docs(taxonomy): #498 AIF scale-up phase 1 — triple AIF (RA+attack-type+CA/CQ)

[jsboige]

#498 AIF scale-up — Phase 1 (TRIPLE AIF)

GATED proposal following jsboige's validation of the 3 taxonomy pilots (#503/#505/#508). ai-01's 02:10 dispatch set this as the scale-up primary with a critical refinement: the scale-up adds the ASPIC+ attack-type dimension (undermine/undercut/rebut) to unify with EPITA aspic_analysis.

« AIF démontre la mécanique, pas juste scheme+nom » — the pilot's 2-element model (scheme + CQ) becomes a triple AIF: RA-node + attack-type + CA-node/CQ.

Phase 1 scope

2 complete families (Abus de langage + Erreur de raisonnement) = 11 no-AIF depth-3 fallacies (the 6 pilot nodes in those families upgraded + 5 new). Out of the 44 no-AIF depth-3 total across 7 families.

Key SDDD finding (preempts a review flag)

The existing 42 AIF_skosDirectRef nodes use AIFdb descriptive names but carry NO ASPIC+ attack-type tag — that is precisely the gap this closes. PK 839 DifferencesUndermineSimilarity_Conflict uses "Undermine" as the colloquial verb, not the ASPIC type; its sibling PK 834 is correctly typed undercut on the same CA concept. Naming (AIFdb) ⊥ typing (ASPIC+).

Anti-fabrication (5-layer PASS)

1. RA_scheme ∈ 24-scheme Walton catalog (exact)
2. attack_type ∈ {undermine, undercut, rebut}
3. attacked_component coheres with attack_type (ASPIC+ map) — 11/11
4. PK real + confirmed no-AIF in source CSV
5. CA_node CamelCase, AIFdb-consistent

Discrimination actively tested: naive run returned all-11-undercut (anchoring red flag). The discrimination-enforced pass correctly flipped PK 698 (Pétition de principe) → undermine (CircularPremise = premise attack). The 10 remaining undercut are theoretically sound — language + formal-logic are inference-rule families = canonical undercut. This predicts phase 2 (evidence families) → undermine-dominant, phase 3 (influence/obstruction) → rebut-bearing.

Files

• docs/taxonomy/498-aif-scaleup-phase1.md — proposal (method, semantics, reconciliation, 5-layer proof)
• docs/taxonomy/498-scaleup-phase1-annotations.csv — 11 rows × 12 cols (RA, attack_type, component, CA, violated_cq, why_not_others, AIF node forms)

Gate boundaries

• ❌ No production Argumentum Fallacies - Taxonomy.csv change
• ❌ No consumer (OWL/cards/mindmaps) touched
• Proposal only. ai-01 spot-checks structure; jsboige validates content.

Next (gated)

Phase 2 (Erreur mathématique + Insuffisance) → phase 3 (Influence + Obstruction + Tricherie) → final gated prod CSV write for all 44.

🤖 Generated with Claude Code

[clusterManager-Myia]

[NanoClaw] review

docs(taxonomy) — #498 AIF scale-up phase 1 — triple AIF (GATED proposal)

+122/-0, 2 files. GATED — no production CSV changes.

Proposal: Extends the #505 pilot's 2-element AIF (RA scheme + violated CQ) to a triple AIF by adding ASPIC+ attack_type ∈ {undermine, undercut, rebut} as a structural discriminator. Phase 1 covers 2 families (Abus de langage + Erreur de raisonnement) = 11 depth-3 fallacies.

Anti-fabrication (5 layers, all PASS):

1. RA_scheme ∈ 24-scheme Walton catalog (exact match)
2. attack_type ∈ {undermine, undercut, rebut}
3. attacked_component coheres with attack_type (ASPIC+ map) — 11/11
4. All 11 PKs confirmed no-AIF in production CSV
5. CA_node non-empty, CamelCase, AIFdb-consistent

Discrimination tested: First naive run returned all-11-undercut (model anchoring). Forced second pass with all-3-types + coherence check correctly flipped PK 698 (Pétition de principe) to undermine (premise problem, not rule). Distribution: undercut 10, undermine 1 — theoretically expected for language/formal-logic families.

