Skip to content

docs(taxonomy): #498 AIF scale-up phase 3 — Influence + Tricherie + Obstruction (27 fallacies, triple-AIF, rebut-type appears) #546

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
jsboige merged 1 commit into from
Jun 19, 2026
Merged

docs(taxonomy): #498 AIF scale-up phase 3 — Influence + Tricherie + Obstruction (27 fallacies, triple-AIF, rebut-type appears) #546

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
120 changes: 120 additions & 0 deletions docs/taxonomy/498-aif-scaleup-phase3.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,120 @@
# #498 — AIF Scale-up, Phase 3: Influence + Tricherie + Obstruction depth-3 (27 fallacies, triple-AIF) (GATED proposal)

**Issue:** [#498 — AIF generative scale-up: triple-AIF (RA-node + attack-type + CA-node/CQ) for all depth-3 fallacies](https://github.com/ArgumentumGames/Argumentum/issues/498)
**Author:** Claude Code @ myia-po-2024 (worker)
**Date:** 2026-06-19
**Base:** master `909d04c3`
**Status:** **GATED PROPOSAL** — ai-01 reviews structure, jsboige validates content. **No production Fallacies CSV change until jsboige approves.** This document + the annotation CSV are the proposition.

**Calibration:** [`498-aif-scaleup-phase1.md`](498-aif-scaleup-phase1.md) — Phase 1 (11 fallacies, Erreur de raisonnement + Abus de langage) **merged** (`cae93dc8`). [`498-aif-scaleup-phase2.md`](498-aif-scaleup-phase2.md) — Phase 2 (12 fallacies, Erreur mathématique + Insuffisance) **OPEN** (PR #542, CLEAN). This is **Phase 3: the rhetorical/divergence families** (Influence + Tricherie + Obstruction, 27 fallacies) — the rebut-bearing cluster. With phases 1+2 (23 fallacies) and phase 3 (27), AIF depth-3 coverage reaches **50/63 (79%)**.

---

## 1. Scope of Phase 3

Phase 3 extends the triple-AIF annotation to the **rhetorical/divergence families** — the fallacies of persuasion, deception, and discussion-derailment. These are the **rebut-bearing** families: where phases 1–2 (evidence/premise + rule/language families) were undermine/undercut-dominant, phase 3 is where the **rebut** attack-type (absent in phases 1–2) finally appears — for the direct-counter-claim and tu-quoque cases of Obstruction (`Relativisme abusif`, `Évasion`, `Procès en incohérence`).

**27 fallacies** = the 9 depth-3 leaves of each of the 3 families (Influence pks 177–511, Tricherie pks 889–1242, Obstruction pks 1282–1398). All 27 carry no AIF yet — these three families are entirely unannotated before this batch.

| AIF phase | Families | Fallacies | attack-type signature | status |
|---|---|---|---|---|
| Phase 1 | Erreur raisonn. + Abus langage | 11 | undercut-dominant | ✅ merged `cae93dc8` |
| Phase 2 | Erreur math + Insuffisance | 12 | undermine 6 / undercut 6 / rebut 0 | ⏳ PR #542 OPEN CLEAN |
| **Phase 3 (this)** | **Influence + Tricherie + Obstruction** | **27** | **undermine 20 / undercut 4 / rebut 3** | **⏳ this batch** |
| Phase 4 (mopup) | stragglers (4 Abus lang + 3 Err raisonn + 3 Err math + 3 Insuff) | 13 | — | next |
| **Total depth-3** | 7 families | 63 | — | **50/63 after phase 3 (79%)** |

## 2. Schema — identical 12-col triple-AIF schema (phases 1–2, unchanged)

```
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
```

Same conventions as phases 1–2: `RA_scheme` = exact Walton scheme (24-catalog); `attack_type` ∈ {undermine→premise, undercut→inference_rule, rebut→conclusion} (ASPIC+ coherence, attack_type ↔ attacked_component); `CA_node` = CamelCase AIFdb conflict concept; `AIF_RA_node` = scheme + `_Inference`; `AIF_CA_node` = CA_node + `_Conflict`; `violated_cq`/`why_not_others`/`justification` = FR. The eventual prod write (AIF cells to the production Fallacies CSV) is the separate final gated step.

