How do I find the evidence behind an AI result?

Open Evidence for a task and time range, then follow records back to the work they describe. Related task, run, and skill pages provide the detailed workflow views.

Record	What it establishes
Comparison	Which task and dataset versions were tested, the candidates and scoring settings, and the returned answers
Release	The selected model and task version, snapshotted validation and routing settings, and the comparison link when supplied
Task request	The input, model response, release, rule results, and retrieval attribution recorded for that request
Agent step	What the integration submitted, Orlo’s decision, and the result reported by the runtime
Approval	The exact held call, decision, person who decided, and timing
Skill retrieval	Which person and app retrieved a version, with its content hash
Reviewed example	The correction or saved step added to a dataset version and its origin

Inspect one result

Find the request or run, confirm the task and version, and read the checks or permission decision. Follow the release to its comparison when one exists. For a source-backed answer, inspect the passages retrieved for that request rather than assuming the current documents are unchanged.

A release without a comparison link does not establish prior evaluation. A skill retrieval proves delivery, not that a model followed it. An imported trace records reported behavior, not prior enforcement. Use these distinctions when presenting the evidence.

Give reviewers appropriate access

Evidence auditor provides history and metadata without event content. Add other presets only when a reviewer needs the corresponding task, skill, or agent details. Presets cover an area across the organization, not one resource. Who does what explains the choices.

Preserve what your process needs

Agree on retention and export requirements for your deployment before relying on a record for a formal review. Access to records is not a claim of tamper-proof storage, certification, or a completed audit package. Use the documented API for supported reads and follow your organization’s evidence handling process.