Findings, severity, confidence

Every module in the product speaks the same finding schema, which is why a Site review judgement and a Speed & errors measurement can sit in one list without one of them lying.

What a finding is made of

Field	What it is
Title	Plain language, describing the effect on a visitor rather than the rule that fired.
Surface	The page it was found on.
Evidence	What the browser saw, with the selector and usually a frame.
Probe / rule	How it is known: which check, which rule, which device.
Severity	Ship-blocking, medium, or low.
Confidence	How sure the tool is, as a percentage it has to be able to justify.
Fix prompt	Paste-ready, ending in an acceptance check.

How severity is decided

Ship-blocking means a first-time visitor could bounce on it, or could not complete the thing the page exists for. That is a claim about consequence, not about how hard it was to detect.

The findings list: every open finding with its surface, severity and confidence score. — Every open finding, with the surface it was found on and how sure the run is about it.

Confirmation is the only route to ship-blocking. A behavioural claim written by a model reading your page is re-driven in a real browser before it can carry that severity:

Confirmed: the steps ran again and the same thing happened. What the browser saw is attached.
Refuted: struck from the report, and listed under what the gate dropped, so a claim that failed is visible rather than missing.
Unverifiable: some claims no browser can settle. Those keep the word unverified on them and are capped below ship-blocking, so one can never block a release on its own.

What confidence means

It is not a feeling. It is a number the tool has to be able to defend, and you can ask it to:

node cli.cjs actions why <id>

That prints the evidence behind the score and behind the item’s rank: which probe fired, whether it was re-driven, whether this class of finding has been right before on this product.

Whether the percentages hold up

node cli.cjs calibration

The tool’s report card on itself, from data already on disk and no model calls: whether the confidence percentages match your own verdicts, and (by capture-recapture over repeated runs) how much of your product no run has met yet.

It refuses to print a rate it does not have the evidence for. Early on it will tell you there are not enough verdicts yet, which is the correct answer and not a bug.