"Skill Use or Skill Theater": agents report using skills that provably didn't change their decisions.

The BACKTRACE framework pairs every skill-conditioned answer with a matched no-skill counterfactual, intervenes on skill meaning, wording, identity, content, and assignment, then elicits the agent's attribution only after the answer is committed. Stated skill use stays stable while actual causal reliance varies wildly: both silent uptake and purely performative use. Every observational detector failed: direct claims, text mentions, trace similarity, LLM judge. In multi-agent settings, skill influence survived communication after its source was lost, and no-skill teams named skills that were never supplied. If you're auditing an agent-skill marketplace by reading traces, you're auditing theater.