NEXT_PUBLIC_EVALGATE_EVIDENCE_UI_V1=1 before building. The default is 0;
changing this public presentation switch requires rebuilding the web application.
It does not grant API access or enable a provider.
Open a trace to inspect its observed span hierarchy, chronological transcript,
canonical tool calls and recorded subagent identifiers. Tool success,
authorization and confirmation are separate observations. Unsealed revisions
have no canonical tool evidence yet. More calls load against the same revision;
an investigation never executes or retries a historical tool call.
In an evaluation, select a candidate and an optional comparison run. The
comparison covers every returned case, including passing, missing and unmeasured
outcomes. Filters and ordering apply to the full returned run snapshot before
display paging. Repeated attempts remain in the expandable case evidence.
Selecting a comparison does not publish it as the baseline. Server decision
receipts and saved reports retain their run identity; an aggregate score cannot
authorize release.
Knowledge search shows the returned lexical or vector method, raw score,
document/chunk identity and source content. A retrieval score is not judged
relevance. Results stay bound to the submitted query; permission failures,
stale embeddings, loading and empty results have separate states. The returned
semantic-search status and reason remain visible alongside any word matches;
word matches do not imply semantic search was available or successful.
Blind Arena review uses completed, stored outputs. A saved preference, tie or
skip reveals the corresponding metadata in that review. Output text can identify
its source and authorized match history remains separately accessible; this is a
presentation for preference collection, not an anti-tampering assessment.
Preference records do not qualify a model or authorize promotion.
Review discussions keep the authenticated author and reviewed material revision.
Posting retries reuse the same request identity. Replying, resolving or reopening
a thread does not change the approval or promotion decision. Refresh after the
review material changes before writing against its new revision.
Evaluation schedules use five-field UTC cron expressions with numeric lists,
ranges and steps. A valid schedule must have a next occurrence within one year.
The persisted next occurrence and last dispatch are distinct from the preview.
Retries identify the same persisted due occurrence; missed occurrences are
skipped to the next future occurrence. Dispatch does not prove a completed run.
Trend brushing, replay and activity views use returned observations. Missing
dates and missing measurements remain unknown. Simulation type and sampling
rate remain visible. The status page’s dependency snapshot and process uptime
do not establish historical availability.
Document extraction results
To inspect source citations in an existing evaluation case, record an object with this optional output shape (or its JSON string representation):evalgate api <operation-id>. The presentation switch does not change server
permissions, release policy or the availability boundaries in
Feature status.