

Placeholder, one section each when implemented: d1 (LiquidAI, local), jev, Cloudflare clef, any LLM as fallback.

Where it plugs in

• ../03-analyze/METHOD.md intent labels, per-session classification.
• ../04-auto-analysis/quality.md per-event checks.
• ../02-measure/ optionally at ingest, to label events as they arrive.