v1.12.0

Added

• Model routing (scripts/model-route.mjs) and the routing doc. Score a coding task before any tokens are spent on it and recommend a cheap, mid, or premium model tier. otito answers the repository half deterministically (AX, containment, canonical risk flags); a System One model answers the request half with calibrated probabilities over three narrow typed questions. Host agnostic by design: --tier-only prints the tier for any host to map, --host <id> prints that host's model name, and maps live in .otito/model-route.json so nothing about the scoring is specific to one IDE.
  • The router runs before work starts and never touches the gate. It decides how much model to spend; the gate decides whether a change may merge. A failed, unreachable, or unkeyed model call falls back to a local estimate that labels itself uncalibrated, and costs a tier, never a verdict. The gate never asks.
  • It ships advisory: it prints a recommendation and does not pick a model. The share weights and the band thresholds were chosen by judgement and have never been compared to an outcome, which is the same critique the calibration thesis makes of inferRisk. Promoting it past advisory needs a backtest measuring regret, tasks routed cheap that ended in a revert or repair, against the spend avoided.
  • Dogfooded on a production Next.js application, where the first pass routed every request to premium and surfaced two defects worth recording. A question whose answer never moves carries no information however well calibrated it is: asked as a yes/no, "is this request ambiguous enough that a wrong reading produces the wrong change?" returned 0.57 to 0.81 for every request including a typo fix. And band thresholds belong to the quantity they were drawn for: subtracting flat points from AX while still banding at AX's own 45/75 is sufficient on its own to push a whole corpus into the most expensive tier.

Changed

• docs/index.md lists the calibration thesis. Doc 17 shipped in 1.11.0 without its row in the documentation pack table.