Python parity of typescript/examples/06-skill-routing. A model-powered system becomes inspectable and comparable once it is a graph. Two routers, compared:

Graph A  task -> naive keyword route -> skill            (no structured context)
Graph B  task -> extract requirements -> select skill    (structured context)

• Each transform records signals (tokens, cost_usd) via ctx.record_signal.
• An exact_match_evaluator scores chosen-skill vs expected-skill.
• The report shows accuracy, cost, and where error enters the trace, plus two compare_runs cases (one B wins, one honest A-wins).

Unlike the TypeScript original, the aggregate line does not print wall-clock duration (the one nondeterministic signal). compare_runs still aggregates it under duration_ms for the caller — it's just not printed, so the output is byte-stable.

python3 python/examples/06-skill-routing/main.py

Captured output: expected-output.txt.

• RunContext.record_signal → TraceStep.metadata (the neutral signal carrier).
• compare_runs(a, b) → status/score/error verdict + aggregated signal deltas + diverged_at_step.
• The accuracy-vs-capacity tradeoff, localized in the trace to select-skill.