Strata and PolyStrata Inference Engines Feature Hidden 'Answer Room' Behavior

Strata and PolyStrata LLM inference engines contain a hidden 'answer_room' behavior that silently caps thinking tokens at 75% of the max_tokens limit. It then injects a wrap-up message directly into the model's output stream without notifying the client or logging an error. This hidden modification alters model responses and disrupts reasoning transcripts, affecting benchmarking, reproducibility, and task performance on complex problems. Practitioners comparing these engines to others need to be aware that models served through Strata receive unrequested internal nudges. Triggered by default when thinking is enabled and max_tokens is at least 1024 without an explicit reasoning budget, the engine injects text mimicking the model's voice to force a wrap-up. This abruptly cuts short the chain-of-thought process and can degrade performance on difficult reasoning tasks.