Model discovery plus one documented synchronous prediction boundary.
SkyDNN AI · computer vision evidence
Inference is a response. Trust is a workflow.
Discover and lock an entitled model, bind one immutable asset identity, validate synchronous labels, confidence, and geometry, then review and redact derived evidence without inventing provider state or leaking media and credentials.
Rectangles, polygons, and polylines validated against source dimensions.
Registration through review, decision, retention, and redaction.
No image, video, URL, token, face, plate, or internal model path is accepted.
Contract before orchestration
Keep the synchronous provider call honest.
The public SkyDNN repository establishes a bearer-authenticated OpenAPI 3 surface at ai.mappls.com with synchronous multipart POST /predict and model-discovery GET operations. Its prose also mentions /predict/polygon, but that path is absent from the reviewed OpenAPI document and therefore remains selection-required rather than an implemented contract.
Current Mappls material describes satellite imagery, street photographs, and video frames returning confidence, labels, and bounding-box, polygon, or polyline geometry. It also advertises JavaScript integration and human-in-the-loop annotation workflows, but does not make those exact SDK or review-state contracts available in the reviewed public repository.
Reviewed API surface
Four contracts you can trace. One path you must confirm.
Every operation retains the evidence level of the source that established it.
/server/whoamiDiscover the model keys currently exposed by the entitled server.
/modelsList available model details.
/models/{api_model_key}Inspect one selected model before inference.
/predictSubmit one multipart image and receive a synchronous JSON inference response.
/predict/polygonMentioned in prose but absent from the reviewed OpenAPI; confirm before use.
Application-owned evidence lifecycle
Seven accountable stages around one response.
The host owns purpose, asset identity, model locks, policy, human decisions, retention, audit, and recovery.
Asset steward
registered
Purpose, lawful basis, asset hash, media class, capture time, location context, retention deadline, and no raw media in the case store.
Vision service
model locked
Entitled server identity, selected model key, model fingerprint, classes, policy version, and input compatibility.
Application worker
inference requested
One idempotent attempt identity, exact asset and model fingerprints, and a secret-safe provider request boundary.
Application adapter
inferred
Schema-valid status, timing, output hash, labels, confidence, normalized geometry, and provider provenance separated from business policy.
Decision service
review pending
Versioned thresholds, reason codes, uncertainty band, required reviewer role, and immutable inference evidence.
Independent reviewer
accepted or rejected
Attributable decision, reason, policy/model/output identities, separation of duties, and no rewritten inference.
Privacy worker
redacted
Derived detection detail removed at retention deadline while minimal hashes, decision, and audit evidence remain.
Normalize before policy
Validate the shape of every claim.
Provider output enters decision logic only after its label exists in the locked model vocabulary, confidence is finite and bounded, and every coordinate fits the source dimensions.
- rectangleObject detection
x, y, width, height
- polygonSegmentation
three or more bounded points
- polylineLane or line marking
two or more ordered points
Durable records
Persist provenance, not pixels.
Vision case: external identity, purpose, state, optimistic version, policy, and retention boundary.
Asset envelope: opaque asset reference, SHA-256 content identity, media class, dimensions, capture time, location context, and consent or lawful-basis reference.
Model lock: server reference, API model key, model fingerprint, class vocabulary, compatibility decision, and lock time.
Inference attempt: request identity, asset/model fingerprints, lifecycle timing, outcome, safe error class, normalized result, and response hash.
Policy and review: thresholds, uncertainty reasons, reviewer, disposition, explanation, and immutable linkage to one inference attempt.
Retention, audit, command receipt, and transactional outbox records committed with every transition.
Eight real operating scenarios
Model output is only one part of the system.
Each industry needs its own source context, policy, evaluation, reviewer, retention, and downstream identity.
Building-footprint change triage
Satellite source/date, polygon validity, confidence band, temporal comparison, and planner review.
Lane and road-furniture survey
Sequence/frame identity, camera context, polyline or box normalization, duplicate suppression, and field verification.
Shop-sign and POI discovery
Purpose-limited imagery, text/box confidence, Mappls place reconciliation, reviewer confirmation, and privacy controls.
ADAS dataset quality gate
Model/version lock, scenario coverage, false-negative review, no live safety actuation, and release acceptance outside this API.
Roadside asset inventory
Asset class vocabulary, geometry projection, route/date provenance, human verification, and maintenance-system identity.
Vegetation and water segmentation
Imagery resolution/date, area-of-interest identity, polygon topology, seasonal baseline, and agronomist review.
Visual damage triage
Explicit consent, no biometric reuse, confidence as triage only, adjuster review, retention, and appeal evidence.
Face and plate redaction quality
Sensitive-class policy, miss-rate sampling, access restriction, derived-output retention, and independent audit.
Downloadable reference
Run the full review journey without uploading an image.
The maintained Vision Evidence Desk uses synthetic rectangles and polygons to prove model locks, validation, uncertainty, independent decisions, replay safety, restart recovery, and redaction. Its ten tests make no Mappls model-quality or live-entitlement claim.