PRD 5.14 rescris cu pivotul arhitectural: LLM doar etichetator OFFLINE,
runtime = clasificator local fara API (fuzzy + embeddings), baza de
cunostinte GOLD partajata cross-account (validarea unui service ajuta
toate). Decizia 8 (corpus per-cont) SUPERSEDED.
Tooling nou OpenRouter (free, familia NVIDIA Nemotron): or_common.py
(client + corpus pe frecventa, cheie din .env) + or_modeltest.py
(comparatie modele, acord ensemble vs Groq). Masurat: super-120b +
nano-9b fiabile, 3/3 unanim pe 87% volum; ultra-550b aruncat.
Corpus real (4 CSV service, coloana NR=frecventa) + etichete Groq
bootstrap incluse ca date de masurare.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
