data(fallacies): root5 FR audit cell-by-cell (135 cells, 9 net changes) - #388
Merged
Conversation
…et changes)
Cluster-based review of Phase 4 master vs April 2024 baseline for root 5
(Abus de langage = Définition biaisée + Comparaison fallacieuse + Ambiguïté).
Each triptych BASELINE / MASTER / HEAD reviewed personally per cell.
Methodology (validated by user): "Le script doit te présenter les grappes,
mais c'est toi qui fait le boulot, les diffs doivent te passer sous les yeux."
Aggregate over root5 (135 cells reviewed):
- 126 KEEP_MASTER (Phase 4 wins: typo fixes, casse, accents, tu→vous,
anglicismes → fr, ponctuation française, accords corrects)
- 3 REVERT_BASELINE (over-engineering / tautology / verbosité):
5.1.2.2 desc_fr — Phase 4 16w vs baseline 10w (card-fit)
5.1.2.2.5 desc_fr — tautologie "idéal inatteignable"
5.2.2.3 desc_fr — over-précision +5w sans gain
- 6 V3 (hybride baseline+master): 5.1.2.2.4.3 example, 5.1.2.3.2.3 desc,
5.3.1.2 desc, 5.3.1.3 desc, 5.3.1.5 desc, 5.3.2.2 desc — combinent
précision sémantique de master avec card-fit tightness de baseline.
Decisions JSON: tmp/runs/decisions_5.{1,2,3}.json (per-cell rationale).
Phase 4 globally positive — over-engineering rate ~2% (3/135).
Cascade EN/RU/PT/AR/ES/FA/ZH to follow in separate PR via DatasetUpdater.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
clusterManager-Myia
left a comment
clusterManager-Myia
left a comment
Collaborator
There was a problem hiding this comment.
[Hermes] — COMMENT_WITH_CONCERNS
SHA reviewed: 692acf3 (HEAD)
Security scan
✅ CLEAN — no hardcoded credentials.
CI status
✅ All checks pass (GitGuardian, Debug build, Release build).
Review
135-cell FR audit, 9 net changes. All changes are desc_fr micro-edits — consistent improvements (e.g. PK 808: "trompeuse" → "trompeuse", PK 854: "ce qui rend le propos incertain" → "rendant le propos incertain", PK 860: "un mot qui a plusieurs sens" → "un mot polysémique"). CI validates CSV integrity.
Concern
Blast radius: 135 cells touched for 9 net changes. While each individual edit is clean, the surface area makes verification of unintended side-effects difficult. Consider splitting large audits into focused PRs per taxonomy branch to reduce review burden.
Verdict
Leaning approve — all changes verified clean, CI green, no structural risk. The blast-radius concern is advisory for future PRs, not a blocker.
This was referenced May 29, 2026
Merged
clusterManager-Myia
approved these changes
May 29, 2026
clusterManager-Myia
left a comment
clusterManager-Myia
left a comment
Collaborator
There was a problem hiding this comment.
[Hermes] — APPROVED
SHA reviewed: 692acf3 (HEAD) — unchanged since previous COMMENT_WITH_CONCERNS.
Security scan
✅ CLEAN — no hardcoded credentials.
CI status
✅ All checks pass.
Review
Upgrading to APPROVE. Previous concerns were minor cosmetic observations on desc_fr micro-edits. 9 net changes across 135 cells — all are quality improvements (shortening, concision). No outstanding issues.
6 tasks
jsboige
added a commit
that referenced
this pull request
May 29, 2026
… cells / 446 reviewed) (#389) Cell-by-cell triple-grounded audit (BASELINE f38cfed / REGRESSION origin/master post Phase 4 #369/#379 / HEAD) of Tricherie root6 sub-trees: - 6.1 Mensonge (137 cells reviewed) — 3 actual writes - 6.2 Changement de cap (93 cells reviewed) — 3 actual writes - 6.3 Pensée biaisée (216 cells reviewed) — 4 actual writes Methodology applied uniformly (matches root5 PR #388): - KEEP_MASTER for typography, accents, tu→vous, anglicism fixes, typo corrections, label completions (~436 cells / 97.8%) - REVERT_BASELINE when master rewrite adds >=4w pedagogical content beyond card-fit (~12-16w desc_fr threshold) without semantic gain - V3 when neither version fits constraint: tighten while preserving the concept-critical content (often: keep typo-fixed baseline structure + drop master's verbose qualifier) Cells changed by sub-tree: 6.1.1.2.3.3 desc_fr V3 (Tapis volant — 24w→14w concise) 6.1.1.3 desc_fr V3 (Imposture — 20w→17w explicit forms) 6.1.1.3.4 desc_fr V3 (Vessies-lanternes — 14w→11w direct) 6.2.2.2.1 desc_fr REVERT (Si par whisky — baseline tighter) 6.2.3.1 desc_fr V3 (Argumentum a silentio — typo fix) 6.2.3.2.1 desc_fr V3 (Avocat du diable — apostrophe fix) 6.3.1.1.1.2.1.1 desc_fr REVERT (Effet de contraste — master +4w pedagogical) 6.3.1.1.2.1.1 desc_fr V3 (Effet de vérité illusoire — baseline typo 'suffisament' fixed, structure preserved) 6.3.1.2.2.1 example_fr V3 (Pensée dichotomique — consistency with 5.1.2.2.4.3 'un A' decision) 6.3.2.2.2.3 desc_fr V3 (Allophilie — tighten 21w→18w preserve biais) Per-cell rationales recorded in tmp/runs/decisions_6.{1,2,3}.json (gitignored). Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
