Who, and Which Run
The two facts a record has to keep separate.
When an AI agent finishes a task and reports “done,” a record of that moment usually contains less than it looks like. There’s a timestamp, a log, maybe a trace. What’s often missing is the ability to answer two plain questions without cross-referencing three systems by hand: who was this agent, and which specific governed run am I looking at?
Those sound like the same question. They aren’t — and the gap between them is where “the agent said it was done” quietly stops being checkable.
We learned this the direct way, building our own evidence layer. Identity and run-correlation were being reconstructed after the fact from platform logs rather than bound into the record when it was written. Reconstruction is a plausible story. It is not evidence. So we changed the shape of the record itself.
In the build we just completed, the session model binds two distinct server-side facts. One identifies the agent. One identifies the governed run — a correlation the server issues, that the caller cannot choose, and that becomes write-once once established. They stay separate on purpose. Collapse them into a single field and you lose the ability to say “this outcome belongs to this run and no other.” This is built locally; it is not a production claim until the migration is deliberately applied and the path is deployed.
The discipline that falls out of this is the part we care about most: an outcome with no matching receipt never renders as success. Inside a short grace window it reads as pending — the receipt may still be landing. After that window it reads as contradicted, and it says so, loudly. Absence of evidence is never dressed up as a result.
Authentic individual artifacts are not the same thing as an established end-to-end outcome. A very detailed log is still a detailed claim; a record that preserves those relationships explicitly is a different thing.
If you’re evaluating any agent-governance tooling — ours included — ask three separate questions: which agent was involved, which governed run does this outcome belong to, and what authority does the record actually establish? You should not have to stitch together log sources to answer them.
We’re not claiming this is shipped yet. We’re showing the design principle we are enforcing: preserve agent identity, run identity, authority, outcome, and evidence as separate facts so the final record does not have to reconstruct them after the fact.
When this ships to production, we’ll follow up with the record — not just the note.