## 3. The triple-AIF paradigm (honored per fallacy, same as phases 1–2)

A fallacy is a **defeasible exception** to a legitimate Walton scheme. The triple-AIF encodes HOW it defeats the scheme's legitimate inference: (1) **RA_scheme** = the canonical Walton scheme the fallacy mimics/abuses (the legitimate move it counterfeits); (2) **attack_type** = the ASPIC+ defeat mode — *undermine* attacks a premise's truth/acceptability, *undercut* attacks the inference rule's applicability, *rebut* raises a counter-conclusion; (3) **CA_node** = the specific critical question violated. The `why_not_others` field forces active per-case discrimination, preventing collapse to a single attack type.

## 4. Method & anti-fabrication guarantee (identical to phases 1–2)

- **Model:** gpt-5.5 (OpenAI direct, key live 2026-06-19). `POST /v1/responses`, `reasoning:{effort:"low"}`, `max_output_tokens:7000`, no `temperature`. Answer read from `output[].content[].text`.
- **Grounding catalog:** the **24 canonical Walton schemes** (Walton/Reed/Macagno 2008, reused verbatim from the phase-1 dataset). The prompt forbids referencing any scheme outside this catalog.
- **ASPIC+ coherence as hard constraint:** every row's `attack_type` ↔ `attacked_component` pair coheres per the undermine→premise / undercut→inference_rule / rebut→conclusion map.
- **Verification (three independent layers, all re-checked against the REAL corpus):**
1. *Catalog membership* — every `RA_scheme` ∈ the 24 Walton set; `attack_type` ∈ the 3 types; node suffixes `_Inference`/`_Conflict`.
2. *Ground-truth* — every PK re-verified against the real **1408-row** Fallacies CSV (`PK` + `text_fr` + `Famille` + `depth`); pk ↔ name ↔ family ↔ depth=3 cross-checked; all 27 present, unique, no missing/extra.
3. *ASPIC+ coherence* — attack_type ↔ attacked_component per the map.
- **Result: 27/27 annotated, 0 violations across all three layers.** [PASS]
- **Anti-fab validator #518** (`tools/validate_taxonomy_annotations.py`): **`✓ CLEAN — 27 rows, 0 HARD, 0 WARN`** (kind=aif-scaleup).

Generation/verify script: `tmp/498_phase3_gen.py` (ephemeral, not committed — derived from the phase-2 `tmp/498_phase2_gen.py`). Dataset: `tmp/498_phase3_dataset.json` (ephemeral). Raw model output: `tmp/498_phase3_gen_result.json` (ephemeral).

## 5. The 27 rows (full set in the CSV)

### 5.1 Distribution: undermine 20 / undercut 4 / rebut 3

| Family | undermine | undercut | rebut | signature |
|---|---|---|---|---|
| Influence (9) | 9 | 0 | 0 | pure persuasion = bad/loaded premise |
| Tricherie (9) | 6 | 3 | 0 | deception = false premise + double-standard rule shifts |
| Obstruction (9) | 5 | 1 | **3** | derailment = source attacks + counter-claims/tu-quoque |

**Phase progression confirms the ASPIC+ prediction.** Phases 1–2 (evidence/premise + rule families) produced **rebut 0/23** — no fallacy raised a direct counter-conclusion; they were all bad premises (undermine) or inapplicable rules (undercut). Phase 3 introduces the **rebut type (3 cases)**, exactly where the rhetorical/divergence families license it: `Relativisme abusif` 1282 (a competing-truth counter-claim), `Évasion` 1313 (substituting a different conclusion), `Procès en incohérence` 1361 (tu-quoque counter). The undermine dominance (20/27) is the Tricherie/Obstruction signature — these families counterfeit legitimate schemes via false/biased premises (Mensonge, fausse attribution, attention sélective, biais, attaque personnelle, empoisonnement du puits) and bare assertions (Preuve par assertion).

