skillfactor-pipeline fc509d0e4c feat(gate): ad relevance gate with blocklist + gemma3 judge, eval 29/30 ...

Two-stage gate (deterministic role-family blocklist, then LLM judge against
the ESCO description). Eval set: 30 real titles from the v1 corpus, 15 of
them the actual designer contamination. Ship threshold 27/30, achieved 29.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PDKeXvpT6tENSvyQGLV1Uq

2026-07-09 06:30:34 +02:00

..

eval-set.json

feat(gate): ad relevance gate with blocklist + gemma3 judge, eval 29/30

2026-07-09 06:30:34 +02:00

run_eval.py

feat(gate): ad relevance gate with blocklist + gemma3 judge, eval 29/30

2026-07-09 06:30:34 +02:00