Audit skill impact with paired testing

Audit skill behavior via paired probe with and without skill loaded

Detect when a skill changes behavior below pass-rate threshold—surface anchoring, template copy, excess planning, off-task drift.

• · probe_task_bank
• · skill_name
• · task_descriptions with phases

• · SIP report markdown
• · phase_comparison table
• · retire/refine/keep recommendation

• · Skill has been active ≥3 sessions
• · Probe-task bank curated (3-5 tasks)
• · Phase decomposition rules defined

• · Workflow guides (loop, schedule) trivially show task-recovery
• · No baseline if skill just created

• · Based on arxiv 2605.11946v1 (CTA)
• · Detects 522 behavioral changes while pass-rate moves +0.3%