data(fallacies): FR clarity + cascade drift infrastructure (gpt-5.5) by jsboige · Pull Request #369 · ArgumentumGames/Argumentum

jsboige · 2026-05-28T19:02:20Z

Summary

Phase 4 Fallacies FR clarity review using PKHierarchicalChar hierarchical chunks with gpt-5.5
955/1408 PKs modified (67.8%), 1651/4224 cells changed (39.1%)
desc_fr: 700/1408 (49.7%) — simplified complex phrasing, "Vous" formal register
example_fr: 744/1408 (52.8%) — more vivid, concrete examples
text_fr: 207/1408 (14.7%) — minor name refinements only
MaxChildren raised 8→20 for less chunk fragmentation
Cascade drift remap infrastructure for Fallacies/Scenarii/Virtues/Rules + Scenarii FR clarity prompt + Logger.cs hardening

Quality check

Sampled paths 1.1.1, 3.2.1 — "Vous" register preserved, technical terms intact, descriptions clearer
Hierarchical chunking (PKHierarchyLevel=2) provides family context for consistent edits

Config changes

DatasetUpdaterRootConfig.cs: MaxChildren 8→20 for Fallacies tasks, added 4 cascade task configs
New prompts (10 files): PromptFallaciesFrClarityUser/Assistant.txt, PromptFallacies/Virtues/Scenarii/RulesCascadeDriftUser/Assistant.txt, PromptScenariiFrClarityUser/Assistant.txt

Cascade drift design — progressive convergence pattern

The cascade prompts use explicit progressive convergence rules rather than "skip if accurate" framing (the latter caused 3/4 cascades to converge brusquely to 0 changes in early experiments — diagnosed as prompt bug, not real convergence):

Function-calling exclusive (UpdateRecord(path, field, value))
Max 1 cell per node per pass, with field priorisation:
- Fallacies: example_* > desc_* > text_*
- Virtues: remark_* > description_* > title_*
- Scenarii: enjeu/issue > contexte/context > suggestion > titre/title
- Rules: 1 language per rule per pass (text only)
Tangible drift criteria: semantic, style, embodied fidelity, cross-lang coherence

Smoke test validation (Fallacies cascade, chunk size 12, TakeChunkNb=1)

Pass	Source	UpdateRecord calls	Records touched	Cells/record
1	master	10	10 distinct	1
2	pass 1 output	9	9 distinct (rotated to different languages/fields)	1

Gradient confirms progressive convergence, not brusque-to-zero. Estimated full Fallacies cascade run: ~$58 (1408 records / 12 = 117 chunks × ~$0.50).

Logger.cs hardening

LogException now logs raw exception type+message FIRST so the underlying error is always visible, even when Spectre.Console's ExceptionFormatter itself throws (known IndexOutOfRangeException in stack-frame parsing that previously masked the real exception during cascade runs). Spectre.WriteException now wrapped in try/catch with stack fallback.

Next step

Coordinator decision on running the full cascade (~$58 est) once data(virtues): FR clarity pass gpt-5.5 (223/223, 516 cells) #367 (Virtues) and data(scenarii): FR clarity pass gpt-5.5 — 162/167 records #368 (Scenarii) FR clarity merge
Cascade tasks all left Enabled = false — deliberate activation only

🤖 Generated with Claude Code

Self-review fixes (commit `2e5e110`)

During final self-review I caught 2 bugs in the cascade infra commit:

Fallacies FR clarity had Enabled=true (smoke test artifact) — would have re-run on next dotnet run, wasting budget. Set to Enabled=false.
Scenarii cascade drift remap was duplicated (lines 1805 + 1967). The duplicate used PKHierarchicalChar+ChunkSize=4 (copy-pasted from Fallacies config), which is wrong for Scenarii (flat PKs 1.1.1 require SequentialChunks ChunkSize=12 per MEMORY.md). Kept the first (correct) entry only.

Build: 0 errors. All cascade/clarity tasks now Enabled=false (deliberate activation only).

…ecord calls) SequentialChunks ChunkSize=12, 14 chunks, gpt-5.5 function calling. FR fields reviewed: titre, contexte, enjeu, suggestion. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Phase 4 Fallacies FR clarity review using PKHierarchicalChar chunks. - desc_fr: 700/1408 changed (49.7%) — simplified complex phrasing, "Vous" register - example_fr: 744/1408 changed (52.8%) — more vivid, concrete examples - text_fr: 207/1408 changed (14.7%) — minor name refinements - MaxChildren raised 8→20 for less fragmentation - Added Fallacies cascade drift remap task config + prompts Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

clusterManager-Myia

[Hermes] — APPROVED

Revue de données + config converter : corrections typographiques et mises à jour LLM.

