Prompts, classes, and the verification pass

Lexi builds classes from your sentence, so say what the field crew would say: the objects, the defects, and the context. 'Detect insulators and rust corrosion on the pins' produces two clean classes with the relationship understood.

Boxes answer where and how many, and they are the default. Polylines fit linear assets like pipelines and power lines. Polygons and masks are for questions where the exact shape is the answer, and they cost more per label because the answer carries more information and takes longer to draw and verify. Start with boxes; upgrade the classes that prove they need more.

tip · Shape type drives what the work costs. Tell us the mix you expect and we will price it with you before any labeling starts.

Pre-labels arrive as drafts and our team works them: confirm what is right, fix what is close, reject what is wrong. A verified frame is one an expert has actually seen, which is the difference between a dataset and a guess. Where we are not certain about a frame we flag it for review rather than guess, and those frames land in 'Awaiting your review', a queue you can work or leave with us. Every other frame reads 'Expert-verified' where our pipeline made the call and 'Verified by you' where you did, so you can always tell whose judgment a label carries. The review controls in the product are for that queue and for corrections you want made, not for doing the pass yourself.

tip · Reviewers anchor on drafts. A plausible wrong label gets accepted more often than a blank frame gets a missed one drawn in. And because corrections train the next pass, an accepted mistake does not stay a single mistake; it seeds the drafts that follow. Two habits in our process break the loop: a small random sample re-checked from scratch each session, and a deliberate second look at frames where the draft found nothing.

Every correction feeds the next pre-labeling pass on your footage, ours on the verification pass and yours in 'Awaiting your review'. Expect the second batch to arrive noticeably cleaner than the first; the gap between draft and verdict narrows with every session.

Renaming or splitting classes mid-project fragments your dataset. Settle the vocabulary with your team early, the words the alerts will use later, then send it to us so we label against it consistently. Treat the class definitions as a living document. An edge case will force a ruling sooner or later. Does surface staining count as rust? Write the ruling down and send it to us, so it reaches the people doing the labeling. The spec is also the ceiling: a model cannot be more consistent than the labels that taught it, so two verifiers disagreeing on a definition bounds the accuracy before training even starts.

tip · If the field crew would not recognize the class name, the alert built on it will confuse them too.