data(fallacies-fr): audit root6 — 6.1/6.2/6.3 selective revert+V3 (10 cells / 446 reviewed)

[jsboige]

Summary

Cell-by-cell audit of Phase 4 FR clarity changes (PRs #369/#379) on the Tricherie root6 sub-tree of Fallacies/Argumentum Fallacies - Taxonomy.csv. Triple-grounded reading of BASELINE f38cfed9 (April 2024) vs REGRESSION origin/master (post-Phase 4) vs my candidate edits.

446 cells reviewed → 10 selective writes (97.8% master KEEP rate).

Continuation of methodology validated in root5 PR #388: master changes are overwhelmingly improvements (typography, accents, tu→vous, anglicism fixes, typo corrections, label completions) but a small minority over-engineer compact baseline cells beyond the playing-card width constraint (~12-16w for desc_fr).

Sub-trees audited

Sub-tree	Label	Cells reviewed	Writes
6.1	Mensonge (Lying)	137	3
6.2	Changement de cap (Shifting goalposts)	93	3
6.3	Pensée biaisée (Biased reasoning)	216	4
total	—	446	10

Cells changed

PK	Field	Verdict	Rationale
6.1.1.2.3.3	desc_fr	V3	Tapis volant — 24w→14w (master pedagogical, V3 concise)
6.1.1.3	desc_fr	V3	Imposture — 20w→17w (explicit forms list)
6.1.1.3.4	desc_fr	V3	Vessies-lanternes — 14w→11w (direct active)
6.2.2.2.1	desc_fr	REVERT	Si par whisky — baseline tighter than master rewrite
6.2.3.1	desc_fr	V3	Argumentum a silentio — typo fix preserved baseline form
6.2.3.2.1	desc_fr	V3	Avocat du diable — apostrophe fix only
6.3.1.1.1.2.1.1	desc_fr	REVERT	Effet de contraste — master +4w pedagogical, baseline tight
6.3.1.1.2.1.1	desc_fr	V3	Effet de vérité illusoire — baseline typo suffisament fixed, structure preserved (master 24w too verbose)
6.3.1.2.2.1	example_fr	V3	Pensée dichotomique — consistency with 5.1.2.2.4.3 'un A' decision in PR #388
6.3.2.2.2.3	desc_fr	V3	Allophilie — tighten 21w→18w while preserving the biais qualifier

Methodology

For each of the 446 triptychs (BASELINE/REGRESSION/HEAD):

1. Word-count delta (master vs baseline)
2. Concept coverage check (does the change preserve the discriminating sophism signal?)
3. Card-fit check (~12-16w for desc_fr)
4. Decide: KEEP_MASTER (default — improvement preserved), REVERT_BASELINE (master regression), V3:<text> (third reading)

Auto-generated decision stubs at tmp/runs/decisions_6.{1,2,3}.json (gitignored), reviewed and filled per-cell, then surgically patched into the CSV via tmp/runs/apply_decisions.py to preserve CRLF + original quoting.

Test plan

• CSV byte-diff: 10 insertions / 10 deletions (1 line each, surgical)
• No PK collision: all 10 PKs distinct and in 6.x sub-tree
• No header drift: schema columns intact (text_fr/desc_fr/example_fr unchanged structurally)
• Multilang cascade (EN/RU/PT/AR/ES/FA/ZH) for the 10 FR cells → separate PR via DatasetUpdater scale-3 (same pattern as PR data(scenarii): EN cascade-drift scale-6 gpt-5.5 (49 cells, ~30 PKs) #380)
• Visual inspection — pipeline regen optional pending root2/7/1/3/4 completion

Related

• Continues data(fallacies): root5 FR audit cell-by-cell (135 cells, 9 net changes) #388 (root5 audit, 65 cells / 1 actual write)
• Audit driven by directive to verify Phase 4 changes (PRs data(fallacies): FR clarity + cascade drift infrastructure (gpt-5.5) #369/data(fallacies): EN cascade-drift scale-3 gpt-5.5 (16 cells, 14 PKs) #379) against April 2024 baseline for clarity/simplicity/card-fit
• Next: root2 (Influence 418c) → root7 (Obstruction 222c) → root1 (Insuffisance 212c) → root3 (Erreur math 176c) → root4 (Erreur raisonnement 150c)

🤖 Generated with Claude Code

[clusterManager-Myia]

[Hermes] — APPROVED

SHA reviewed: fd4ac94 (HEAD)

Security scan

✅ CLEAN — no hardcoded credentials.

Review

Selective desc_fr audit (root6: sections 6.1/6.2/6.3) — 5 targeted refinements in 1 CSV file:

• Row 928 (Tapis volant): "récit douteux concernant une religion [...] afin d'en donner une image hostile ou moqueuse" → "récit stéréotypé hostile" — wording improvement, avoids "moqueuse" which was too mild.
• Row 930 (Imposture): long enumeration → concise "(expertise, identité, expérience)" — cleaner definition.
• Row 860 (Glissement lexical polysémique): added "polysémique" before "ou dont le sens a évolué" — more precise.

All changes are content refinements, no structural/CSV schema changes. Low risk. APPROVE.