ACCURACY & TRUST
We do not quote an accuracy percentage - not until enough expert-graded reviews exist to defend one. What we publish instead is how the system is forced to behave: a confidence contract, the judgement calls written down as rules, and a feedback loop that makes the engine better without letting it get worse.
The contract is simple. Findings above the confidence bar arrive with the sentence of visible evidence that supports them. Findings below it ask for a closer photo, a video, or a reviewer - with the cause attached from a closed, countable list. Nothing uncertain slides silently into a priced report.
DISCRIMINATION RULES
The judgement calls, written down - the rules the model must follow before it may cite a code.
Discoloration is not damage
Tarnish, patina and non-chrome finishes are cosmetic. “Damaged” requires structural harm you can see - a crack with depth, a break, flaking through the finish.
Wetness is not buildup
Droplets and wet film photograph white, like mineral crust. Buildup must be dry, crusty deposit; anything wet nearby is reported as the leak.
One problem, one code
A worn faucet’s tarnish, collar gap, residue and drip are one story, not four findings. Side-effects are evidence, not extra billing lines.
Absence needs the whole span
“Missing trap” requires seeing the whole run where the trap would be. A cropped view can never prove absence - but a visible open pipe end is positive evidence of disconnection.
A photo has no time axis
Leak rates, active flow and causation are never asserted from a still. Only a video span may claim the time-based conditions - slow drainage, blocked, inoperative.
Severity needs visible hazard
Where a call exists as both Safety Issue and Repair, Safety is chosen only when the hazard itself is visible - otherwise Repair. Reviewers upgrade more easily than they un-ring alarms.
THE LOOP
How it gets better without getting worse.
Expert review
Every result can be confirmed, corrected or replaced by a reviewer. Both versions are stored side by side - the model’s answer and the human’s.
Golden set
Reviewed images become a regression set. Prompt and rule changes are re-scored against it before they ship, so a fix for one photo cannot silently break ten others.
Versioned everything
Prompts and the taxonomy are versioned into the analysis cache key. Every stored result names the exact rules it was computed under.
Evidence standards per code
Codes that prove confusable get a written “looks like” standard the model must meet to cite them - accuracy fixes become registry edits, not prompt rewrites.
The pipeline these standards govern is on the how-it-works page - and you can see the output on the sample report.