Span-01The first hyper-parallel reasoning classifierbuilt for unseen challenges
User frustration
Laziness
Unsafe
Jailbreak
Win
Escalation
Refund
True single-forward classification with extraordinary behavior detection across long production traces.
Span-01 brings frontier-level judgment to every trace without token-by-token generation. Span-01 Lite makes the same workflow completely free.
Jev proved that decision models can make classification fast and practical.
But can a classifier preserve frontier-level reasoning, generalize to unseen definitions, and carry that reasoning across long traces in one true forward pass?
Span-01 is built to reason over new natural-language definitions, then return direct probabilities across every definition in parallel.
Reason first. Classify once.
The difficult part of classification is not returning a label. It is applying a definition the model has never seen to context it has never encountered.
We train general classification reasoning with RLAIF before specializing Span-01 for behavior detection. Instead of memorizing a fixed taxonomy, it executes unseen behavior definitions across new behaviors and domains.
At inference, hyper-parallel definition branches evaluate multiple unseen behaviors over the same trace context. Hybrid attention is central to preserving that reasoning in long, complex traces.
One true forward pass. No token-by-token generation. Span-01 returns direct present, absent, and not_observable probabilities.
Frontier-level judgment.
The first test was whether a dedicated classifier could retain the judgment quality of frontier models. Across nine models and both English and multilingual behavior detection, Span-01 ranks first overall. Span-01 Lite also outperforms Jev and Sonnet 5.
Behavior benchmark
Overall F1
Quality vs. cost
Overall behavior F1 vs. normalized inference cost
Overall is the unweighted mean of English and multilingual F1. Raindrop uses macro F1 (n=61), normalized from $0.003/event at 1,000 tokens/event; the $299/month base fee is excluded.
Built for the behaviors production teams actually monitor.
Aggregate scores can hide the behaviors that matter in production. We separated safety, reliability, grounding, agent behavior, and response quality to see where the signal holds. Span-01 outperforms Jev and Sonnet 5 overall, while a full frontier model remains the strongest ceiling.
Production behavior benchmark
Behavior detection across safety, reliability, grounding, and quality
| Behavior domain | Span-01 | Jev | Sonnet 5 | GPT-6 Sol |
|---|---|---|---|---|
| Jailbreak and prompt injection | 0.779 | 0.752 | 0.709 | 0.878 |
| Safety and refusals | 0.803 | 0.751 | 0.785 | 0.911 |
| Privacy and secrets | 1.000 | 0.948 | 0.901 | 0.935 |
| Hallucination and grounding | 0.796 | 0.673 | 0.756 | 0.903 |
| Agent and tool reliability | 0.845 | 0.691 | 0.677 | 0.861 |
| Task and instruction following | 0.702 | 0.671 | 0.621 | 0.821 |
| Response quality | 0.771 | 0.691 | 0.722 | 0.956 |
| Overall | 0.806 | 0.716 | 0.719 | 0.885 |
Decision models judged by Span-01.
We then inverted the benchmark and used Span-01 as the evaluation layer for 11 emerging decision models across accuracy, consistency, injection resistance, and calibration.
Span-01 does not appear in this ranking. It produced the evaluation signal!
Decision models judged by Span-01
11 models across accuracy, consistency, injection resistance, and calibration
| Model | ||||
|---|---|---|---|---|
| Jev 1.13.0 | 0.932 | 0.021 | 0.063 | 0.045 |
| Tev1 4B | 0.901 | 0.052 | 0.042 | 0.031 |
| Kev 4B | 0.833 | 0.060 | 0.078 | 0.070 |
| Decider 2B v10 | 0.747 | 0.128 | 0.208 | 0.074 |
| AlexWortega OpenJev 4B v5 | 0.744 | 0.116 | 0.089 | 0.118 |
| Bosun v3.1 1.7B | 0.680 | 0.208 | 0.323 | 0.089 |
| Tiny-Jev 0.6B | 0.646 | 0.128 | 0.224 | 0.267 |
| Laya | 0.542 | 0.298 | 0.396 | 0.092 |
| MoJev 0.85B | 0.501 | 0.150 | 0.307 | 0.129 |
| Open-Jev DeBERTa v3 large | 0.472 | 0.194 | 0.385 | 0.147 |
| Von 1.2 | 0.422 | 0.242 | 0.297 | 0.071 |
Paired flip rate measures consistency under equivalent variants. Injection ASR is attack success rate. ECE is expected calibration error.
Try Span-01 against Jev.
Judge every trace.
Now it's time to get rid of LLM-as-a-judge.
Span-01 is public now. Follow the quickstart to run it on your own traces and behavior definitions.