Agent Trust Lab · Decision workspace

Choose with evidence.
Review before action.

Describe the capability and constraints you need. Trust Live compares versioned provider evidence and makes unknowns visible for human review.

Trust Live is read-only: it never edits a repository, requests write permission, or calls a provider. Synthetic fixtures demonstrate the decision flow and are not real provider recommendations. Evidence Lab measurements are not yet mapped to individual providers.

1 · Measure and inspect evidence2 · Review a task-specific decision

Decision requirements

Only the selected capability and constraints are sent to this app's recommendation endpoint.

This produces a deterministic review result. It never calls a provider tool.

Human review boundary
Recommendations do not authorize an action. Safe-Call remains a separate review step; provider execution and payments are disabled.

Decision result

Your results will appear here

No provider call

Set your requirements on the left, then review a clear comparison here.

Eligible

Matches the selected capability and evidence requirements.

Needs human review

Important provider facts are missing or still uncertain.

Not eligible

A provider conflicts with one of your task constraints.

These are possible result types, not recommendations for your current selection.

Evidence Lab measurementsMeasure how agents selected tools in preserved benchmark cases. Browse the Evidence Lab. These records are not yet mapped to Trust Live provider identities.
Trust Live provider evidenceCovers capability, source, observed behavior, and Safe-Call constraints shown with each candidate. Recommendations currently use this evidence lane only.

Version evidence · synthetic replay

See how a capability change affects a decision

This replay uses a synthetic provider fixture. It is not a live provider update, audit, or claim about any real endpoint. The version change is evaluated through the same Trust Live recommendation engine.

Version evidence · pinned source snapshots

Review a real source change

This compares locally validated GitHub MCP Server source snapshots for issue_read:get. It does not identify the version currently running on GitHub’s hosted endpoint.

The two evidence types remain separately identified. Selection measurements are not provider safety certifications, and missing provider facts remain unknown. Official Registry data quality grades describe catalog records only; Trust Live task recommendations do not use or inherit those grades. View Registry data grades. Privacy · Terms · Support.