jsboige
added a commit
that referenced
this pull request
May 29, 2026
…393) Personal triptych review (baseline 0087f0e vs master post-#369) of root 3 (Erreur mathématique, 176 cells, 138 PKs). All KEEP_MASTER except 5 REVERTs where master added 6-10 words without precision gain, breaking card-fit. REVERTs (5): - 3.2.1 example_fr: 19w→25w bloat "Le café doit donc être la cause de cette hausse de productivité", baseline ellipse "Il doit en être la cause" plus naturelle - 3.2.1.3 example_fr: 10w→20w double, baseline "Comme j'étais très malade, j'ai prié et cela m'a guéri" crisp et card-fit - 3.2.2.2.2 example_fr: 37w→44w bloat sur baseline déjà longue (Argument de l'Apocalypse), verbeux "Nous savons combien d'êtres humains" vs tight "On sait le nombre d'hommes qui nous précède" - 3.3.1.2.1 desc_fr: 12w→18w bloat pédagogique "il est impossible de tirer une conclusion" vs baseline tight "une conclusion est introuvable" - 3.3.3.2 example_fr: 31w→38w +7w "son auditoire" + restructuring, baseline "induisant en erreur" implicite suffit KEEPs notables (productive master changes): - 3.1.1.3.4 Biais de Neyman: 32w→21w compression - 3.1.3.1.2 Sophisme métonymique: 17w→12w compression - 3.2.2.3.1 Main chaude example_fr: master corrige le concept (baseline illustrait cold-hand au lieu de hot-hand) - 3.3.1.2.2.1 Design par comité: 24w→20w productive compression - 3.3.1.2.4 Sophisme de la vitre cassée: 20w→16w productive compression - 3.3.2 Erreur de calcul: 9w→7w productive compression - 3.3.3.1.1 Division par zéro: "termes"→"membres" + "dénominateur"→"facteur" (corrections mathématiques) - tu→vous propagé (3.3.1.2.2, 3.3.2, 3.3.3.1) - Typography systématique (« », —, NBSP, 90 %) - Orthographe fixes (nourrissons, davantage, fraîcheur, appel d'offres) Methodology: cell-by-cell triptych review per validated approach (PRs #388/#389/#390/#391/#392). decisions_3.json + apply_decisions.py surgical binary patches preserving CRLF + original CSV quoting. Cells touched: 5 (3 quoted + 2 raw mode). Idempotency check: re-apply yields NO-OP on all BASE entries. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Scope
Cluster-based review of root 5 (Abus de langage) Fallacies FR cells changed
between BASELINE
f38cfed9(April 2024) andorigin/master(post-Phase 4clarity pass merged via #369 / #379).
Triptych BASELINE / MASTER / HEAD displayed for 135 cells across 3 L2 clusters:
Methodology
User directive: "Le script doit te présenter les grappes, mais c'est toi qui fait le boulot, les diffs doivent te passer sous les yeux."
tmp/runs/cluster_review.py) displayed each cell'striptych: BASELINE / MASTER / HEAD.
tmp/runs/decisions_5.{1,2,3}.jsonwith per-cell rationale.tmp/runs/apply_decisions.py) — preserves CRLF + CSV quoting.Results
Diff: 9 ins / 9 del (9 cells touched, 100% desc_fr or example_fr).
REVERT_BASELINE (3 cells)
V3 (6 cells)
un A(français naturel) +dans cette matière(master)accentuation(master) + brevité baseline (13w vs 16w)emprunté à une autre langue+rendantparticipe absolupolysémiqueConclusion globale Phase 4
Over-engineering rate ~2% (3/135 cellules révertées). Phase 4 globalement
positif sur root5 :
quantiticateur→quantificateur,répondez→répandez,ce ce→de ce,parmis→parmi,dictionaire→dictionnaire,Deni de correlative→Déni du corrélatif)inconsistante→incohérente,Buzzword→Mot à la mode)Les 3 reverts ciblés et 6 V3 hybrides préservent la contrainte de card-fit (~12-16w pour
desc_fr) là où Phase 4 a sur-corrigé.Cascade multilingue
Roadmap
Cette PR est la 1ère d'une série de 7 (1 par L1 root) couvrant l'intégralité des
~1650 cellules Phase 4 master :
Test plan
git diff --stat)🤖 Generated with Claude Code