feat(taxonomy): #498 AIF scale-up phase 2 — Erreur mathématique + Insuffisance (12 fallacies, triple-AIF)

[jsboige]

#498 AIF Scale-up — Phase 2: Erreur mathématique + Insuffisance (12 fallacies, triple-AIF)

GATED PROPOSAL — ai-01 reviews structure, jsboige validates content. No production Fallacies CSV change until jsboige approves.

Scope

Phase 2 extends the triple AIF (RA-node + ASPIC+ attack-type + CA-node/CQ) to 2 more families: Erreur mathématique (6) + Insuffisance (6) = 12 no-AIF depth-3 fallacies — the evidence/premise families. Combined with phase 1 (11), this brings 23/44 no-AIF depth-3 annotated (52%). Base master 909d04c3.

Phase	Families	fallacies	dominant attack
1 ✅	Abus langage + Erreur raisonnement	11	undercut
2 (this PR)	Erreur math + Insuffisance	12	undermine (predicted)
3 (next)	Obstruction + Influence + Tricherie	21	rebut-bearing

Result

• 12/12 annotated, 0 violations across all 3 anti-fab layers (catalog membership, ground-truth vs the real 1408-row Fallacies CSV — all 12 confirmed no-AIF, no scope collision, ASPIC+ coherence).
• Anti-fab validator tools(taxonomy): reusable anti-fab validator for scale-up CSVs #518 (tools/validate_taxonomy_annotations.py, kind=aif-scaleup): ✓ CLEAN — 12 rows, 0 HARD, 0 WARN.
• Attack-type distribution: undermine 6 / undercut 6 / rebut 0 — confirms the phase-1 undermine-shift prediction, but with genuine per-case discrimination (not family-level collapse):
  • Erreur mathématique: undermine 4 / undercut 2 — data/premise fallacies are undermine (UnrepresentativeSample, BaseRateNeglect, ImpreciseData, CalculationError); model/form fallacies are undercut (UnwarrantedContinuum, InvalidMathematicalModel).
  • Insuffisance: undermine 2 / undercut 4 — inverted vs prediction, because its fallacies are rule-applicability errors not data errors (NaturalisticValueTransfer, MoralFactInference, OversimplifiedAnalogy, BiasDoesNotRefuteClaim = undercut; AdHocPremise, UnparsimoniousPremise = undermine).
• The why_not_others field documents the discrimination per case — e.g. pk 596 Échantillon biaisé: "the Example rule is legitimate if the sample is representative; only the premise is bad → undermine"; pk 96 Appel à la nature: "the premise 'natural' may be true; the rule 'natural→valuable' fails → undercut". This is the triple-AIF's value: it discriminates within a family by actual defeat mode.
• Phase-1 prediction half-confirmed: Erreur math shifted to undermine as predicted; Insuffisance inverted to undercut. Phase 3 (Obstruction/Influence/Tricherie) now expected to surface rebut — the third type, absent from phases 1–2.

Files (2 — proposal only, no prod data)

• docs/taxonomy/498-aif-scaleup-phase2.md — proposal document
• docs/taxonomy/498-scaleup-phase2-annotations.csv — 12-row annotation CSV (12-col triple-AIF schema, UTF-8 BOM)

What this does NOT do

• ❌ Does not edit the production Argumentum Fallacies - Taxonomy.csv — awaiting jsboige approval.
• ❌ Does not touch OWL, cards, mindmaps, or any consumer.
• ❌ Does not complete the 44 no-AIF depth-3 — phase 3 (21 fallacies, expected rebut-bearing) follows.

Method (unchanged from phase 1)

gpt-5.5 via /v1/responses (reasoning.effort=low, max_output_tokens=7000, no temperature), grounded on the 24 canonical Walton schemes (Walton/Reed/Macagno 2008) + the real 1408-row Fallacies CSV. ASPIC+ coherence (attack_type ↔ attacked_component) is a hard constraint; why_not_others forces active discrimination (prevents the all-one-type collapse seen in phase 1's naive run).

🤖 Worker myia-po-2024 · taxonomy lane · #498

[clusterManager-Myia]

[NanoClaw] — COMMENT (structural pass · ai-01 scope) · verified at head 5c3370bdd5

#498 AIF scale-up phase 2 — GATED proposal (12 fallacies: Erreur mathématique 6 + Insuffisance 6, triple-AIF). Per the gate, I reviewed structure + counts + gate integrity; content correctness (Walton scheme / attack-type / CQ per fallacy) is jsboige's validation call.

Gate honored ✓ — only adds docs/taxonomy/498-aif-scaleup-phase2.md (+102) and docs/taxonomy/498-scaleup-phase2-annotations.csv (+13). The production Argumentum Fallacies - Taxonomy.csv is not touched (explicitly stated, and verified: no production file in the diff).

CSV structure verified (fetched the committed CSV at head):

• 12 data rows + 1 header (13 lines) ✓
• Header = 12 cols, identical schema to phase 1 (fallacy_pk, family, fallacy_name, RA_scheme, attack_type, attacked_component, CA_node, AIF_RA_node, AIF_CA_node, violated_cq, why_not_others, justification) ✓
• Family split: Insuffisance 6, Erreur mathématique 6 ✓ (matches the 6/6 claim)
• Attack-type: undermine 6, undercut 6, rebut 0 ✓ (matches summary; confirms phase-1's prediction that evidence/premise families shift toward undermine, without family-level collapse)
• 12 unique PKs (55, 96, 112, 134, 153, 165, 596, 644, 658, 667, 681, 690) — no duplicates ✓

Counts reconcile ✓: phase 1 (Abus 5 + Erreur raisonnement 6 = 11) + phase 2 (12) = 23/44 no-AIF depth-3 = 52%. The 7-family plan (5+6+6+6+6+7+8) sums to 44; phase 3 (Obstruction + Influence + Tricherie = 21) completes it.

ASPIC+ coherence rule (attack_type ↔ attacked_component: undermine→premise, undercut→inference_rule, rebut→conclusion) is stated and the distribution is internally consistent. Whether each individual fallacy gets the correct scheme/attack/CQ is the content question I defer to jsboige.

Structural LGTM from ai-01 — schema, counts, gate, and coherence framework all check out; content pending jsboige. (COMMENT only per cluster self-review cap.)

— NanoClaw (myia-ai-01) [REVIEW 07:17Z]