Points vérifiés :

Sécurité : scan clean. Les occurrences de "secrets" dans les exemples fallacies sont du contenu pédagogique, pas des credentials. GitGuardian ✓.
CI : build Debug ✓, build Release ✓.
Taxonomy CSV : ~900 lignes modifiées. Changements principaux = corrections typo FR (espaces insécables, virgules, rewording). Pas de changements structurels. Header : colonne "latin" renommée en "latin1".
Scenarii CSV : corrections typographiques mineures FR.
C# DatasetUpdaterRootConfig : 3 nouvelles configs de batch LLM (2 Scenarii disabled + 1 Fallacies enabled). DTOs simples, pas de logique métier.
Ressources prompts : fichiers texte pour batchs LLM, pas de secrets.

Aucun risque fonctionnel. CI verte.

…rgence strategy User-crafted prompt with: - Max 1 cell/node/pass (example > desc > text priority) - Tangible drift criteria (semantic, style, fidelity, cross-lang consistency) - Expected 0-30% update rate per chunk (not massive rewrite) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…ning Completes the cascade-drift infrastructure referenced by the Fallacies/ Scenarii/Virtues/Rules cascade task configs. The Fallacies cascade prompts already shipped in 5463c91; this commit adds the missing prompt pairs and the Scenarii FR clarity prompt referenced by line 1778 of the config. Prompts use the progressive convergence pattern validated via smoke test: - Function-calling exclusive (UpdateRecord(path, field, value)) - Max 1 cell per node per pass, with field priorisation - Fallacies: example_* > desc_* > text_* - Virtues: remark_* > description_* > title_* - Scenarii: enjeu/issue > contexte/context > suggestion > titre/title - Rules: 1 language per rule per pass (text only) - Tangible drift criteria: semantic, style, embodied fidelity, cross-lang coherence (vs "skip if accurate" which converges brusquely to zero) Smoke test (Fallacies cascade, chunk size 12, TakeChunkNb=1): - Pass 1: 10 UpdateRecord calls / 10 distinct records, 1 cell each - Pass 2 (with pass 1 output as source): 9 UpdateRecord / 9 records, rotated to DIFFERENT languages/fields than pass 1 - Gradient confirms progressive convergence, not brusque-to-zero Logger.cs hardening: - LogException now logs raw type+message FIRST so the underlying error is always visible, even when Spectre.Console's ExceptionFormatter itself throws (known IndexOutOfRangeException in stack-frame parsing that previously masked the real exception) - Spectre.WriteException now wrapped in try/catch with stack fallback Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…narii cascade Self-review of PR #369 found 2 bugs in the cascade infra commit: 1. "Fallacies FR clarity review gpt-5.5" had Enabled=true (smoke test artifact). Disabled to prevent accidental re-run that would waste budget. 2. "Scenarii cascade drift remap gpt-5.5" appeared TWICE — the duplicate (lines 1967-2024) was a copy-paste from the Fallacies cascade config that used PKHierarchicalChar+ChunkSize=4, which is wrong for Scenarii (flat PKs 1.1.1 require SequentialChunks ChunkSize=12 per MEMORY.md). Kept the first (correct SequentialChunks) entry only. Build: 0 errors, all tasks now Enabled=false (deliberate activation only). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

