Found first: a primary source the press has not covered yet.
A new paper tests whether Orthrus speculative decoding actually produces outputs identical to standard autoregressive inference. Under BF16, the precision used in production deployments, exact trajectory equivalence holds in only 43 to 45% of cases, according to the paper. Under FP32 it holds on every evaluated prompt.
What the source says
Koziev, Sinev, and Oseledets evaluated two Orthrus checkpoints across 1,190 prompts from 12 domains: the checkpoint released by the Orthrus authors and one they trained independently. In BF16, the Orthrus authors' checkpoint matched the reference trajectory 45% of the time; the independently trained checkpoint matched 43% of the time. FP32 produced exact matches on all evaluated prompts. The probability of exact matching correlated strongly with the response-conditional perplexity of the reference model. No systematic degradation appeared on lm-eval-harness benchmarks despite the trajectory divergence.
Why it matters
Orthrus's central claim is that its intra-model consensus mechanism makes speculative decoding lossless, producing token sequences identical to vanilla autoregressive inference. That guarantee holds in FP32, but in BF16 the outputs diverge in more than half of all runs. The paper draws a distinction that matters for anyone evaluating speculative decoding systems: passing downstream benchmarks is not the same as generating the same output. The losslessness claim is technically accurate under a precision setting that standard production inference does not use.