Skip to content

docs(taxonomy): #498 AIF generative layer pilot — GATED proposal (18-fallacy, gpt-5.5, anti-fab verified) #505

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 17, 2026
Merged

docs(taxonomy): #498 AIF generative layer pilot — GATED proposal (18-fallacy, gpt-5.5, anti-fab verified) #505

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
90 changes: 90 additions & 0 deletions docs/taxonomy/498-aif-generative-pilot.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,90 @@
# #498 — AIF Generative Layer Pilot (GATED proposal)

**Issue:** [#498 — AIF génératif: structure d'exception pour sophismes sans analogue AIF](https://github.com/ArgumentumGames/Argumentum/issues/498)
**Author:** Claude Code @ myia-po-2023 (worker)
**Date:** 2026-06-16
**Base:** master `d780774e`
**Status:** **GATED PROPOSAL** — ai-01 reviews structure, jsboige validates content. **No production Fallacies CSV change until jsboige approves.** This document + the pilot CSV are the proposal.

---

## 1. The gap & the reformulation (per ai-01)

Only **42/1408** Fallacies nodes carry a direct `AIF_skosDirectRef` (2.9% direct; ~31% effective via grappe-level inheritance). ai-01's reformulation (verified read-only by po-2024) settled the scope:

> *"Le chantier n'est PAS « augmenter la couverture » (impossible depuis AIF, ~20 schemes natifs), mais **génératif** — pour les sophismes sans analogue AIF, rédiger la **structure d'exception** (scheme + CQ) au lieu d'un nom latin."*

AIF's value is **not** its ~20-node catalog — it's the **exception structure**: a fallacy is a *defeasible exception to a legitimate Walton scheme*, escaping one of the scheme's critical questions (CQ). A latin name is consumable by any LLM 0-shot; the *« exception to scheme S / violated CQ Q »* mechanism is not — that's the generative layer this pilot produces.

## 2. Approach — exception-structure annotation

For each pilot fallacy, generate four fields:

| Field | Meaning | Grounding |
|-------|---------|-----------|
| `walton_scheme` | The canonical Walton scheme this fallacy is an exception to | Must be an exact name from the canonical catalog (§4) |
| `violated_critical_question` | The specific CQ of that scheme the fallacy violates | Articulated in the style of Walton's CQs (FR) |
| `exception_mechanism` | How the fallacy derails the scheme by evading that CQ | Generated, FR |
| `justification` | One line linking them | Generated, FR |

This **deliberately does not** invent new AIF corpus node names. It documents the generative exception structure directly — the high-value, non-trivial layer.

## 3. The pilot — 18 fallacies (no prior AIF, all 7 families)

**Selection:** 18 well-known depth-3 fallacies that currently have **no** `AIF_skosDirectRef`, ~2-3 per family, chosen where the Walton exception mapping is demonstrable.

| PK | Fallacy | Walton scheme (exception to) | Violated critical question |
|----|---------|------------------------------|----------------------------|
| 55 | Sauvetage ad hoc | Argument from Rule | L'exception invoquée est-elle prévue par des raisons indépendantes, ou introduite seulement pour sauver la conclusion ? |
| 96 | Appel à la nature | Argument from Values | La valeur « naturel » est-elle pertinente pour établir que la chose est meilleure, malgré sécurité/efficacité/justice ? |
| 112 | Sophisme moraliste | Argument from Values | La valeur morale invoquée est-elle pertinente pour la vérité factuelle, ou seulement pour l'évaluer/condamner ? |
| 322 | Repoussoir | Argument from Values | La valeur négative rendant l'idée repoussante est-elle pertinent pour l'évaluer rationnellement ? |
| 340 | Appel aux conséquences | Argument from Consequences | Les conséquences invoquées sont-elles pertinentes pour la vérité/falsité de la proposition ? |
| 596 | Échantillon biaisé | Argument from Example | L'échantillon est-il représentatif de la population cible de la conclusion ? |
| 644 | Probabilités faussées | Argument from Sign | La probabilité est-elle fiable, correctement calculée et pertinente ? |
| 698 | Pétition de principe | Argument from Commitment | La prémisse-engagement est-elle acceptée indépendamment de la conclusion à prouver ? |
| 707 | Inversion de causalité | Argument from Cause to Effect | La direction causale est-elle correctement établie, plutôt qu'inversée ? |
| 784 | Syllogisme invalide | Argument from Rule | La règle + le cas autorisent-ils réellement la conclusion tirée ? |
| 834 | Comparaison abusive | Argument from Analogy | Existe-t-il des différences pertinentes rendant l'analogie trompeuse ? |
| 847 | Amphibologie | Argument from Verbal Classification | La formulation attribue-t-elle clairement, sans ambiguïté, le prédicat au cas ? |
| 855 | Équivoque | Argument from Verbal Classification | Le terme a-t-il le même sens dans les prémisses et la conclusion ? |
| 889 | Mensonge | Argument from Witness Testimony | Le témoin est-il sincère et dit-il la vérité telle qu'il la connaît ? |
| 942 | Fausse attribution | Argument from Expert Opinion | La source a-t-elle affirmé ceci et possède-t-elle l'expertise pertinente ? |
| 974 | Exigence renforcée | Argument from Commitment | L'engagement est-il stable, explicite et cohérent tout au long de l'échange ? |
| 1352 | Empoisonnement du puits | Argument from Bias | Le biais allégué est-il pertinent pour la fiabilité sur *cette* affirmation précise ? |
| 1361 | Procès en incohérence | Argument from Commitment | Les engagements attribués impliquent-ils réellement la contradiction alléguée ? |