PR #369 delivered the cascade prompts for Rules and Virtues (PromptRulesCascadeDrift{User,Assistant}.txt, PromptVirtuesCascadeDrift {User,Assistant}.txt) but only the Scenarii + Fallacies cascade task configs reached master. This closes that gap so all 4 datasets have matching infrastructure. Both tasks default Enabled=false. Settings mirror the existing cascade tasks: - SequentialChunks ChunkSize=4 (small datasets vs Scenarii ChunkSize=12) - gpt-5.5, function-calling, MaxOutputTokens 4096/8192 - TakeChunkNb=1 ready for smoke-test convergence checks - FieldsToUpdate covers 7 languages (EN/RU/PT/ES/AR/FA/ZH), excludes the FR source (locked by FR clarity passes) Build: 0 errors, 18 known warnings. Tests: 120/125 pass (5 skip GDrive/GUI). Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…379) Cascade EN-only realignment after FR clarity pass (PR #369). TakeChunkNb=3, ChunkSize=4, PKHierarchyLevel=2, scope: text_en/desc_en/example_en. Distribution: desc_en=9, example_en=6, text_en=1 (priority example>desc>text honored). Notable correction: path 1.1 text_en "Sloppy argument" -> "Hasty generalization" (actual concept name; previous EN was wrong terminology). Other quality samples (FR -> EN-before -> EN-after): - 1.1.1.2 desc: generic "Arguing that a conclusion is right because something must be done" -> "Defending one's point of view by insisting on the need to act, even without evidence that the proposed action is the right one" - 1.1.1.3 desc: legalistic "Demanding the acceptance of a statement as true..." -> "Asking for a claim to be accepted without providing any reason or logical explanation" - 1.1.3 desc: "Making unjustified explanatory hypotheses in order to prove that a counter-argument is not actually disproving one's point" -> "Improvising unfounded justifications in the face of valid objections in order to escape criticism" - 1.1.1.2 example: "the mice keep multiplying in my basement; I think we just need to flood it" -> "We can't let the mice keep multiplying in the cellar. So we have to flood it." (direct mirror of FR) 1408/1408 records preserved. Zero scope leaks. EN impersonal style preserved. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…6.1.1.2) Cell-by-cell triptych review of 6 desc_fr cells modified by the Phase 4 FR clarity pass (PRs #366-#369). Verdict: 3 baseline restores, 3 Phase 4 keeps (no-op), 0 AI proposals accepted. REVERT to baseline f38cfed (Phase 4 collapsed scope toward one child): - PK 7.3 Ad hominem : restore "Vous visez votre adversaire lui-même..." (covers 3 children, not just Attaque personnelle) - PK 3.1.3 Transfert illicite : restore "...ou l'inverse" (preserves composition⇄division duality) - PK 6.1.1.2 Factoïde : restore "faits largement acceptés mais inexacts" (covers Désinformation/Rumeur/ Mythe/Mème, not just Rumeur) KEEP Phase 4 (genuine improvements): - PK 4.2.3 Erreur de modalité : accessible vocabulary - PK 4.3.1.1 Prémisses non pertinentes : fixes typo prémices→prémisses - PK 3.3 Conclusion math invalide : singulier + concis Methodology: tmp/runs/triptych.py displays baseline / Phase 4 master / HEAD side-by-side per PK with word counts and direct-children labels for scope verification. Validated by jsboige ("OK pour la répartition, tu peux rester sur cette ligne pour le reste") — to be extended to full Phase 4 Fallacies audit (~1650 cells / ~955 PKs) via cluster review. Surgical binary patch via tmp/runs/revert_cells.py: 3 insertions / 3 deletions. CRLF + quoting preserved. No collateral. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…6.1.1.2) (#387) Cell-by-cell triptych review of 6 desc_fr cells modified by the Phase 4 FR clarity pass (PRs #366-#369). Verdict: 3 baseline restores, 3 Phase 4 keeps (no-op), 0 AI proposals accepted. REVERT to baseline f38cfed (Phase 4 collapsed scope toward one child): - PK 7.3 Ad hominem : restore "Vous visez votre adversaire lui-même..." (covers 3 children, not just Attaque personnelle) - PK 3.1.3 Transfert illicite : restore "...ou l'inverse" (preserves composition⇄division duality) - PK 6.1.1.2 Factoïde : restore "faits largement acceptés mais inexacts" (covers Désinformation/Rumeur/ Mythe/Mème, not just Rumeur) KEEP Phase 4 (genuine improvements): - PK 4.2.3 Erreur de modalité : accessible vocabulary - PK 4.3.1.1 Prémisses non pertinentes : fixes typo prémices→prémisses - PK 3.3 Conclusion math invalide : singulier + concis Methodology: tmp/runs/triptych.py displays baseline / Phase 4 master / HEAD side-by-side per PK with word counts and direct-children labels for scope verification. Validated by jsboige ("OK pour la répartition, tu peux rester sur cette ligne pour le reste") — to be extended to full Phase 4 Fallacies audit (~1650 cells / ~955 PKs) via cluster review. Surgical binary patch via tmp/runs/revert_cells.py: 3 insertions / 3 deletions. CRLF + quoting preserved. No collateral. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

