Detects behavioral misalignment in AI agents at the trajectory level rather than the language level: instead of inspecting what an agent says, ATLAS inspects the sequence of actions it takes, encoded under the 6A Trace Schema — a semantic layer above OpenTelemetry with provenance markers separating declared from behavior-inferred intent. A 7–14B LLM backbone pretrained on 6A-encoded trajectory corpora, with fine-tuned detection heads, scores deviation in real time, attributes the decisive step, and forecasts failure before it lands.