docs(taxonomy): #499 Virtues phase 2 batch 5 — Argument pertinent leaves (gated)

[jsboige]

#499 — Virtues Scale-up, Phase 2, batch 5: Argument pertinent (gated proposal)

No production CSV change. Gated proposal — ai-01 reviews structure, jsboige validates content. This PR adds the Phase-2 depth-3..7 leaf annotation for the Argument pertinent family only.

Scope

Batch 5 of Phase 2: Argument pertinent (segment 1, 29 leaves: 9 depth-3 + 13 depth-4 + 5 depth-5 + 2 depth-6) — the largest single batch so far. Mirrors Insuffisance (Fallacy family 1) — the relevance/sufficiency family (empty argument, bare assertion, hasty conclusion, biased sample, appeal to authority).

Method (identical to batches 1–4)

• Model: gpt-5.5 (/v1/responses, reasoning.effort=low, no temperature). max_output_tokens raised 4500 → 7000 for this batch — 4500 truncated mid-JSON on the 29-node generation; 7000 produced complete output. Note for the sub-batches of the 2 large families (Raisonnement valide 51, Échange enrichissant 40): budget 7000 or sub-batch to avoid truncation.
• Grounding catalog: 7 Fallacy families + 63 named depth-3 fallacies (reused verbatim from Phase 1). Prompt forbids any PK outside the catalog.
• Family mirror hard constraint: prevented_family_pk = 1 for every row.

Anti-fabrication guarantee

Three independent layers, all re-checked against the REAL corpus (not the prompt catalog):

1. Catalog membership — every opposed PK ∈ the 63 depth-3 set.
2. Ground-truth — every PK re-verified against the real 1408-row Fallacies CSV (text_fr), opposed-PK↔name cross-checked character-for-character.
3. Mirror consistency — every row's prevented_family_pk = 1.

Result: 29/29 annotated, 0 violations across all three layers. [PASS]

Anti-fab validator #518 (tools/validate_taxonomy_annotations.py, kind=virtues): ✓ CLEAN — 29 rows, 0 HARD, 0 WARN.

Semantic check

Scheme distribution is the broadest yet (10 distinct schemes): Position to Know ×6, Expert Opinion ×5, Sign ×4, Bias ×4, Cause to Effect ×4, Example ×2. The Position-to-Know + Expert-Opinion dominance (11/29) is the expected signature for the source-evaluation family. The most-opposed fallacies — Argument d'autorité 71 and Fausse attribution 942, each appearing 11 times — are the family's semantic anchor: a relevant argument requires the source be both competent (not Argument d'autorité) and faithfully reported (not Fausse attribution). The Cause-to-Effect cluster (×4: pk 24 Hypothèse plausible, pk 31 Représentation parcimonieuse, pk 32 Rasoir d'Ockham, pk 33 Rasoir de Hanlon) opposes over-complication and pseudo-explanation. Depth reaches d6 (pk 18 Expert du domaine, pk 23 Preuves empiriques) — the deepest leaves annotated so far. Confirms the pipeline tracks each family's genuine topical range.

Files

• docs/taxonomy/499-virtues-scaleup-phase2-argument.md — proposal
• docs/taxonomy/499-scaleup-phase2-argument-annotations.csv — 29-row annotation (10-col presentation schema, UTF-8 BOM for parity)

Generation scripts / raw model output kept ephemeral in tmp/ (not committed).

Phase 2 progress

Family	leaves	status
Langage exact	14	✅ batch 1 merged (fc2013fc)
Rigueur mathématique	16	✅ batch 2 merged (62a28efc)
Honnêteté intellectuelle	23	⏳ batch 3 (PR #534)
Présentation intègre	21	⏳ batch 4 (PR #535)
Argument pertinent	29	✅ this PR
Échange enrichissant	40	next (sub-batched)
Raisonnement valide	51	next (sub-batched)

Cumulative: 103/194 leaves (53%) — past halfway. No prod write until jsboige approves the full Phase-2 set.

Closes nothing standalone — part of #499.

🤖 Generated with Claude Code

[clusterManager-Myia]

[NanoClaw]

LGTM (structure + anti-fab). Virtues phase 2 batch 5 — famille « Argument pertinent » (29 leaves, depth 3-6). Même méthode, même schéma 10-col, même garantie anti-fabrication que les batches 1-4 (déjà reviewés).

Vérifications :

• Anti-fab validator #518 : CLEAN — 29 rows, 0 HARD, 0 WARN.
• 3 layers de vérification contre le VRAI corpus (Fallacies 1408 lignes), pas le prompt : appartenance catalogue (63 depth-3 fallacies), ground-truth PK↔text_fr char-par-char, cohérence miroir.
• Mirror : les 29 leaves ont prevented_family_pk=1 (Insuffisance) — miroir correct de « Argument pertinent ».
• Schémas Walton plausibles (Sign, Rule, Example, Expert Opinion, Position to Know, Bias, Witness Testimony, Verbal Classification, Cause to Effect pour Ockham/Hanlon).
• Pas de changement prod : GATED proposal, n'ajoute que le CSV annotation + doc de proposition. « No prod Virtues CSV change until jsboige approves ».

Security scan : seul hit = « key live 2026-06-18 » dans la description méthodologique (référence au fait que la clé API OpenAI était active), aucune valeur de secret/token réelle. False positive. La sémantique de contenu (mapping Walton) reste le gate de jsboige comme le doc le stipule — mon review est structural. RAS structurel.