Skip to content

data(fallacies): simplify 27 desc_fr for younger audience - #351

Merged
jsboige merged 2 commits into
masterfrom
data/fr-simplification-pass1
May 24, 2026
Merged

data(fallacies): simplify 27 desc_fr for younger audience#351
jsboige merged 2 commits into
masterfrom
data/fr-simplification-pass1

Conversation

@jsboige

@jsboige jsboige commented May 24, 2026

Copy link
Copy Markdown
Contributor

Summary

  • DatasetUpdater pass 1 simplifying French descriptions (desc_fr) for a younger audience
  • 27 descriptions rewritten using clearer, more concrete language via gpt-5.5
  • All changes maintain the "Vous" (second person) form consistently

Examples of changes

PK Name Before After
281 Antilogie "Afin de perturber votre auditoire, vous avancez des arguments qui se contredisent eux-mêmes." "Vous tenez des propos qui se contredisent eux-mêmes pour créer un effet rhétorique."
916 Rumeur "Votre argument s'appuie sur une histoire erronnée qui se propage." "Vous appuyez votre argument sur une information non vérifiée qui circule de personne en personne."
729 Négation de l'antécédent "Vous rejetez une conclusion simplement parce qu'une cause possible ne s'est pas réalisée, confondant ainsi cause et condition." "Vous concluez qu'une conséquence est fausse simplement parce qu'une de ses conditions suffisantes n'est pas remplie."

Verification

  • Only desc_fr field modified (verified programmatically)
  • 1408 rows × 102 columns preserved
  • Column count validated on every row (0 mismatches)
  • Config changes reverted (SkipConfigFile = true, DatasetUpdater disabled)

Test plan

  • CSV structural integrity (102 columns, 1408 rows)
  • Visual review of the 27 simplified descriptions
  • Confirm no regression in pipeline output

🤖 Generated with Claude Code

DatasetUpdater pass using gpt-5.4 to simplify French descriptions.
All 27 changes use clearer, more concrete language while maintaining
the "Vous" second-person form. Examples:
- PK 281: "Afin de perturber votre auditoire..." -> "Vous tenez des
  propos qui se contredisent eux-memes pour creer un effet rhetorique."
- PK 916: "Votre argument s'appuie sur une histoire erronnee..." ->
  "Vous appuyez votre argument sur une information non verifiee..."

Pass 2 ran (952 API calls) but produced 0 additional desc_fr changes,
confirming pass 1 simplifications were sufficient.

No other fields modified. 1408 rows, 102 columns preserved.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@jsboige
jsboige force-pushed the data/fr-simplification-pass1 branch from 093dbbf to d21199a Compare May 24, 2026 02:42

@clusterManager-Myia clusterManager-Myia left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

27 desc_fr simplifications verified — clearer phrasing, typo fix (danns), less aggressive framing (perturber/troubler removed). Only desc_fr column changed, all other columns preserved. Clean data improvement for younger audience.

…268, 750)

Hand reworks on top of the gpt-5.5 pass 1, to preserve sense distinctions
with parent/sibling/child nodes (validated by jsboige):

- PK 268 (Adjonction): "supprimez un mot attendu..." collided with sibling
  Ellipse (PK 273, omission). Reworded to "Vous ajoutez et reliez plusieurs
  mots ou idees pour creer un effet de style." (adjonction = adding/joining).
- PK 750 (Erreur de modalite, parent node): added "ou obligatoire" so the
  parent definition encompasses its deontic child (PK 755).

Gate: only desc_fr changed, 27 cells total vs origin/master, 1409 rows x
102 cols, 0 malformed rows.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
@myia-ai-01

Copy link
Copy Markdown

Added 1 commit (42915d0) with 2 manual taxonomy-distinction reworks on top of the gpt-5.5 pass, validated by @jsboige:

  • PK 268 (Adjonction) — the LLM rewrite ("supprimez un mot attendu…") collided in sense with sibling Ellipse (PK 273, omission). Reworded to "Vous ajoutez et reliez plusieurs mots ou idées pour créer un effet de style." (adjonction = adding/joining, distinct from omission).
  • PK 750 (Erreur de modalité, parent node) — added "ou obligatoire" so the parent definition encompasses its deontic child (PK 755).

