Skip to content

data(rules): EN cascade scale-3 pass gpt-5.5 (24/24 rules) - #376

Merged
jsboige merged 1 commit into
masterfrom
data/rules-cascade-en-scale3
May 28, 2026
Merged

data(rules): EN cascade scale-3 pass gpt-5.5 (24/24 rules)#376
jsboige merged 1 commit into
masterfrom
data/rules-cascade-en-scale3

Conversation

@jsboige

@jsboige jsboige commented May 28, 2026

Copy link
Copy Markdown
Contributor

Summary

First scale-up of cascade EN-only pattern (validated in smoke #372). TakeChunkNb=3 covers all 24 rules in 3 SequentialChunks. ~47s wall, ~$0.80 estimated.

Pattern (unchanged from #371/#372)

  • FR readonly anchor (Text) + EN editable target (Text_en)
  • Function-calling exclusive: UpdateRecord(pk, field, value)
  • Per-chunk priority ordering: 1 lang/pass per rule, EN-only

Stats

  • Chunks: 3 / 3 (full coverage)
  • CSV diff: 76 ins / 77 del, all on Text_en cells
  • Non-EN edits: 0 (verified by per-language char scan)
  • BOM: preserved (Rules convention)

Quality samples (self-review)

Before After Notes
"Facility" "Setup" Literal MT of "Installation"
"Roll of the English Channel" "Round sequence" "Channel" was MT artifact from "manche"
"speed running" "The speed race" Speed run was unidiomatic
"Material" "Components" Game vocab consistency
"pickaxe" / "pickler" "drawer" "piocheur" = card drawer, not mining tool
"sleeve" (manche) "round" Repeated MT bug
"By comming by the neighbor's left of the pickler" "Starting with the player to the drawer's left" Grammar + game term
"smooth talker" "smooth talker" (preserved) Where rule references embromador role explicitly

Validation

  • ✅ All edits sharpen EN to mirror FR clarity (PR data(rules): FR clarity pass gpt-5.5 (19/24) #366)
  • ✅ Game terminology unified (drawer/embromador/round)
  • ✅ No grammatical regressions
  • ✅ 0 scope leakage to other languages
  • ✅ Markdown headings (##, ###) preserved

Test plan

  • Build + run cascade EN-only with TakeChunkNb=3, ChunkSize=8
  • Diff inspection: only Text_en cells modified
  • BOM restored to Rules CSV
  • Visual inspection of generated cards (post-merge, blocked on pipeline regen)

Part of Phase 4 scale-up campaign (smokes #372-#375 validated, now scaling).

🤖 Generated with Claude Code

@jsboige @claude
data(rules): EN cascade scale-3 pass gpt-5.5 (3 chunks, 24/24 rules c…
9198358 
…overed)

Cascade EN-only drift pass with TakeChunkNb=3 (full coverage of 24 rules
across 3 SequentialChunks). Per-chunk function-calling with priority
ordering Text_en (only editable lang per pass). FR/RU/PT/AR/ES/ZH/FA
included as readonly context.

Quality fixes (samples):
- "Facility" (literal MT of "Installation") → "Setup"
- "Roll of the English Channel" → "Round sequence"
- "speed running" → "The speed race"
- "Material" → "Components"
- "pickaxe"/"pickler" → "drawer" (game-term consistency)
- "sleeve" (manche) → "round"
- "By comming by the neighbor's left of the pickler" → "Starting with the player to the drawer's left"
- "smooth talker" preserved where rule references the embromador role

Stats: 76 ins / 77 del, all on Text_en cells; 0 non-EN modifications verified.
BOM preserved (Rules CSV convention).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
@jsboige
jsboige merged commit e57d2aa into master May 28, 2026
3 checks passed
@jsboige jsboige mentioned this pull request May 28, 2026
Closed
jsboige added a commit that referenced this pull request May 28, 2026
@jsboige @claude
data(scenarii): EN cascade full-pass gpt-5.5 (14 chunks, 109 cells / …
1811afc 
…109 records) (#382)

Continuation of cascade EN-only wave (PRs #377 scale-3, #380 scale-6).
TakeChunkNb=-1 full-pass over all 167 records → produces 14 chunks of size 12
(plus 1 final chunk of size 11).

Effective edits: 109 cells across 109 distinct records (65% record coverage),
0 scope leaks outside EN fields. Distribution:
  - issue:          57 cells
  - context:        25 cells
  - suggestion_en:  20 cells
  - title:           7 cells

Quality highlights (verified samples):
  - 1.3.2: "Mr. Chair" → "Mr. President" (Truman = POTUS, factual fix)
  - 1.3.4: matches official UN translation "All human beings are born free
    and equal in dignity and rights" (UDHR Article 1 phrasing)
  - 2.1.4: "Grumpy" → "Snow White with the Dwarfs" (title was wrong; the
    scenario is the whole scene, not the dwarf's name)
  - 2.1.6: "You're the Yeti" → "The smooth talker is a shaggy..." (EN
    impersonnel voice — aligns with jsboige directive)
  - 2.1.3: gender/agent consistency fix on Scheherazade narrative

Cascade pattern outcome:
  - Rules    (scale-3)  : full coverage 24/24    → #376
  - Virtues  (scale-3,6): 16, then 0 (converged)  → #378
  - Fallacies(scale-3,6): 16, then 0 (converged)  → #379
  - Scenarii (scale-3,6,FP): 28, 49, 109 (still productive) → #377, #380, this

Pipeline regen required to produce updated EN PDFs and cascade RU/PT/AR/ES/FA/ZH.

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This was referenced May 29, 2026
Closed
Closed
@jsboige
jsboige deleted the data/rules-cascade-en-scale3 branch June 1, 2026 21:41
This was referenced Jul 1, 2026
Closed
Merged
jsboige added a commit that referenced this pull request Jul 2, 2026
@jsboige Your @claude
fix(i18n): refonte traduction Rules — purge embromador + retraduction…
2079f0c 
… PT/RU + glossaire verrouillé (#633) (#640)

Corrige les 23 findings HIGH du dataset Rules (audit ai-01 2026-07-02) :
vieille MT ~2021 jamais remplacée (Text_pt/Text_ru) + contamination LLM
« embromador » (mot PT) injectée dans FR/EN/ES par les passes #366/#372/#376.

Cards/Rules/Argumentum Rules - Cards.csv (15 records) :
- Text_pt : retraduction intégrale depuis FR (gpt-5.5 Responses API)
- Text_ru : Rules_07 titre (Бинго аргументативной миксологии), Rules_12
  formatage headers, Rules_15 retraduction (header vide + phrase FR incrustée),
  « Ход партии » → « Ход раунда » ×3 (cohérence раунд)
- Text_en : Rules_09 « The Last Smooth Talker » (beau parleur), Rules_10
  « rounds » (manche ≠ sleeves)
- Purge embromador FR→baratineur / EN→smooth talker / ES→embaucador (11 cells,
  PT conservé — embromador = canon PT du rôle baratineur)

Cards/Rules/Argumentum Rules - Cards Print and Play.csv (6 records) :
- Text_pt + Text_ru : retraduction intégrale (PP divergeait silencieusement du
  principal ; couverture #306 reportée : « A escola dos mentirosos »)

Taxonomy (2 cellules MEDIUM) :
- Fallacies PK615 example_ar : russe (copie example_ru) → arabe MSA
- Virtues pk71 description_ru : fuite française → russe

Prévention récidive — glossaire verrouillé dans les prompts DatasetUpdater :
- PromptRulesFrClarityUser.txt : « embromador » listé à tort comme terme FR →
  corrigé en canonique « baratineur » + directive anti-contamination étrangère
- PromptRulesCascadeDriftUser.txt : ligne « terminologie incohérente
  (embromador, baratineur) » induisait l'harmonisation → remplacée par le
  glossaire canonique multi-langue (piocheur/baratineur/manche/pli/atout/saynète)

Scanner (scan_translations.py) : 22 findings → 7 (les 7 résiduels = faux
positifs documentés : noms latins canoniques Credo quia absurdum, emprunts
Gish Gallop/Creepypasta/Whataboutism, mnémoniques Celarent/Camestres, nom
propre Johnny Johnny). 0 HIGH Rules, 0 MEDIUM taxonomy résiduel.

NOTE : GDrive (source autoritative, item #6 du scope) NON mis à jour — upload
gated jsboige. Sans sync GDrive, le prochain pull baseline réimportera le
garbage (cause racine du clobber #211 par 7fae40c).

Co-authored-by: Your <your.email@example.com>
Co-authored-by: Claude-Code <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@jsboige