Observation
What is directly visible in the image or video.
Science & evidence
What we can show, and how we know it
This page explains the evidence vocabulary used across the public site and summarizes what the current service pages document.
AI-generated editorial image — temporary placeholder pending licensed field photography.
Observation, inference, and confirmation are not the same
Keeping these three apart is the core of how IntelCit reports results honestly.
Inference
What a model or workflow proposes from the observation, always with uncertainty.
Confirmation
What a person decides after review. The system never confirms itself.
A six-step framework for evaluating a service
These are the checks IntelCit uses as a public evaluation framework. Completion and available evidence vary by service.
Define the task
State exactly what the service is asked to observe and what it is not.
Capture conditions
Record how the input was captured — device, angle, lighting, and setting.
Label and ground truth
Establish what the correct answer is for the test material being used.
Run the model or workflow
Produce the counts, classes, or candidate events the service returns.
Human review
A person checks the output against the source and decides what to accept.
Record limitations and version
Note what failed, what is uncertain, and which version produced the result.
Plain-language metric glossary
The words behind detection results, without the statistics jargon.
| Term | What it means |
|---|---|
| Precision | Of the things the service flagged, how many were actually correct. |
| Recall | Of the things that were really there, how many the service found. |
| False positive | The service flagged something that was not actually there. |
| False negative | The service missed something that was actually there. |
| Confidence vs validated accuracy | A confidence score is the model's own estimate for one output. It is not a validated accuracy figure. |
What has been demonstrated, per service
A registry-driven summary of the current evidence and delivery stage for all five services.
Where results get uncertain
These conditions can make visual analysis less reliable and should be considered during review.
- Lighting — low light, glare, and strong shadows reduce reliability.
- Occlusion — objects hidden behind people, vehicles, or foliage may be missed.
- Motion — fast movement and motion blur weaken detection and tracking.
- Angle — steep or unusual camera angles change what is visible.
- Season and weather — rain, snow, wet surfaces, and seasonal cover change appearance.
- Very small objects — small or distant items are the hardest to detect.
Boundaries we keep
These limits hold everywhere IntelCit is described publicly.
- CAI is IntelCit's internal cleanliness scale — 1 is nearly clean and 4 is very dirty. It is not a validated or certified standard, and it is not a GUQS mapping.
- A vehicle-littering output is a candidate event for mandatory human review. It is never an offence, fine, citation, or automatic enforcement decision, and it stays incomplete_for_enforcement.
- IntelCit does not publish accuracy percentages, ROI, savings, city or customer case studies, certifications, or unsupported sample sizes.
Versioning and human review
Where version information is available, evidence should identify it. The public service pages do not yet publish a complete model and workflow version history.
Available version information should stay beside the evidence it produced; missing version information remains a documented limitation.
Human review is a required step, not an optional add-on. IntelCit supports a decision; it does not replace one.
Evaluate the evidence with your own material
A guided demo can walk through the method and the current evidence for the service most relevant to your team.