compare-corpora

compare-corpora — diff two compiled papers

Manual-only. The MCP server serves one corpus at a time; this skill flips the bound research dir between calls via bind_research_dir. Output is a single comparison-report.md in the user's cwd.

Inputs

$ARGUMENTS is two --research-dir paths separated by space (or two repo roots; the skill walks for research/ if given a repo root):

/paper-compiler:compare-corpora /path/to/jepa_impl /path/to/dreamer_impl

If only one path is given, ask the user for the second before proceeding.

Procedure

1. Validate both artifacts. For each side:

   • Confirm <path>/research/build-manifest.json exists.
   • bind_research_dir(path="<path>/research") then get_paper_context() → read title, paper_id, atom counts, communities.

2. Build a side-by-side category map. Run trace_dependency(component=<c>) for each c in {method, objective, data, procedure, evaluation, baseline, theory} on both sides; collect atoms by (category, canonical_name).

3. Align atoms across corpora. Two atoms are an "aligned pair" if their canonical names (lowercased, punctuation-stripped) match. The compiler's stable atom_uid includes the defining paper id, so two papers defining the same concept have different uids — alignment is by name + category only.

   • BOTH — concept appears in both papers.
   • A_ONLY — only in paper A.
   • B_ONLY — only in paper B.

4. Surface contradictions. For each side, query edges with best_role='contradicts' OR citation_intent='contrasts' via the MCP server. Highlight any case where paper A contradicts paper B (or vice versa) by checking if the contradicted-edge target paper is the other side's target_paper_id.

5. Community overlap. list_communities() for both sides; if any community labels are similar (cosine on title? simple lowercase substring match suffices), flag it — the two papers cluster the same neighborhood.

6. Emit comparison-report.md in the user's cwd (NOT in either research/ artifact, since this is a side-output):

   # Comparison: <paper-A> vs <paper-B>
   Compiled A: <date>, B: <date>
   
   ## Summary
   - A: <N_atoms> atoms, <N_communities> communities, <coverage>% coverage
   - B: <N_atoms> atoms, <N_communities> communities, <coverage>% coverage
   - Aligned atoms (BOTH): <count>
   - A-only: <count>; B-only: <count>
   
   ## Per-category alignment
   ### method
   | BOTH | <atom_a.uid> ↔ <atom_b.uid> | <name> |
   | A_ONLY | <atom_a.uid> | <name> |
   | B_ONLY | <atom_b.uid> | <name> |
   
   ### objective, data, procedure, evaluation, baseline, theory
   ... (same shape)
   
   ## Contradictions
   - Paper A's atom `<uid>` (<name>) contradicts paper B's chunk_id=<N> ...
   - (or "_clean — no cross-paper contradictions detected_")
   
   ## Community overlap
   - A-community-0 (JEPA World Models) overlaps B-community-2 (Latent Predictive Models)
   

7. Append a compare event to both papers' wiki/log.md.

Rules

• Serial only. This skill runs one paper at a time; concurrent compare-corpora calls would corrupt the server's bound research dir.
• No DB writes. All output lands in the user's cwd. Neither corpus is mutated.
• Cite uids. Every atom in the report carries its full atom_uid — they're stable, and the user can get_evidence(atom_uid=...) from either bound research dir to deep-dive.