Respan: Span-01

Span-01 is Respan’s behavior-scoring model for conversations and agent traces. It reads a text interaction together with behavior definitions written in plain language and returns probabilities for each definition. The output distinguishes a behavior that is present, absent or not observable in the supplied trace. This gives an evaluation workflow a numerical signal with an explicit way to represent missing evidence, rather than a generated explanation to parse.

Define the behavior you want to observe

Behaviors are supplied by the caller. Prompt injection, unsupported claims, secrets exposure, repetitive loops, tool misuse and user frustration are possible definitions, not a fixed taxonomy built into the model. State what counts as the behavior and what does not. A narrow definition can distinguish a user asking for a human handoff from merely expressing disappointment. Another can distinguish an assistant acknowledging a problem from promising an action it cannot perform.

Use stable identifiers with the definitions, so results remain associated with the behavior being evaluated. Keep the criterion focused on something the trace can show. If the behavior requires a source document, a tool result or a later customer response, include that evidence in the text interaction where it belongs. A definition alone cannot establish a fact absent from the trace.

Prepare the preceding context and target turn

The documented input separates preceding messages in span.input from the target turn in span.output. Each message’s content is a string. Supply the actual conversation context and the turn you want scored, preserving the speaker roles and the order needed to interpret it. This is a text-scoring profile; prepared document text or a transcript can be evidence, while media from which that text was obtained is not itself the documented input.

For a support interaction, keep the customer’s request with the assistant’s proposed reply. For a tool-using agent, retain the relevant tool observation and the answer that relied on it. An illustrative grounding definition might ask whether the assistant asserts a conclusion that the supplied observation does not support. The useful result concerns the defined behavior in that span, not a complete diagnosis of the whole application.

Read three probabilities for every definition

Each result includes the behavior identifier and p_present, p_absent and p_not_observable. The three probabilities sum to approximately 1. Present estimates that the behavior occurred; absent estimates that it did not; not observable estimates that the supplied span lacks the evidence required to decide. These are separate meanings and should stay separate when results are interpreted.

For example, a span that ends with an assistant saying it will solve a problem may show a promise. It cannot yet show that the customer accepted the eventual solution if no customer reply is present. Treat that missing-evidence result as unknown rather than converting it to absent. A low probability of presence is not automatically strong evidence of absence when a significant share of the distribution is not observable.

One-pass scoring across behavior definitions

Respan describes Span-01 as a reasoning classifier trained first for general classification with RLAIF and then specialized for behavior detection. Each definition runs as a parallel branch over the same trace, and the model returns the behavior probabilities in one forward pass. Hybrid attention supports its treatment of long, complex traces. It does not generate tokens, write a narrative judgment or produce a free-form explanation of its classification.

The concept guide and scoring reference document no per-request behavior-count cap. Every definition is still part of the supplied evaluation, so keep each one clear and relevant rather than treating the absence of a count cap as evidence of unlimited input capacity. The result contains one entry per behavior in request order, identified by its id. Multiple definitions describe multiple independent observations; they are not mutually exclusive category choices that must add up to one across behaviors.

Preparing a useful evaluation artifact

Combine the returned probabilities with task-specific thresholds and review rules outside the model. The producer recommends using labeled conversations to choose those thresholds and inspecting borderline presence or high not-observable results. A support workflow can record frustration, a proposed escalation and an unsupported promise separately. An agent-quality workflow can track repeated actions or whether a response used the evidence made available to it.

Keep the text span, behavior definition, result identifier and full three-way distribution together. When the definition changes, recognize that it changes what the probability measures. Review apparent errors against the actual conversation instead of treating a numerical result as proof of an unobserved event. A useful artifact is a behavior record tied to the trace and criterion, with uncertainty retained for later inspection or an application’s next decision.