model_mismatch

Detects: a premium model used for a step that's really just picking a tool — a step a cheaper model would very likely handle equally well.

Public routing guidance suggests 60–70% of agent calls suit a smaller model; this detector looks for the clearest, most defensible slice of that: calls whose whole job was a short, structured tool-selection decision.

How it works

A call is a downgrade candidate when its iteration produced a tool call (its job was picking a tool, not open-ended reasoning) and its own output is small (150 tokens or fewer by default) — the "tool-only step, short structured output" archetype. There's no task-type classifier in this codebase, so this narrow structural signal is the only "clearly trivial step" indicator available without guessing at intent.

By default, this detector produces nothing at all without evidence. require_quality_map defaults to true: recommending a model swap is consequential enough that the project's own design treats "no evidence the cheaper model actually works here" as a reason to stay silent, not merely to caveat the finding. Supply a --quality quality.json with a downgrade_evals entry ("tool_select@<candidate-model>": {"pass_rate": 0.97}) showing the candidate model passes at or above the configured threshold, and the detector starts reporting real, dollar-quantified findings.

examples/quality.json is a working, real example of the full shape (both tasks for the score gate and downgrade_evals for this detector) — try it:

uv run wattage report examples/sample_trace.json --quality examples/quality.json

Fix

Route this specific step to the configured cheaper candidate model (claude-haiku-4-5 for Anthropic, gpt-5.6-luna for OpenAI, by default — configurable per provider).

Quality risk: review

A model downgrade is one of the two waste patterns the project's own scoring rules explicitly single out as needing evidence before it can count toward your Token Efficiency score — the other is reasoning_overspend.