### 5.2 The 27 rows

| Fallacy (pk) | Family | Walton scheme | attack → component | CA_node | violated CQ (FR) |
|---|---|---|---|---|---|
| Langage persuasif (177) | Influence | Verbal Classification | undermine → premise | LoadedLanguage | Les termes qui cadrent le débat sont-ils neutres, acceptables et non chargés de présupposés contestables ? |
| Humour (219) | Influence | Popular Opinion | undermine → premise | EmotionalAppeal | La sympathie ou l’amusement suscité par l’orateur repose-t-il sur une raison pertinente et non sur une simple séduction émotionnelle ? |
| Poésie (247) | Influence | Values | undermine → premise | RhetoricalEmbellishment | Les valeurs invoquées sont-elles appuyées par des raisons pertinentes plutôt que par un embellissement stylistique destiné à séduire ? |
| Connivence (300) | Influence | Values | undermine → premise | EmotionalAlliance | L’appel à une valeur commune repose-t-il sur une raison partageable plutôt que sur une alliance émotionnelle avec l’auditoire ? |
| Repoussoir (322) | Influence | Bias | undermine → premise | DerogatoryPortrayal | La caractérisation négative de l’idée ou de ses défenseurs est-elle pertinente et étayée plutôt que simplement repoussante ? |
| Appel aux conséquences (340) | Influence | Consequences | undermine → premise | RelevanceOfConsequences | Les conséquences invoquées sont-elles réelles, pertinentes et suffisantes pour juger la thèse elle-même ? |
| Conditionnement (357) | Influence | Sign | undermine → premise | AssociativeConditioning | Le signe ou l’association invoqué indique-t-il réellement la propriété conclue plutôt qu’une émotion artificiellement conditionnée ? |
| Jeu de pouvoir (420) | Influence | Commitment | undermine → premise | CoercedCommitment | L’engagement attribué à l’interlocuteur est-il librement accepté plutôt qu’imposé par une pression affective ou psychologique ? |
| Influence non verbale (511) | Influence | Sign | undermine → premise | NonVerbalManipulation | Les signes non verbaux utilisés sont-ils des indices pertinents de la thèse plutôt que des moyens de déstabilisation ou de pression ? |
| Mensonge (889) | Tricherie | Position to Know | undermine → premise | PremiseTruth | L’assertion présentée comme connue est-elle vraie et sincèrement soutenue par quelqu’un en position de la savoir ? |
| Fausse attribution (942) | Tricherie | Expert Opinion | undermine → premise | SourceCredibility | La source citée existe-t-elle, est-elle correctement identifiée et possède-t-elle l’expertise pertinente ? |
| Attention sélective (953) | Tricherie | Example | undermine → premise | SelectiveEvidence | Les exemples ou faits retenus sont-ils représentatifs de l’ensemble des données pertinentes, y compris celles qui contredisent la thèse ? |
| Exigence renforcée (974) | Tricherie | Rule | undercut → inference_rule | MovingGoalposts | Le critère ou la règle d’évaluation reste-t-il stable une fois que l’interlocuteur y a satisfait ? |
| Vouloir le beurre et l’argent du beurre (992) | Tricherie | Commitment | undercut → inference_rule | EquivocalCommitment | La position ou l’engagement invoqué est-il suffisamment clair pour permettre d’en tirer une conséquence argumentative contrôlable ? |
| Exigence relâchée (1011) | Tricherie | Rule | undercut → inference_rule | LoweredStandards | Le critère d’évaluation appliqué est-il maintenu au même niveau au lieu d’être abaissé pour sauver la position défendue ? |
| Biais naturels (1024) | Tricherie | Cause to Effect | undermine → premise | AnthropocentricBias | La relation causale ou téléologique invoquée repose-t-elle sur des faits plutôt que sur une projection anthropocentrique ? |
| Biais culturels (1174) | Tricherie | Values | undermine → premise | CulturalBias | L’évaluation proposée repose-t-elle sur des valeurs justifiées plutôt que sur les présupposés particuliers d’un contexte culturel ? |
| Biais théoriques (1242) | Tricherie | Rule | undermine → premise | TheoreticalBias | La règle ou le concept théorique appliqué repose-t-il sur une prémisse adaptée au cas plutôt que sur un cadre rigide et déformant ? |
| Relativisme abusif (1282) | Obstruction | Commitment | **rebut → conclusion** | CounterConclusionConflict | L’affirmation relativiste contredit-elle indûment la conclusion commune en niant qu’une vérité discutable puisse valoir au-delà d’une perspective individuelle ? |
| Pseudo-explication (1287) | Obstruction | Cause to Effect | undermine → premise | PseudoExplanation | L’explication proposée identifie-t-elle réellement des causes pertinentes plutôt que de reformuler ou d’escamoter ce qui doit être expliqué ? |
| Preuve par assertion (1297) | Obstruction | Position to Know | undermine → premise | BareAssertion | L’assertion répétée est-elle appuyée par des raisons ou des connaissances indépendantes plutôt que par sa seule insistance ? |
| Évasion (1313) | Obstruction | Practical Reasoning | **rebut → conclusion** | IssueAvoidance | La réponse proposée traite-t-elle la question à résoudre au lieu de substituer une autre conclusion ou un autre agenda de discussion ? |
| Complication exagérée (1345) | Obstruction | Practical Reasoning | undercut → inference_rule | ExcessiveComplexity | La complexité invoquée empêche-t-elle réellement l’application d’un raisonnement pratique ou sert-elle à rendre indécidable une question traitable ? |
| Empoisonnement du puits (1352) | Obstruction | Bias | undermine → premise | PoisoningTheWell | L’information préjudiciable invoquée établit-elle un biais pertinent de la source plutôt qu’un discrédit préalable et non probant ? |
| Procès en incohérence (1361) | Obstruction | Commitment | **rebut → conclusion** | TuQuoque | L’incohérence alléguée de l’interlocuteur réfute-t-elle réellement sa conclusion plutôt que de lui opposer seulement son comportement ou ses engagements ? |
| Sophisme génétique (1371) | Obstruction | Bias | undermine → premise | GeneticAdHominem | L’origine de l’idée ou de la personne établit-elle un biais pertinent qui affecte le contenu de la thèse ? |
| Attaque personnelle (1398) | Obstruction | Bias | undermine → premise | PersonalAttack | Le trait personnel attaqué établit-il un biais ou une incompétence pertinente relativement à l’argument présenté ? |