… 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>

… reviewed) (#390) Root 2 (Influence — 418 cells) cell-by-cell audit of Phase 4 #369 (FR clarity gpt-5.5). Re-read every triptych baseline (April 2024 f38cfed) vs master vs HEAD, decided per cell. Verdicts breakdown: - KEEP_MASTER: 399 cells (typo/anglicism/typography/tu→vous fixes) - REVERT_BASELINE: 18 cells (master expansion exceeds card-fit ~12-16w constraint without adding precision) - V3: 1 cell (2.3.2.3.1 desc_fr — baseline 8w too laconic, master 19w verbose, V3 14w preserves precision) REVERT highlights: - 2.3.1.1 Cadrage desc_fr (9→17w near-double, baseline tight enough) - 2.3.2.1 Charme personnel desc_fr (13→25w near-double) - 2.3.3.2 Chronémique desc_fr (15→31w, master over-explains a basic concept) - 2.3.1.1.1 Objectif caché example_fr (master reveals the hidden strategy, contradicting the fallacy itself) - 2.3.3.3 Oculésique example_fr (concept replacement, baseline more illustrative) - 11 others — bloat without precision gain Method: tmp/runs/cluster_review.py shows triptych, decisions_2.json records verdicts, apply_decisions.py surgical binary patch preserves CRLF + CSV quoting. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Audit cellule-par-cellule des 212 cellules régressives sous racine 1 (Insuffisance) générées par le passage Phase 4 FR clarity (#369). Décisions: 209 KEEP_MASTER (gains productifs préservés) + 3 REVERT_BASELINE (bloat pédagogique sans gain de précision). REVERT_BASELINE (3 cellules): - 1.1.1.1.5 desc_fr (Appel à la complexité): 20w→23w master ajoutait "ce qui vous permet de" sans précision - 1.3.1.1.4 desc_fr (Angélisme): 15w→21w bloat "contraintes" + "en jeu" borderline card-fit - 1.3.1.3.2 desc_fr (Anatopisme): 19w→28w bloat "supposez ou jugez que...alors qu'ils ne l'étaient pas" KEEP_MASTER (209 cellules) — patterns productifs préservés: - tu→vous normalization - French typography (« », —, ', espaces insécables) - Anglicism fixes (modèles → motifs) - Historical accuracy (Jules César "dictateur à vie" ≠ "empereur") - Concept improvements (1.3.2 "plutôt qu'en examinant l'argument" = essence ad-hominem) - Précision Morgan canon ("intentions, raisonnements ou caractéristiques") - Lamppost effect compression 35w→20w (master = vrai gain) - Typo fixes (mlieu→milieu, lis→lit, chomage→chômage) - Grammar (Quoiqu'ils→Quoi qu'ils, sentaient→pouvaient) Test plan: visual inspection des cartes Tarot FR sur les 3 PKs reverted; verifier card-fit ≤16w sur desc_fr. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…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>

