tracealign-alignment-failure-attribution

Trace unsafe LLM outputs back to conflicting beliefs in training data using suffix-array matching and Belief Conflict Index.

Independent GitHub snapshot · Automated classification · Original documentation governs