By name: replay-graph-llm-review description: Reviews replay, graph projection, reporting, and LLM evidence-package behavior without treating generated or inferred output as verified evidence.
Skill: Replay Graph LLM Review
Description
Reviews replay, graph projection, reporting, and LLM evidence-package behavior without treating generated or inferred output as verified evidence.
Instructions
- Locate replay, graph, report, or LLM-related code and documentation.
- Identify the explicit evidence inputs.
- Check how missing or incomplete evidence is represented.
- Check deterministic ordering and provenance preservation.
- Check that graph projections and reports remain derived views.
- Check that LLM prompts are built from structured evidence and that LLM output is labeled as generated analysis.
- Define tests that do not require external graph databases or live LLM providers unless explicitly isolated.
Expected Inputs
- domain and application evidence models
- graph, replay, report, or LLM adapters
- prompt templates or report templates
- related tests
- QUALITY.md
Expected Outputs
- evidence-integrity findings
- uncertainty-handling risks
- provider-isolation risks
- test plan
- verification commands
Stop Conditions
Stop if:
- runtime execution would be inferred from static facts alone
- graph edges would be invented
- replay steps would hide missing trace events
- LLM output would be treated as confirmed fact
- live providers would be required for default unit tests