…22 cells) Cell-by-cell audit of all 222 cells changed in root7 (Obstruction) since PR #369/#379 (Phase 4 FR clarity + cascade drift). Verdicts: - 215 KEEP_MASTER — typography (« », apostrophes courbes, em-dash), tu→vous, anglicism fixes (Mauvais Support→Fondement erroné, Pathos gambit→Argument du pathos, leaders→dirigeants), typo fixes (rever→revers, géner→gêner, chechez→cherchez, anneries→âneries, parmis→parmi, nonchalence→nonchalance, idiosyncrasique→idiosyncratique, Bulverisme→Bulvérisme, Evasion→Évasion), grammar (accords, articles), case fixes, substantivation des titres. - 7 REVERT_BASELINE — master rewrites add 3-9 words without precision gain, violating ~12-16w card-fit constraint: * 7.1.1 desc_fr 12w→20w (redondant avec titre) * 7.1.2 desc_fr 12w→21w (pseudo-explication déjà clair) * 7.1.3.1 desc_fr 13w→18w (over-explanation) * 7.2.1.1.2 example_fr 24w→30w (narrative bloat) * 7.2.1.3.1.2 desc_fr 12w→19w (redondance avec titre) * 7.2.2.1 desc_fr 15w→18w (embarrasser→circumlocution) * 7.3.1 example_fr 20w→24w (perte du punch 'ma morale le tolère') Notable preserved master corrections: - 7.3.2.2 example_fr — fixes logical inconsistency: baseline had "Fred a tort d'affirmer que les prisonniers NE peuvent PAS se réinsérer ...car c'est un ancien prisonnier repenti" (logically incoherent — a repenti would advocate FOR reintegration). Master removes "ne...pas". - 7.2.3.3.1 DARVO desc_fr — master expands to match the actual concept (Deny, Attack, Reverse Victim and Offender) instead of vague baseline. - 7.2.3.4.1 example_fr — fixes corrupted "n' es" spacing. - 7.3.2.1.5 example_fr — anglicism fix "La prochaine chose qu'on sait" (calque) → "À ce rythme" (idiomatic French). Cascade EN/RU/PT/AR/ES/FA/ZH to be triggered separately after all 7 roots complete (scale-3 DatasetUpdater). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…22 cells) (#391) Cell-by-cell audit of all 222 cells changed in root7 (Obstruction) since PR #369/#379 (Phase 4 FR clarity + cascade drift). Verdicts: - 215 KEEP_MASTER — typography (« », apostrophes courbes, em-dash), tu→vous, anglicism fixes (Mauvais Support→Fondement erroné, Pathos gambit→Argument du pathos, leaders→dirigeants), typo fixes (rever→revers, géner→gêner, chechez→cherchez, anneries→âneries, parmis→parmi, nonchalence→nonchalance, idiosyncrasique→idiosyncratique, Bulverisme→Bulvérisme, Evasion→Évasion), grammar (accords, articles), case fixes, substantivation des titres. - 7 REVERT_BASELINE — master rewrites add 3-9 words without precision gain, violating ~12-16w card-fit constraint: * 7.1.1 desc_fr 12w→20w (redondant avec titre) * 7.1.2 desc_fr 12w→21w (pseudo-explication déjà clair) * 7.1.3.1 desc_fr 13w→18w (over-explanation) * 7.2.1.1.2 example_fr 24w→30w (narrative bloat) * 7.2.1.3.1.2 desc_fr 12w→19w (redondance avec titre) * 7.2.2.1 desc_fr 15w→18w (embarrasser→circumlocution) * 7.3.1 example_fr 20w→24w (perte du punch 'ma morale le tolère') Notable preserved master corrections: - 7.3.2.2 example_fr — fixes logical inconsistency: baseline had "Fred a tort d'affirmer que les prisonniers NE peuvent PAS se réinsérer ...car c'est un ancien prisonnier repenti" (logically incoherent — a repenti would advocate FOR reintegration). Master removes "ne...pas". - 7.2.3.3.1 DARVO desc_fr — master expands to match the actual concept (Deny, Attack, Reverse Victim and Offender) instead of vague baseline. - 7.2.3.4.1 example_fr — fixes corrupted "n' es" spacing. - 7.3.2.1.5 example_fr — anglicism fix "La prochaine chose qu'on sait" (calque) → "À ce rythme" (idiomatic French). Cascade EN/RU/PT/AR/ES/FA/ZH to be triggered separately after all 7 roots complete (scale-3 DatasetUpdater). Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

jsboige and others added 2 commits May 28, 2026 19:34

data(scenarii): FR clarity pass gpt-5.5 (162/167 records, 437 UpdateR…

5299150

…ecord calls) SequentialChunks ChunkSize=12, 14 chunks, gpt-5.5 function calling. FR fields reviewed: titre, contexte, enjeu, suggestion. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

clusterManager-Myia reviewed May 28, 2026

View reviewed changes

jsboige and others added 2 commits May 28, 2026 21:44

jsboige changed the title ~~data(fallacies): FR clarity pass gpt-5.5 (955/1408 PKs, 39.1% cells)~~ data(fallacies): FR clarity + cascade drift infrastructure (gpt-5.5) May 28, 2026

jsboige merged commit 9d45b4f into master May 28, 2026
3 checks passed

jsboige deleted the data/fallacies-fr-clarity branch May 28, 2026 20:06

jsboige mentioned this pull request May 28, 2026

feat(config): add Rules + Virtues cascade drift task configs #370

Merged

4 tasks

jsboige mentioned this pull request May 28, 2026

data(fallacies): EN cascade-drift scale-3 gpt-5.5 (16 cells, 14 PKs) #379

Merged

5 tasks

jsboige mentioned this pull request May 29, 2026

data(fallacies): re-align desc EN/RU/PT onto reworked desc_fr (40 drifted PKs) #355

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

data(fallacies): FR clarity + cascade drift infrastructure (gpt-5.5) - #369

data(fallacies): FR clarity + cascade drift infrastructure (gpt-5.5)#369
jsboige merged 5 commits into
masterfrom
data/fallacies-fr-clarity

jsboige commented May 28, 2026 •

edited

Loading

Uh oh!

clusterManager-Myia left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

jsboige commented May 28, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Quality check

Config changes

Cascade drift design — progressive convergence pattern

Smoke test validation (Fallacies cascade, chunk size 12, TakeChunkNb=1)

Logger.cs hardening

Next step

🤖 Generated with Claude Code

Self-review fixes (commit 2e5e110)

Uh oh!

clusterManager-Myia left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

jsboige commented May 28, 2026 •

edited

Loading

Self-review fixes (commit `2e5e110`)