Gate re-run vs origin/master: only desc_fr changed, 27 cells total, 1409 rows × 102 cols, 0 malformed rows.

@jsboige
jsboige merged commit 7206f2f into master May 24, 2026
3 checks passed
@jsboige
jsboige deleted the data/fr-simplification-pass1 branch May 24, 2026 08:17
jsboige added a commit that referenced this pull request Jun 10, 2026
…uping (#449)

The Mémo Back card collapsed from the spec 7×3×3 taxonomy grid (7 families ×
3 sub-families × 3 sub-sub-families) to an irregular shape. Root cause: the
template grouped the grid via name-coincidence markers
({{#ifCond Famille "==" text_fr}} …) that fired only on rows whose text_fr
happened to equal their own level label. FR rename campaigns (#351 ×27,
#308 ×13) changed text_fr → the coincidences vanished → families collapsed
(Tricherie sub-families = 0, Erreur math/raisonnement = 1, …). The taxonomy
DATA is correct; only the template's grouping was broken.

Fix (lane ai-01 = TEMPLATE, data is fine):
- CardPen main.js: add 3 control-break helpers reading the language-invariant
  FR taxonomy columns (Sous-Famille / Soussousfamille), present in every row
  regardless of render language. They carry NO column token, so the localizer
  (CardSetLocalization.FormatField appends "}}") cannot corrupt the helper call:
    ifFamilyHeader   = Sous-Famille=='' && Soussousfamille==''
    ifSubfamilyHeader= Sous-Famille!=''  && Soussousfamille==''
    ifLeafFirst      = Soussousfamille!='' on first occurrence (control-break)
- Argumentum_Memo_Back_fr.json: swap the 3 name-coincidence conditions for the
  new helpers (display tokens {{Famille}}/{{Sous-Famille}}/{{Soussousfamille}}
  unchanged → localization preserved); fix colorPalette to iterate the rowset
  and emit one swatch per family-root (was fixed indices 0/25/50/…/150, which
  duplicated Influence and omitted Obstruction).

Single shared template → fix propagates to all 8 languages.

Validated three ways on the real injected data (taxonomy WHERE carte∈{1,2},
176 rows): (1) Node + vendored Handlebars → 7 families / 21 sub-families /
63 leaves; (2) OLD-template render proving the regression (10/53, Tricherie=0);
(3) LIVE CardPen engine (Playwright) → 7×3×3 DOM + 7 distinct palette colors
(Obstruction red present, no Influence duplicate) + viewed screenshot.

Refs #443

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
jsboige added a commit that referenced this pull request Jun 10, 2026
…-up) (#450)

* fix(memo): #443 Mémo Back grid — language-invariant control-break grouping

The Mémo Back card collapsed from the spec 7×3×3 taxonomy grid (7 families ×
3 sub-families × 3 sub-sub-families) to an irregular shape. Root cause: the
template grouped the grid via name-coincidence markers
({{#ifCond Famille "==" text_fr}} …) that fired only on rows whose text_fr
happened to equal their own level label. FR rename campaigns (#351 ×27,
#308 ×13) changed text_fr → the coincidences vanished → families collapsed
(Tricherie sub-families = 0, Erreur math/raisonnement = 1, …). The taxonomy
DATA is correct; only the template's grouping was broken.

