Documents & plans (vision)
What it is: FACE’s image work: grading camera frames
(Vision & CCTV), the vision agent (vision), reading
photographed or scanned forms, and proposing elements from a picture of a site plan.
What it does
It reads images and turns them into fields, grades and proposed elements that FACE and a person can check. Its jobs:
- Camera frames. Grades the inspectable object in a frame: what it is, its
material, a damage grade (
PRISTINE,DAMAGED,DEFECTIVEorDESTROYED) and an estimated damage percentage. Optional detection draws one box per object of the graded class. - Forms. Reads a photographed or scanned document, such as a bill of lading, packing list or delivery note, when text extraction alone can’t.
- Site plans. Proposes zones, docks and yard slots from a picture of a plan, for an operator to confirm.
What it reads
Frames pushed to FACE or sampled from RTSP/HTTP camera feeds, document images, and plan images. Frames are downscaled before the model sees them.
What it produces
- A graded frame with per-object results, each box labelled with the stage that drew it, and notes that explain any absence. See Vision & CCTV.
- Values read from a document image, each recorded with the outcome of the read
(
READ,NOT_PRESENT_ON_DOCUMENT,ILLEGIBLEorEXTRACTION_FAILED). You attach a document with the clip or camera on Ask, and Why shows what was attached to the run. See Evidence documents. - Proposed plan elements, which stay proposals until an operator confirms them, in the Site plans panel on Actionable Twins › Map. See Twin ontology & graph.
Human oversight
- Nothing a vision model proposes from a plan is written to the twin graph until an operator confirms it. See Twin ontology & graph.
- A grade is input for a person. It never approves, rejects or releases anything.
- Any action taken on what a frame or a form shows goes through the same action-card approval as everything else.
Model
The deployment’s vision tier, started on demand.
The live camera view uses a local detector only and never calls the vision model.
Where it runs and data handling
On the deployment’s own model plane. No outside AI service is called on any plan. Images travel in their own typed field, never as text inside a prompt. See Sovereign models.
Guardrails
- A person is not a subject. People never appear in the graded objects, and a frame with no inspectable object gets no grade rather than a low one.
- A detector’s class is never written into a graded label.
- Refuses to downgrade, hide or fake a camera assessment.
- Refuses to steer a sensor cue into unlocking a crane, blinding a camera or standing down a safety interlock, and refuses to plant false events into the shared operational picture.
- A model’s confidence in a value it read from a document is recorded, never used as a truth threshold.
How to read its results
- A grade and a damage percentage are the vision model’s estimate from downscaled pixels. They aren’t a physical inspection.
- A retained evidence frame shows the reviewer what was assessed. It isn’t a forensic record.
- It identifies no one.