Connect. Check. Keep the proof.
Run the real local engine against a staging adapter. Only fresh synthetic test records are created and deleted. No application credentials are sent to the public Trace website.
- 01 Connect staging
- 02 Review coverage
- 03 Understand the result
Connect an adapter
The inventory-only connection check does not create or delete records.
Review the coverage
“Checked” means planned API observations; it does not mean independent discovery of every store. The result shows which observations actually completed.
Connect to see the plan.
A missing component is never treated as proof of zero records.
Connections are immutable. Use a new name if you change the URL, profile, credentials variable, or coverage.
Your result will appear here.
Verification outcome, coverage, and cleanup are reported separately.
Staging evidence and regressions
Completed reports survive restart. Resume a running/interrupted job using its ID.
Saved regressions
Deleted should
mean deleted.
Test what your application forgets.
Keep the evidence. Catch the regression.
Configure a run
SYNTHETIC DATA ONLYInspect the evidence
READYProof starts with a run.
Choose a fixture and test sequences.
Results come from observed store contents,
retrieval responses and preserved controls.
Precisely scoped.
Honestly reported.
What we verify
API-visible source records, vector records and cached answers after deletion and controlled worker completion. An unrelated record must survive.
What we do not claim
Disk or backup erasure, model unlearning, historical chat deletion, production coverage, or independently reproduced upstream vulnerabilities.
The evidence ledger.
Reports are append-only through Trace. Saved test definitions are content-addressed. Local file access is not a security boundary.
Every run leaves a lesson.
Confirmed outcomes inform which test runs first. This is a transparent history heuristic—not model training or evidence of generalization.
Calculated per backend over the latest 500 reports. Inconclusive tests do not count as successes. Fixture mode influences these observations; this is not a production risk estimate. Test assertions are never relaxed.
Evidence in your workflow.
Connect an MCP-compatible client to the same local engine. No model subscription is required to use Trace's CLI or dashboard.
Local stdio connection
Install the optional MCP extra first: python -m pip install -e ".[mcp]". This configuration uses the Python executable running this dashboard. Restart your client after adding it.