From a trace to a decision
See the failure.
Measure the proposed fix.
An AI observability and evaluation platform connects records of live AI behavior with repeatable tests of proposed changes. Stimulir captures or imports traces, preserves available prompt and model lineage, and uses selected evidence for controlled evaluations. Teams can inspect quality, latency, and cost before deciding whether to promote a candidate.
Updated
Explore the workflow in the docs- 01Capture a trace
- 02Fix the cohort
- 03Compare a change
- 04Review the evidence
Production promotion requires separate human approval.
One practical workflow
Turn a recurring summary failure into a reviewed change
Follow one support-summary workflow from observed requests to a prompt revision. Keep the trace evidence and the experiment boundary visible throughout the process.
- 1
Observe the request
Inspect the input, output, prompt version, provider/model, tags, latency, tokens, and cost where available. Repair missing lineage before evaluating.
- 2
Curate the evidence
Select relevant traces, or import an existing JSON, JSONL, or CSV export. Review sensitive data and take an immutable evaluation snapshot.
- 3
Evaluate the revision
Compare the deployed instruction with one prompt candidate against the same task criteria. Inspect regressions as well as aggregate measurements.
- 4
Keep a review boundary
Review the result and recommendation. Applying a promotion is an explicit, separate action; collecting traces does not authorize a production change.
Choose the right layer
What each approach tells you
| Capability | Question it answers | Evidence to inspect |
|---|---|---|
| Observability | What happened in the application? | Recorded inputs, outputs, lineage, and operational measurements. |
| Evaluation | How does a candidate compare? | A fixed cohort, clear criteria, incumbent and candidate results. |
| Promotion review | Should this change reach production? | Comparable evidence, remaining failures, and explicit human authorization. |
Verify the details
Documentation you can inspect.
The published quickstart describes tagged requests and trace inspection. The data-assets reference explains supported uploads and the distinction between current evaluation datasets and legacy trace-import controls. Field availability depends on the source export. The Improve reference documents the path from inspection to reviewed results; the Help center provides setup and troubleshooting guidance.
Bring one live AI failure.
30 minutes with the founder. Bring one live AI failure; leave with an improvement plan.