Inspectable Scores: What a Trace Is and Is Not
An inspectable score needs evidence. This note defines trace metadata and the payloads that trace rows do not duplicate.
Abstract
When capture is enabled, Versalist can attach a Trace to an Episode. A Trace is a read-only view of captured behavior metadata. This note defines the view and its data limits. It also explains the separate outputs that an Episode can retain.
The honest question
When a platform reports a score, users need the evidence behind it. A rubric explains the judgment. Captured event metadata shows the ordered calls.
When trace capture is enabled for a run type, its episodes can include a Trace. Capture is best-effort and can be incomplete.
What a Trace is
A Trace is a read-only view of one Episode. It shows ordered events with status, model, latency, and token counts. Its summary shows dropped and known truncated states. The system does not present a known incomplete view as complete.
What trace events do not duplicate
`trace_events` stores bounded metadata and hashes. These rows do not duplicate raw prompts, model completions, or full transcripts. The episode executor can retain public-case agent and evaluator outputs in `episode_steps`.
Service-run private cases do not retain those outputs. This design limits payload storage in trace tables. It does not create a platform-wide no-retention guarantee.
What the claim is not
A Trace supports behavior inspection. It is not a tamper-proof audit log. It does not prove score integrity or support payload replay.
Where this goes
Future runtimes can add event types after they execute the related behavior. Future views can also compare captured events across Episodes.
- Cite this note
- Versalist Research (2026). Inspectable Scores: What a Trace Is and Is Not.
- Publication record
- Published May 21, 2026
More from the research notebook
Evaluation Generation as a Post-Selection Step
Most generated challenges never ship, and evaluations written for them are wasted work, so the pipeline defers evaluation generation until promotion. What that buys, what regenerate-not-edit means in practice, and what the generated module is not.
Read note →Skills as Accuracy PrimitivesSkills as Accuracy Primitives
On Versalist, a skill is a named and versioned instruction bundle. Clients can record observations, and proposed changes use recorded history for scored comparisons.
Read note →Agent Training MethodologyThe Adversary Agent: Hostile Review as a Pipeline Step
A second agent, prompted to assume the code is broken until proven otherwise, reviews changes it did not write. We describe the prompt shape, how the step is enforced, and what happened the day it returned nothing.
Read note →