Replay a fix before it touches a live system.
Replay real sessions, including past remediation work, against a new model, prompt, or skill, roll out on evidence, and keep the telemetry in your own stack.
| Invoice exceptions | Current model | Candidate model |
|---|---|---|
| Sessions replayed | 120 | 120 |
| Completed correctly | 104 | 109 |
| Prompt injection cases | Held | Held |
| Result | No regression on the replayed sessions. | |
Replay real sessions
Turn yesterday’s work into a test for tomorrow’s change.
Test a policy change first
Replay past decisions against a proposed policy to see what it would have allowed or blocked.
Done means verified
Coding work can require an independent review and passing checks tied to a specific commit.
Telemetry you own
Export OpenTelemetry traces with model and cost details, without conversation content.
Ship a change with evidence
Shared by every surface.
Use it with Dex on every surface, alongside company context that existing agents read over MCP.