Findings and reliability

Cranley’s rule is that a number goes from a calculation to a result object to the screen. It never passes through a model’s text.

What a result holds

Part	What it is
Verdict	pass, warn, fail or unknown, with one sentence
Metrics	Measured values, each with a unit and a source
Checks	One rule each: what was measured, the threshold, and the status
Findings	The things worth a look, each anchored to a room, a doorway, a wall, a point in a panorama or a place on the map
Assumptions	What the tool assumed, and whether the person can edit it

The source on every value

Each metric and each finding carries a source with four parts.

Part	Values
Layer	matterport, vexcel, costar, drone, model or phone
Method	geometry, vision, lookup, simulation, sensor or manual
Confidence	high, medium or low
Note	A sentence about what the value rests on

The app says this in plain words: “Measured from the scan”, “Seen in aerial photos”, “From the property record”, with a mark for how sure. The tool sets the label for each value. A model never does.

For example, the door widths tool labels a doorway’s width as geometry with medium confidence, and its note says that a hinged door’s leaf takes a few centimetres off the opening.

Unknown is an answer

When a scan cannot answer, the tool says so. The verdict is unknown, and the sentence gives the reason: no wall graph, no door data, no location, unscanned space. Every plan carries a quality block that tools check first. In the app, unknown is hatched and never shown as a pass.

When we ran every tool on a 15-place demo set through the real API, 266 runs answered, 62 said “cannot answer here”, 1 failed on a bad input and 0 were judged wrong.

What the agent may say

The agent’s closing message is graded before the person sees it. The grader checks that the message:

1. is at most three short sentences
2. contains no measurement, count, time, distance or money
3. says a check was skipped only when the transcript shows it
4. names where to look first

A real turn once claimed that none of the checks ran, after twelve runs. That mistake is why the grader exists.

What waits for a yes

Slow or invasive steps wait for the person: a smoke simulation, a read of every panorama by the vision model, a plan edit, a roll-up of a whole city. The choices are “Run it”, “Skip it” and “Always run it here”.

What was not checked

Every PDF report ends with a list of what was not checked. The fields are defined in the result format.