choosing-modelslisted

# Choosing Models Capacity dispatch. choosing-tools decides which skill owns a task; this skill decides how much model the task gets. The output is a **(model, effort) pair** — model is capacity, effort is thinking depth, and the two are chosen together, per task, at the moment work is delegated or priced. This is a **flexible** skill: the procedure below is the default shape of the decision; the judgment inside each step is yours. ## Where the decision is taken Consulting the rubric costs what the deciding session costs, and that runs opposite to what it is worth: across three deciding tiers the measured premium per task spans two orders of magnitude, while agreement with the session's unaided choice rises as the decider gets dearer. So: - **A strong-tier session routing a single task skips the rubric.** It lands on the tier it would have chosen unaided, at the highest price of any deciding context — a break-even no correction rate reaches. - **Score batches, and score them at the weak tier.** The fixed cost amortises there, and that is the only deciding context where the scoring changed any decision at all. It is also the tier where the rubric was seen misapplied, so check an emitted tier against the thresholds. Numbers: `models.toml`. ## The procedure 1. **Score the task** with [references/scoring-rubric.md](references/scoring-rubric.md), at authoring or spawn time. The rubric owns *how to score* and never moves without calibration evidence. 2. **Map