Several are textbook-classic mappings (Appeal to Consequences→Consequences, False Analogy→Analogy, Equivocation/Amphiboly→Verbal Classification, Poisoning the Well→Bias, Causal Inversion→Cause to Effect, Biased Sample→Example). Full `exception_mechanism` + `justification` per row: [`498-pilot-annotations.csv`](498-pilot-annotations.csv).

## 4. Method & anti-fabrication guarantee

- **Grounding catalog (canonical Walton schemes, Walton/Reed/Macagno 2008):** Argument from Position to Know · Expert Opinion · Witness Testimony · Popular Opinion · Example · Analogy · Correlation to Cause · Cause to Effect · Sign · Consequences · Slippery Slope · Rule · Practical Reasoning · Commitment · Values · Verbal Classification · Bias · Danger · Need for Help · Ignorance · Composition/Division · Precedent · Gradualism · Waste. The model must pick an **exact** name.
- **Model / endpoint:** gpt-5.5 via **`/v1/responses`** with `reasoning:{effort:"low"}` — the cluster's proven method for reliable content output (chat/completions burns the budget in reasoning_tokens and can return empty; effort:low yielded clean JSON at **~68 reasoning tokens/call**). One call per fallacy.
- **Verification (3 layers):**
1. `walton_scheme` ∈ catalog (reject fabricated scheme names),
2. `fallacy_pk` ∈ real 1408-row CSV (reject hallucinated PKs),
3. all 18 confirmed **no prior AIF** in the source CSV (the generative-layer claim holds).
- **Result: 0 PK violations, 0 scheme violations. PASS — all schemes canonical, all PKs real.** Scheme distribution is healthy: **12 distinct schemes** across 18 fallacies (no over-concentration).

Generation script: `tmp/498_pilot_gen.py` (ephemeral). Raw output: `tmp/498_pilot_annotations.json` (available on request).

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

- ❌ Does **not** edit the production `Argumentum Fallacies - Taxonomy.csv` (no `AIF_skos*` cells filled, no rows changed).
- ❌ Does **not** invent AIF corpus node names — it documents the exception structure, sidestepping AIF's ~20-node ceiling per ai-01's directive.
- ❌ Does **not** touch the OWL ontology, cards, mindmaps, or any consumer.
- ❌ Does **not** claim final content — it is the **calibration exemplar** for scaling to the ~1366 no-AIF fallacies.

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

1. **jsboige validates** (a) the exception-structure framing (scheme + CQ, not AIF-node mapping), (b) the 18 scheme assignments in §3, (c) the CQ phrasings.
2. On approval: populate the ~1366 no-AIF Fallacies nodes via a DatasetUpdater gpt-5.5 task, with this pilot as few-shot calibration + the same catalog-grounding + 3-layer verify. The `AIF_skosMappingType` cell would hold the scheme + violated CQ (compact form), `AIF_skosDirectRef` left to a curated AIF-corpus match where one genuinely exists.
3. Propagate to the OWL ontology (AIF annotations on fallacy nodes) → feeds the **EPITA Restitution Epic #1134** (the readable report that *narrates* the exception rather than citing a latin name).

## 7. Consumer benefit

The exception structure is the layer with the highest downstream value: it lets a consumer (EPITA harness, CoursIA pedagogy, the restitution report) explain *why* a fallacy fails — "it is an appeal to consequences that evades CQ: *are the consequences relevant to the proposition's truth?*" — instead of naming it « argumentum ad consequentiam ». ai-01's cross-workspace synchro (EPITA × Argumentum × CoursIA) frames this as the readable debouché of the formal verification EPITA runs.

---

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