audit-against-research

audit-against-research — router

Cross-check the repo against the compiled context. Each atom category has its own audit sub-skill with a tight tool whitelist. Per-atom verdicts: IMPLEMENTED / PARTIAL / MISSING / DIVERGENT.

Procedure

1. Load structured context via MCP — do not Read research/research.md directly:

   mcp__paper-compiler__get_paper_context()
   mcp__paper-compiler__graph_stats()
   

2. Enumerate required atoms. For each category present (skip categories the paper doesn't use — a pure theory paper has no procedure atoms):

   • mcp__paper-compiler__trace_dependency(component=<category>) returns the ranked atoms.
   • Cap at top 25 by priority.

3. Run the per-category audit sub-skill for each category present:

   • audit-method — algorithmic/structural unit
   • audit-objective — loss / Hamiltonian / yield / fitness
   • audit-data — datasets + preprocessing
   • audit-procedure — training loop / integrator / experimental procedure + parameters
   • audit-evaluation — metrics & statistical tests
   • audit-baseline — published comparisons
   • audit-theory — theorems / assumptions / principles

4. Cross-check open assumptions. mcp__paper-compiler__list_missing_details(). Every open assumption must show as a visible choice in the code — comment, named config value, or TODO.

5. Write audit-report.md at the repo root using the template at references/audit-report-template.md.

Rules

• Flag DIVERGENT only with evidence — cite atom_uid (stable across rebuilds) + chunk_id + paper_id. Stylistic disagreement is not divergence.
• Do not auto-fix. The user decides.
• Keep the report short — the user reads it next session.