Backend: merge strategy — category defaults are applied as baseline
(ai_generated=0), AI impacts override/extend them (ai_generated=1).
All N category defaults appear in the impacts list; AI covers them
explicitly with event-specific score/direction adjustments.
Prompt: instruct AI to cover ALL default instruments.
Frontend: remove default impact chips under category selector —
they now appear directly in the instrument impacts list.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The AI was returning 'geopolitical' (the event type/family) instead of
a category name like 'Guerre — Moyen-Orient'. Two fixes:
1. Prompt: rename 'Type' → 'Famille', list categories as CAT_N: "name",
explicit warning not to return a generic type, show valid name examples
2. Code: reject any response that matches a known type string before matching
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The AI often returns 'Guerre - Moyen-Orient' (ASCII dash) instead of
'Guerre — Moyen-Orient' (em-dash), causing exact match to fail silently.
Fix: normalize unicode, collapse all dash variants, then try
exact → normalized → partial substring matching.
Also log the raw AI response for debugging, increase max_tokens 1200→2000.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Previously: heuristic text matching (sub_type in category.name) — failed
for semantic cases like 'Iran Ceasefire Talks' → 'Guerre — Moyen-Orient'.
Now: all EventCategories are sent in the prompt (name + type + description).
The AI picks matched_category by semantic understanding, returns it in JSON.
- Validates returned name against known categories (case-insensitive)
- Auto-sets event sub_type when AI finds a match and sub_type was empty
- Returns matched_category in evaluate response + shown in UI eval message
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
_find_cat() returned only default_impacts, never the category name.
Now returns (name, defaults) tuple; cat_name → matched_cat_name.
Fixes: 'name cat_name is not defined' on AI evaluate.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
