feat(dataset-updater): salvage v2 cascade prompt criteria (decoupled from stale #384/#385/#386) - #395
Merged
Conversation
Cherry-pick the two v2 prompt-engineering enhancements developed in the now-stale stacked PRs #384/#385/#386, decoupled from their conflicting data payloads: - PromptFallaciesCascadeDriftUser: add 'Drift sémantique structurel' criterion (translated desc_* carries a concept different from desc_fr → realign on FR, which is authoritative). - PromptVirtuesCascadeDriftUser: add 'Re-conceptualisation des titres' criterion (translated title drifted to a different concept than title_fr → realign on the exact FR concept). Both files are already referenced by the cascade task configs merged in #370; tasks remain Enabled=false. Additive, no data touched, no conflict. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
3 tasks
jsboige
added a commit
that referenced
this pull request
May 30, 2026
…cascade (#396) Part A of two-population strategy from ai-01 dispatch (FleetView 18:55). This commit captures the 0-API-cost salvage of FR-stable cells from the 4 quarantined cascade PRs. The remaining 61 PKs / 61 cells where FR was modified post-cascade (AUDIT_FR scope) will be re-cascaded with gpt-5.5 in Part B (separate PR). === Cell salvage (1,476 cells, 0 API calls) === * Virtues (Argumentum Virtues - Taxonomy.csv): 211 cells from #385 - Source: feat/virtues-cascade-v2-multilang (HEAD dc6c486) - Method: full-file checkout (FR untouched, drift = 0) * Fallacies (Argumentum Fallacies - Taxonomy.csv): 960 cells - 942 from #386 (feat/fallacies-cascade-v2-multilang-results, ec4056f) * Binary-safe row-byte transplant (preserves source quoting + CRLF) * Skipped 61 AUDIT_FR PKs (FR modified post-base in master) → Part B * Invariant proven: 0 FR cells touched in #386 vs base 1811afc - 18 from #355 (data/fallacies-desc-realign-40pk, e628994) * 6 orphan PKs (208/281/286/289/730/1074) × 3 cells (desc_en/ru/pt) * Cell-by-cell mapping by column name (header drift on col 33 'latin' → 'latin1' between #355 base and current master) * Scenarii (Argumentum Scenarii - Cards.csv): 300 PT cells - Source: feat/scenarii-pt-llm-refine (HEAD d3c28ad, #384) - Non-regression vs #383 (PR #383 already merged: deterministic enclitique-hyphen + MT-artefact cleanup) - Rule: apply #384 PT value ONLY where master == base for that cell (i.e. #383 cleanup did not touch it). Skipped 44 cells where #383 cleanup wins. - Per-column: title_pt=55, context_pt=90, issue_pt=56, suggestion_pt=99 === Net-new code from #384 === * DatasetUpdater config block (Enabled=false): - "Scenarii PT refine gpt-5.5" task definition - SequentialChunks, ChunkSize=12, gpt-5.5, MaxChildren=8 - Inserted at end of Configs list, after "Fallacies cascade EN-only" * Prompt files (2 net-new): - PromptScenariiPtRefineUser.txt (+63 lines) - PromptScenariiPtRefineAssistant.txt (+8 lines) * EXCLUDED from #384 (already merged via PR #395): - +1 line in PromptFallaciesCascadeDriftUser.txt ("Drift sémantique structurel") - +1 line in PromptVirtuesCascadeDriftUser.txt ("Re-conceptualisation des titres") === Integrity verification === * FR drift vs master HEAD (5cda433): 0 cells across all 3 CSVs * Row counts preserved: Fallacies 1409, Virtues 224, Scenarii 168 * Build: 0 errors (dotnet build, only pre-existing warnings) * CSV format: CRLF + UTF-8 (no BOM on Fallacies/Virtues, BOM on Scenarii) preserved === Files & deltas === * Cards/Fallacies/Argumentum Fallacies - Taxonomy.csv (1306 lines diff) * Cards/Fallacies/Argumentum Virtues - Taxonomy.csv (422 lines diff) * Cards/Scenarii/Argumentum Scenarii - Cards.csv (258 lines diff) * .../DatasetUpdater/DatasetUpdaterRootConfig.cs (+47 lines) * .../Resources/PromptScenariiPtRefineUser.txt (+63 lines, NEW) * .../Resources/PromptScenariiPtRefineAssistant.txt (+8 lines, NEW) Part B (separate PR) will re-cascade the 61 AUDIT_FR PKs × 7 langues (EN/RU/PT/AR/ES/FA/ZH) using gpt-5.5 cascade-drift task to align non-FR translations with the new FR text. 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
1 participant
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.
What
Cherry-picks the clean, reusable prompt-engineering out of the now-stale stacked PRs #384/#385/#386, separated from their conflicting data payloads. These two v2 criteria are the genuinely valuable, conflict-free "good content" from that trio; the data portions are being salvaged surgically on the post-audit baseline by po-2023.
Changes (additive, +2 lines, 0 data touched)
desc_*carries a different concept thandesc_fr(even if more precise), realign on FR (taxonomy FR fait foi). Caught e.g. PK Sophisme du psychologue.title_fr(e.g.title_fr="Indépendance des événements"→title_en="Assessment of independence"), realign on the exact FR concept.Safety
Enabled = false.Context
Coordinator decision (jsboige 2026-05-29): surgically salvage the good content from the stale PR stack rather than rebase. This PR captures the prompt half; the data half (valid multilang cells excluding audit-reverted PKs) follows in a separate po-2023 PR. #384/#385/#386 to be closed as-superseded once both halves land.
🤖 Generated with Claude Code