AGENT EVIDENCE
Activity is not proof.
Agent transcripts are useful, but they describe what the agent says happened. Terrarium keeps that separate from the path an event targets and the changes the repository actually shows.
Three different signals
These signals often arrive together, so it is easy to treat them as the same thing. They are not. Terrarium preserves the distinction instead of turning a stream of confident messages into a false record of completed work.
The agent report
A transcript or hook says the agent read, edited, or tested something. That is useful context, but it is still a report from the agent or its tooling.
The target
A tool call or hook names a file and worktree. That connects the activity to a place in the repository. It shows intent, not a confirmed change.
The confirmed change
The filesystem watcher or Git state shows that a file actually changed. Terrarium observes that separately. It still does not claim the change is correct.
A simple example
REPORTEDedited src/auth.ts
TARGETEDsrc/auth.ts
CONFIRMEDfile modified on disk
NOT PROVENthe change is correct
What review first means
After Terrarium confirms changes on disk, it ranks up to five files by receipted risk, downstream reach, change type, and repetition. Open a row to inspect the exact local file and the reasons it surfaced.
The order is a review priority, not a prediction that the code is broken. If the detailed risk model is unavailable, Terrarium labels the fallback instead of presenting the simpler score as full evidence.
Why the distinction matters
One agent can spend a long time producing tests, explanations, or tool calls without moving the work you actually care about. With several agents, the harder question is often which one touched a file, whether two sessions crossed into the same worktree, and what was truly left on disk.
Terrarium gives you a live map and a replayable record for answering those questions. It is an instrument, not a judge. You still review the diff, run the code, and decide whether the work is good.