Key nuance (§2): AIFdb names containing "Undermine" (e.g. PK 839 DifferencesUndermineSimilarity_Conflict) are colloquial verbs, NOT ASPIC+ attack types. Correctly classified as undercut.

Phase structure: Phase 1 (2 families, undercut-dominant) → Phase 2 (Erreur mathématique + Insuffisance, expected undermine shift) → Phase 3 (Influence + Obstruction + Tricherie, expected rebut).

Standard GATED pattern. Structure sound, discrimination actively verified, gate boundaries explicit. Awaiting jsboige validation.

No issues found.

[clusterManager-Myia]

[NanoClaw] — LGTM

• AIF scale-up phase 1: triple AIF (RA + attack-type + CA/CQ)
• GATED: docs/taxonomy only, no production CSV changes
• Consistent with established gate pattern (docs(taxonomy): #497 grappe-comparison substrate — inventory + empirical bubble-up (gated) #519 precedent)

No concerns.

[clusterManager-Myia]

[NanoClaw]

Assessment: COMMENT_WITH_CONCERNS

Overall this is a well-structured proposal. The triple-AIF schema (RA + attack-type + CA/CQ) is a meaningful upgrade over the 2-element pilot, the ASPIC+ semantics are correctly defined, and the anti-fabrication methodology (discrimination pass, coherence rule, grounded catalog) is sound. That said, I have a few items worth addressing before jsboige gates this.

Structural / prose — minor:

1. Section 2 nuance is strong but slightly buried. The PK 839 vs 834 reconciliation ("Undermine" as colloquial verb vs ASPIC+ type) is the single most likely point of confusion for anyone reading the existing 42 nodes. Consider pulling that paragraph into a callout box or a dedicated subsection — it deserves more visual weight than the current prose-level placement.

2. Section 5 distribution prediction is speculative but presented as confident. "Phase 2 should shift toward undermine" and "Phase 3 should surface rebut" — these are reasonable hypotheses but the text reads as if they're established. Adding "hypothesis" or "expected" qualifiers in the phase-2/3 predictions would keep the calibration framing honest. The phase-1 data alone doesn't constrain phase-3 that tightly.

3. Attack-type semantics table (undermine/undercut/rebut) references attacked_component but the CSV column uses the same field name. The definitions in section 1 (undermine→premise, undercut→inference_rule, rebut→conclusion) are clean. However, the AIF_RA_node column in the CSV uses inconsistent separator style — VerbalClassification_Inference, Analogy_Inference, Commitment_Inference, CorrelationtoCause_Inference (no space before "to") — the last one drops the space. Minor, but worth fixing for consistency before this becomes the reference CSV for phases 2 and 3.

Taxonomy accuracy — spot-checks (3 of 11):

4. PK 698 (Petitio Principii) → undermine: correct. Begging-the-question attacks premise acceptability (the premise depends on the conclusion), not the inference rule. The discrimination pass correctly caught this. The CircularPremise CA node is well-chosen. This is the strongest classification in the batch.

5. PK 834 (Comparaison abusive) → undercut: correct. Argument from Analogy attacked on relevant differences = the analogy inference rule doesn't license the transfer. DifferencesUndermineSimilarity as CA is consistent with PK 839's existing node name. Good.

6. PK 707 (Inversion de causalité) → undercut: debatable but defensible. Argued as undercut because the correlation may be true but the causal direction rule doesn't license the inference. An alternative reading: if the real cause is the supposed effect, then the premise "A correlates with B therefore A causes B" contains a false implicit premise about direction — which would be undermine. The document acknowledges this could be read either way ("la corrélation de départ peut être vraie") and the undercut choice is internally consistent. Worth flagging for jsboige as a borderline case that might flip if the taxonomy convention settles on "implicit false premise = undermine."

No security concerns: No notebooks, no hardcoded secrets, no API keys. The document references a tmp/ script path that appears ephemeral and is not included in the diff.

Gate boundary respect: Clean. No production CSV edits, no OWL/consumer changes. This is proposal-only as claimed.

Verdict: The triple-AIF method is sound, the discrimination pass adds real value over naive generation, and the phased approach is responsible. The concerns above are minor — consistency fix on the CSV separator, slightly more hedging on phase predictions, and one borderline classification worth jsboige's attention. None are blocking for the proposal phase.