AIZOYA/ VisionProof

OpenCV 5 + AWS · Agentic Vision

Visual AI that must verify the outcome.

VisionProof turns visual evidence into a controlled operational workflow. OpenCV changes what the system may do next, consequential actions require human approval, and execution never counts as success until follow-up evidence is checked.

Run public judge demoSee measured resultsRepository

The control loop

See → Reason → Ask → Act → Verify → Prove

No LLM decides authorization. The workflow is deterministic and auditable.

1

See

OpenCV 5 measures visual quality, sharpness, brightness, connected regions, and ROI occupancy.

2

Reason

A deterministic policy maps the measurements to LOW, CONSEQUENTIAL, or EVIDENCE_INSUFFICIENT.

3

Ask

Consequential evidence forces an explicit human approval state.

4

Act

The competition action is simulated only, and only after approval.

5

Verify

A follow-up image is processed again by the same OpenCV pipeline.

6

Prove

A deterministic hash-linked receipt records the evidence and final outcome.

Measured evidence

Results, not promises.

The routing metrics are scoped to the bounded synthetic competition suite. Deployed success and refusal paths were exercised separately against AWS.

1.0
routing accuracy on 30 synthetic contract cases

Deterministic routing

10 clear, 10 blocked, 10 insufficient-evidence cases. Reproducibility evidence, not real-world accuracy.

0.916 ms
mean in-process perception benchmark

OpenCV perception baseline

300 sequential calls on 240×320 synthetic images on the documented GitHub runner. Not AWS end-to-end latency.

2 paths
deployed AWS terminal outcomes

Success and refusal

✓ VERIFIED_RESOLVED

✕ VERIFICATION_FAILED

Judge path

Evaluate VisionProof in minutes.

The main judge path requires no access code. It replays the verified OpenCV decision loop, including both resolved and failed outcomes.

90-second live workflow

  1. Open the live inspection.
  2. Choose the blocked synthetic fixture.
  3. Observe OpenCV metrics and APPROVAL_REQUIRED.
  4. Approve and run the simulated action.
  5. Verify with a clear follow-up for VERIFIED_RESOLVED.
  6. Open the evidence/decision trace and download the receipt.
Run judge demo

Failure path is real

A separate deployed smoke test uses a blocked follow-up after execution. VisionProof refuses to call the mission resolved.

VISIONPROOF NEGATIVE VERIFICATION SMOKE PASSED
Final outcome: VERIFICATION_FAILED

Execution is not success.

OpenCV → action

The visual result changes the workflow.

Clear / supportedLOW

Occupancy below threshold

Action ready
Blocked / supportedCONSEQUENTIAL

Occupancy at or above threshold

Human approval required
Unsupported evidenceEVIDENCE_INSUFFICIENT

Image quality or evidence support fails

No execute path
Blocked follow-upVERIFICATION_FAILED

Condition remains after simulated action

Do not claim success
Clear follow-upVERIFIED_RESOLVED

Condition is resolved in follow-up evidence

Mission may close