Add neutral scene grounding for the dense corpus (2,018 frames) Stage-1 grounding produced by qwen3.6-27b: what is visible, no safety judgment. This is the evidence the deployed SOS pipeline scores, as opposed to the hazard annotations, which were written by a hazard-seeking prompt. Reasoning mode OFF — with it on, the model's untagged chain-of-thought is parsed as scene sentences (56-58 per frame instead of 8-10, carrying the very hazard words the grounding rule forbids). D4 lint: 3 warnings / 2,018 frames; 0 overlong.
#7
by brad - opened
No description provided.
brad changed pull request status to merged