Understand the work05 / 20
Check the evidence
On this page
An Agent saying “implemented” is not enough. Every explanation comes with its evidence: the code it read, the tests it ran, run records and screenshots, one click away to check. When someone asks, you can point to the exact place.
What evidence contains
Open “Evidence” in the note. The usual kinds of material are:
| Material | Shown as | When clicked |
|---|---|---|
| A code location | A function name, or a “file:line” chip | The code panel opens on the right at those lines |
| Command output, such as a test run | A chip | The full output opens in the auxiliary bar on the right |
| A code diff | A chip | Opens in the auxiliary bar on the right |
| Screenshots and other images | A thumbnail | Opens enlarged |
Command output, diffs and screenshots come from the records the Agent left while implementing. Noea keeps which commands it ran and which screenshots it took in each turn, so an explanation can cite them directly.
Look at a code location
A code location in the middle of a sentence usually shows as the name of a function or variable, such as openGrace; otherwise it shows as a “file:line” chip, such as grace.ts:6-18. If one explanation cites two files with the same name, the chip adds the parent folder.
Rest the pointer on a name or chip to see the full path, line range and symbol. Click it and the code panel opens on the right for that file, at the lines cited. If the file is still as one implementation turn left it, that turn's changes are laid over the whole file, so you can see which turn wrote them. The right-click menu can open the file in its default app, show it in Finder, or copy its path.
If the code has moved since, the chip's line numbers and where it takes you follow its current position. If the cited code has changed and only its name can be found, the chip is dimmed; if the file is gone or the code can no longer be found, the chip is dimmed and struck through, and hovering says why. Either way, the explanation most likely needs to be investigated again.
How far evidence goes
When you look at a piece of material, first tell which kind it is:
| Material | What it can confirm | What to keep in mind |
|---|---|---|
| Implementation or test code | What logic the code uses, and which cases the tests define | Test code existing doesn't mean the tests ran or passed |
| Earlier run results | What ran then, and what the result was then | If the code changed since, the result may no longer hold |
| Verification run during this explanation | The result in this run's environment and scope | Cases that pass say nothing about cases not covered |
Explain and Investigation rely mainly on reading code and existing material; they don't rerun tests, builds or the app every time. The Agent verifies something only when reading the code can't settle a fact the conclusion depends on. Good evidence says what it did not verify, such as “Not run against the real payment provider; charge results come from a test double.” When evidence is thin, no implementation was found or the intent is ambiguous, the evidence and attention should say so plainly rather than offer a conclusion that only looks certain.
Keep in mind, too, that an explanation is an investigation record with evidence, not a certificate. Having one Agent check another Agent's implementation doesn't make the result any more certain by itself. The value of evidence is that you can open it and check, not that the explanation says “confirmed”.
Screenshots and the interface
Some behavior can't be settled from code alone, such as whether a notice really shows up in the interface. If the Agent took a screenshot while implementing, the explanation can list it as evidence, for instance “During the grace period, a notice appears at the top of the team settings page.”
Click the thumbnail to enlarge it. Note which page it shows, in what state, and which run left it: a screenshot proves what the interface looked like then; if the code changed since, the interface may have changed too.
Check item by item
When you check evidence, ask of each item:
- Does this reference point at what this part of the intent is about?
- Which cases do the tests cover, and are the edge cases among them?
- Does the result really support the explanation's conclusion?
- Does the evidence say what was not verified?
For example, the grace-period evidence says “ran the grace-period tests: all 6 pass”, while attention notes that “no test covers the case where the third retry succeeds.” Now you know that path rests on reading the code alone, with no run to prove it.
When the evidence isn't enough
If a key case is missing, hover over the paragraph, click “Ask about this” on its left, and ask the Agent to verify it:
Add a test: the third retry succeeds on day 6 of the grace period, so the subscription returns to normal and is not downgraded. Tell me the result when it has run.
When the Agent finishes, Noea explains the turn on its own; if you turned that off, click “Investigation” on the paragraph's left. The new explanation lists this run as evidence.
If the explanation's conclusion doesn't match its evidence, or you think the Agent misread your intent, ask about that here too, and don't change your intent to suit the code. See Write what you settled into intent.