### 5.3 Semantic-fidelity highlights

- **Influence (9/9 undermine)**: every Influence fallacy counterfeits a persuasion scheme via a **loaded or emotion-laden premise** — not a false rule, not a counter-conclusion, but a premise whose acceptability is corrupted (loaded language, emotional appeal, derogatory portrayal, coerced commitment, associative conditioning). The CA-nodes carry this specificity: `LoadedLanguage`, `EmotionalAppeal`, `DerogatoryPortrayal`, `CoercedCommitment`, `AssociativeConditioning`. The schemes disperse across the persuasion catalog (Verbal Classification, Popular Opinion, Values ×2, Bias, Consequences, Sign ×2, Commitment) — Influence abuses *every* persuasion scheme, always at the premise layer.
- **Tricherie (undermine 6 / undercut 3)**: deception splits into **false/biased premises** (Mensonge = PremiseTruth, fausse attribution = SourceCredibility, attention sélective = SelectiveEvidence, biais ×3) and **double-standard rule shifts** (Exigence renforcée = MovingGoalposts, Exigence relâchée = LoweredStandards, Vouloir le beurre = EquivocalCommitment) — the undercut cluster is precisely the *special-pleading* trio where the rule of evaluation itself is the problem.
- **Obstruction (undermine 5 / rebut 3 / undercut 1)**: the **only rebut-bearing family**. The 3 rebut cases are the genuine counter-conclusion fallacies — `Relativisme abusif` (competing-truth), `Évasion` (issue substitution), `Procès en incohérence` (tu-quoque) — each raising a conclusion that conflicts with the RA rather than merely denying a premise. The undermine cases are the credibility attacks (`Attaque personnelle` = PersonalAttack, `Sophisme génétique` = GeneticAdHominem, `Empoisonnement du puits` = PoisoningTheWell) and bare assertions (`Preuve par assertion` = BareAssertion). `Complication exagérée` is the lone undercut (ExcessiveComplexity = the practical-reasoning rule becomes inapplicable).

Machine-readable: [`498-scaleup-phase3-annotations.csv`](498-scaleup-phase3-annotations.csv).

## 6. What this does NOT do (gate boundaries)

- ❌ Does **not** edit the production `Argumentum Fallacies - Taxonomy.csv`. Awaiting jsboige approval.
- ❌ Does **not** touch OWL, EPITA consumer, cards, or mindmaps.
- ❌ Does **not** complete the depth-3 AIF coverage — **13 stragglers remain** (phase 4 mopup: 4 Abus de langage + 3 Erreur de raisonnement + 3 Erreur mathématique + 3 Insuffisance depth-3 leaves not yet annotated).

## 7. Proposed next steps (gated on jsboige)

1. **jsboige validates** the 27 phase-3 rows (paradigm + content + the rebut-type cluster in Obstruction). Note: phase 3 is independent of phase 2's merge — it bases on master and can be reviewed on its own merits.
2. **Phase 4 mopup** (13 stragglers) — the remaining depth-3 fallacies, completing AIF depth-3 coverage to 63/63.
3. On approval of the full AIF set: the **gated prod write** (AIF cells to the production Fallacies CSV) proceeds — writing `RA_scheme`, `attack_type`, `attacked_component`, `CA_node`, `AIF_RA_node`, `AIF_CA_node`, `violated_cq`, `why_not_others`, `justification` to the production CSV.

---

*GATED proposal. Worker signals structure + grounding; ai-01 reviews, jsboige validates content. No production data changed.*
Loading
Loading