Add honest override: future prediction queries return uncertainty response

Fix false positive: allow historical kill-context, block intent-based only

Safety filter: 53/53 tests passing across all critical categories

Expand safety filter — 28/28 test cases passing

Add input/output safety filter and increase max_new_tokens to 512

Upload AlignX llama2_7b checkpoints and model card