Make the goal the measure of good work.
An agent can finish a run without finishing the job. airis keeps the acceptance criteria, the recorded evidence and the review decision in one view, so a team reads what was met, what is still pending and what was only claimed.
Defined before it is claimed.
Every condition names the evidence that can satisfy it before work starts · Sample data
A report is a claim · A check is evidence for one condition · A review is a decision · Sample data
Compared, not applied · A comparison changes nothing until a person applies the change
01 / 04 The measure
Done means the goal was met, not that the run ended.
- 01 Every task carries an acceptance pack: named conditions, the evidence that can satisfy each one and the person who decides.
- 02 Agent reports, recorded checks and human review are kept as separate rows, each with its capture source attached.
- 03 A comparison holds the input, revision, environment and criteria constant, and keeps unfinished and held-out cases in the result.
02 / 04 Define, inspect, compare
Three questions, answered separately.
What must be true, what the evidence shows and whether a change is better are different questions. airis keeps their answers apart, so no one has to infer one from another.
01 Define acceptance
Say what done requires before the work starts.
An acceptance pack turns the goal into conditions a team can check. Each condition names the evidence that can satisfy it and the person who decides when evidence is not enough. The pack is versioned and travels with the task, so every retry starts from the same definition.
- Conditions written from the goal, not from the diff
- An evidence source declared for every condition
- One named reviewer for what a check cannot prove
| Condition | Satisfied by | Decides |
|---|---|---|
| Median CI run under 8 minutes | Recorded CI timing on the candidate branch | Check |
| All 248 suites retained, exit 0 | Final suite run after the last edit | Check |
| Peer review, then rollout | William reviews; Mike decides rollout | William · Mike |
Goal · Ship changes faster and safely · Pack reused for every retry of this task · Sample data
02 Inspect the evidence
Read the claim, the checks and the decision as three things.
At handoff the agent says the task is complete. airis shows that report, but counts only what was recorded: the CI timing captured on the candidate branch, the exit code of the final suite run and the reviewer’s decision. Two of three conditions met is the honest state, and it stays visible until William decides.
- The agent report sits beside the evidence, never inside it
- Each recorded check carries its scope and capture source
- Pending rows stay pending; nothing is filled in
A passing check proves its own condition, not the whole change · Sample data
03 Compare on equal terms
Judge a change by the terms it will be held to.
A prepared change runs against the same twelve CI tasks as current practice: same input, same revision, same environment, same acceptance pack. Unfinished tasks stay in the denominator and four held-out cases are never used during preparation. The result is a comparison a reviewer can trust, not a highlight reel.
- Paired tasks, with unfinished attempts retained
- Held-out cases kept apart from preparation
- Acceptance judged by the same pack and the same reviewer
| Term | Held constant across both arms |
|---|---|
| Input | The same 12 CI tasks, paired one to one |
| Revision | One pinned commit |
| Environment | One runner image, cold cache |
| Acceptance | Pack v1, judged by the same reviewer |
| Unfinished tasks | Retained in the denominator |
| Held-out cases | 4, kept out of preparation |
Compared, not applied · A concept comparison, not a measured airis result
03 / 04 Rules of evidence
What counts as evidence. What is only shown.
Every row says what kind of thing it is. That is what makes an acceptance rate readable, and what keeps an illustrative example from being mistaken for a measured one.
A report is a claim.
The agent’s own summary is displayed next to the evidence. It never satisfies a condition and never moves an acceptance count.
Sample “Complete” shown beside 2 of 3 conditions met
A check proves one condition.
A recorded command result satisfies the row it was captured for, with its scope and source attached. It does not prove the change is correct.
Sample 7m 42s median run · candidate branch only
A review is a decision.
Human acceptance is its own row with a named reviewer. Until it is recorded the task is pending, and pending is shown as pending.
Sample William reviews next · human review pending
A comparison keeps its terms.
Input, revision, environment and criteria stay constant. Unfinished tasks and held-out cases remain in the result, which stays concept data until measured on your own work.
Concept data 7 → 10 of 12 accepted · 4 held-out cases retained
04 / 04 Related
One evidence base, three outcomes.
The acceptance evidence on this page is the same evidence the cost and delivery stories use, for the same sample workspace.
Cost per accepted task keeps a cheaper failure from looking like a saving.
02 · Outcome Delivery speedA complete review packet, prepared while the checks run, shortens the wait for a decision.
Platform agentacctCost, quality and speed for the same tasks, in one connected workspace under airis.
See it with your own acceptance criteria
See which conditions
your agents met.
A guided look at acceptance, evidence and comparison for a sample cohort, and a conversation about how your teams define done.