Fix (lane ai-01 = TEMPLATE, data is fine):
- CardPen main.js: add 3 control-break helpers reading the language-invariant
  FR taxonomy columns (Sous-Famille / Soussousfamille), present in every row
  regardless of render language. They carry NO column token, so the localizer
  (CardSetLocalization.FormatField appends "}}") cannot corrupt the helper call:
    ifFamilyHeader   = Sous-Famille=='' && Soussousfamille==''
    ifSubfamilyHeader= Sous-Famille!=''  && Soussousfamille==''
    ifLeafFirst      = Soussousfamille!='' on first occurrence (control-break)
- Argumentum_Memo_Back_fr.json: swap the 3 name-coincidence conditions for the
  new helpers (display tokens {{Famille}}/{{Sous-Famille}}/{{Soussousfamille}}
  unchanged → localization preserved); fix colorPalette to iterate the rowset
  and emit one swatch per family-root (was fixed indices 0/25/50/…/150, which
  duplicated Influence and omitted Obstruction).

Single shared template → fix propagates to all 8 languages.

Validated three ways on the real injected data (taxonomy WHERE carte∈{1,2},
176 rows): (1) Node + vendored Handlebars → 7 families / 21 sub-families /
63 leaves; (2) OLD-template render proving the regression (10/53, Tricherie=0);
(3) LIVE CardPen engine (Playwright) → 7×3×3 DOM + 7 distinct palette colors
(Obstruction red present, no Influence duplicate) + viewed screenshot.

Refs #443

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(rules): index-independent heading CSS + cover titles (#438 follow-up)

The 24->15 Rules deck restructure (#438) broke the heading CSS, which keyed
banners and number-boxes to card INDEX (nth-of-type + hardcoded content
'01'..'05' + [class~="4"]). After the restructure those indices were wrong:

  1. «Déroulé de la manche» rendered in a colored banner instead of a plain
     big title like «Règles du jeu…».
  2. «Argumentum» rendered big on the sub-game cover cards instead of the rule
     name.
  3. Inconsistent paragraph headers: empty floating number boxes + size
     mismatches on cards 6/10/12/14/15.

Fix (index-independent, semantic headings):
  - h2:first-child = card main title  -> plain, big, no banner
  - other h2       = section sub-head -> plain, thin theme-colored underline
  - h3             = numbered step     -> theme-colored banner; the «1.» number
                     is already in the heading text, so no separate (broken) box
  - removed the leading «# Argumentum» line from the 4 sub-game cover rows
    (Rules_07/09/11/13) across all 8 language columns so «## <rule name>»
    becomes the title; Rules_01 (main cover) keeps «# Argumentum».
  - tightened heading margins/padding so the densest cards do not overflow
    worse than the committed baseline.

Validated live on the real CardPen engine (all 15 cards): zero bannered
titles, covers lead with the rule name, uniform theme-colored step banners;
total text overflow reduced 208px -> 110px vs baseline (cards 11/13 fully
un-clipped, only card 12 +1px = parity).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
jsboige added a commit that referenced this pull request Jul 3, 2026
…iduals NO-OP proof (#661)

Secondaire dispatch ai-01 (msg kp26j6) — data pendant of visual verdict #140.

Rules audit (Cards/Rules/Argumentum Rules - Cards.csv, read-only):
- 15 records x 8 languages = 120/120 cells OK (populated + correct script).
- 0 empty, 0 contamination residual post-#640 (scanner #647 = 0).
- 2 heuristic flags cleared as false positives: Rules_01 ZH (说谎者学校
  present, latin inflation from brand "Argumentum"); embromador (canonical PT
  game-term for baratineur, 77x in Scenarii PT).

Primaire (Fallacies EN residuals) = NO-OP, proven empirically:
- 0 empty desc_en across 1408 rows; PK 371/607 (cited empty) already populated
  (120/79 chars).
- Scanner #647 TOTAL=0; 0 EN identical-to-FR; 0 EN reading as French.
- The #351/#308 drift (40 PK) was resolved in an earlier cycle.

Reporting the no-op rather than fabricating work / spending gpt-5.5 on nothing
(pattern: verify before concluding, [[matcher-no-match-is-not-content-absent]]).

Read-only, 0 CSV write. Base 9c19e51.

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

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants