Changelog

Tracing ingestion

• Live tracing ingestion keeps runs updated as events stream in.
• Parallel processing accelerates ingestion and reduces latency.
• Clearer error messaging when tracing encounters issues.

Transcript groups, metadata, and filtering

• Transcript naming and metadata support for richer context on every run.
• Preview data values in the filter picker so you can see what's available.
• Bug fixes mixed metadata types

Tracing

• Transcript group display in the front-end + some quality of life improvements for dealing with many transcripts in an agent run
• Disable tracing by setting an environment variable
• Miscellaneous improvements: better error messaging, more robust ingestion, support for new formats, the ability to have fine-grained control over which libraries get instrumented, and bug fixes

Ingestion

• Improved speed and reliability when ingesting large Inspect .eval files; uploading logs of multiple GBs now works

Rubrics and judges

• Bring your own API keys for running rubrics with OpenAI/Anthropic/Gemini models
• Choose a specific LLM when running rubrics

Transcript ingestion

• Removed metadata schemas, so ingestion no longer requires predefined schemas
• More robust Inspect one-click ingestion: Docent used to silently drop metadata and scores when using our file upload or drag-and-drop ingestion; this is now fixed.
• Auto-reload agent run list after successful file upload

Tracing (now in public preview)

• Transcript groups: Ability to group transcripts in the Docent UI, with support for hierarchical relationships between groups.
• Incremental loading: Tracing data is now loaded into the collection as it is being run.
• Improved instrumentation: Auto-instrumentation is now more targeted and can also be manually configured.
• Improved stability: Various bug fixes, primarily related to supporting a wider range of models and formats.
• Documentation: Now publicly available.

Charts

• Export charts as PNG and CSV
• Resizable chart area

Misc

• Upgraded to Claude 4 Sonnet from Claude 3.7 Sonnet for rubric evaluation
• Internal infrastructure improvements

Rubrics and judges

• Rubric versioning: Edits to rubrics are now versioned, allowing you to compare results to past iterations

Charts & metadata

• Any numeric metadata usable as a chart measure: Previously, only scores were allowed
• Charting uses latest rubric version to avoid plotting stale judge results

Infrastructure

• Improved scalability of application serving infra

Clustering & Charts

• Clustering metrics: Allow switching between counting total results vs. unique agent runs
• Interactive chart filtering: Click on chart elements to filter data dynamically
• Improved chart creation workflow: Auto-populate new charts with valid data
• Fix charts sharing: Previously did not work

(New) Tracing & instrumentation

This feature is still in private preview

• Auto-instrumentation of OpenAI and Anthropic Python SDKs
• Simple Python SDK with decorators and context managers for tracing agent scaffolds
• Auto transcript splitting logic that converts individual LLM calls into contiguous transcripts

Infrastructure

• VPN networking infrastructure for private access to Docent
• End-to-end PostHog analytics for internal observability