The three plan-named cases in both languages: two same-boilerplate datasheets for different models that must neither merge nor flag as contradiction (reconcile pair cases), and one multi-model datasheet anchored per section (exact subject_entity assertions under the zero-tolerance subject gate); formalize the deliberately failing hr-x002 title-block case. Refresh the eval cache on an explicitly Mistral-routed configuration and commit the fixtures; state the metric effect in the PR per the ratchet rule; update docs/eval/gate-model.md where it says anchoring is not yet gated. The dedicated published anchoring metric stays with item 6.4 per the plan.
The three plan-named cases in both languages: two same-boilerplate datasheets for different models that must neither merge nor flag as contradiction (reconcile pair cases), and one multi-model datasheet anchored per section (exact subject_entity assertions under the zero-tolerance subject gate); formalize the deliberately failing hr-x002 title-block case. Refresh the eval cache on an explicitly Mistral-routed configuration and commit the fixtures; state the metric effect in the PR per the ratchet rule; update docs/eval/gate-model.md where it says anchoring is not yet gated. The dedicated published anchoring metric stays with item 6.4 per the plan.