SysMoBench

Transition validation (TV) for TLA+ specs. Use when: scoring how faithfully an AI-generated spec models a real system, producing per-action pass rates with defensible explanations. The agent acts as the examiner (考官) — writes instrumentation, runs harness, writes TV modules, runs TLC, and interprets scores.

Trace harness generation for SysMoBench. Use when bootstrapping a new task: clone the system into artifacts/<task>/, instrument it to emit NDJSON traces at the task-required granularity, write a run.sh, and produce INSTRUMENTATION.md. One-time work per task; the resulting harness is reused for every spec evaluation via the tv-eval skill.

Bounded repair of TLA+ specs that fail P1 (SANY) or P2 (TLC from Init). Use when a model-generated spec needs to pass P1/P2 so downstream P3 (TV) and P4 (invariant) can score it. Enforces a strict allow-list of mechanical fixes and MUST NOT alter the model's semantic intent (actions, guards, variable set). Every edit must carry a written justification.