Audit de cohérence Wasserstein

Scanne les manuscrits et le code pour détecter les incohérences entre W1 et W2 dans les métriques de Wasserstein, et produit un rapport.

Spar Skills Guide Bot
TestingIntermédiaire
1027/08/2026
Claude Code
#wasserstein#audit#consistency#manuscript#code

Recommandé pour

/wasserstein-audit — Audit Wasserstein Order Consistency

Scan all manuscript drafts and codebase for W₁ vs W₂ inconsistencies. Resolves the blocking Phase 0 deliverable: the legacy P01 manuscript uses W₁ in several places while project convention mandates W₂ as the primary metric.

Usage

/wasserstein-audit
/wasserstein-audit [papers|code|all]

Example: /wasserstein-audit all


What this does

  1. Reads papers/shared/notation.md (canonical notation standard, Wasserstein Audit section)
  2. Scans manuscript .md files under papers/ for W₁, unsubscripted W, ambiguous W_p
  3. Scans Python files under all domain packages for Wasserstein calls without explicit p=2
  4. Cross-references findings against what notation.md already has "verified"
  5. Produces a structured reconciliation report — does not fix anything without confirmation

Manuscript patterns that trigger a flag

| Pattern | Issue | |---|---| | W_1 / $W_1$ / W_{1} | W₁ usage — needs justification or must change to W₂ | | $W$ (unsubscripted) | Always flag — violates notation standard | | bottleneck without alongside Wasserstein | Sole metric violation | | wasserstein without explicit order nearby | Ambiguous |

Code patterns that trigger a flag

| Pattern | Issue | |---|---| | wasserstein_distance(a, b) with no p= arg | Default may not be W₂ in all library versions | | p=1 or order=1 near Wasserstein call | Explicit W₁ usage | | bottleneck_distance(...) | Must not be used as sole metric | | gudhi.wasserstein.wasserstein_distance without order=2 | gudhi default not guaranteed |


Report format

## Wasserstein Audit Report — YYYY-MM-DD

### Manuscript Findings
| File | Line | Found | Issue | Suggested Fix |

### Code Findings
| File | Line | Found | Issue | Suggested Fix |

### Status
- Blocking Phase 0: YES / NO
- Total issues: N (manuscript: N₁, code: N₂)

### Recommended Actions

Library defaults (reference)

  • persim.wasserstein_distance — p=2 by default ✓
  • gudhi.wasserstein.wasserstein_distance — requires explicit order= arg
  • giotto-tda Wasserstein vectorisation — p=2 by default ✓
  • W₁ in theory proofs is acceptable in Methods sections if clearly labelled and distinguished from the computational metric used
Skills similaires