Platform
The audit trail nobody asks for until they do
When someone asks what happened and who signed off, the answer should already be written down. What to log, and what to keep out of it.
AI in the real worldKey takeaways
- Log the decision and its author, not just the detection.
- Write down what was suppressed as carefully as what was raised.
- Keep faces, plates and names out of anything you are not prepared to defend holding.
- A trail that cannot be exported is a trail that does not exist when it is needed.
Nobody asks for the audit trail until the day somebody asks for it, and on that day it is either already written or it is not. There is no version of this where it gets assembled afterwards.
What to log
- The detection: camera, time, model and version, score.
- The decision: who acknowledged it, when, and what they closed it as.
- The configuration at that moment, the zone and rule as they were, not as they are now.
- Every change to that configuration, with an author and a timestamp.
The third one is the one that gets missed and the one that matters. An incident reviewed against today's rules, six weeks after the rules changed, is being judged by a standard that did not exist when it happened.
Especially the suppressions
A system that quietly drops detections and only records what it showed you is a system that cannot answer the one question an investigation actually asks: was there anything you did not tell us about? Log suppressions with the rule that caused them.
The useful question is never 'what did it find'. It is 'what did it see and decide not to mention'.
What to keep out of it
Retention is a liability as much as an asset. Identity should live in the log only where there is a stated reason to hold it and a date on which it leaves, and the default for everything else, faces, plates, names, is that the trail records that a person was present, not which person.
Frequently asked
How long should incident records be kept?
As long as the policy you can point at says, and not a day longer by accident. The failure mode is almost always indefinite retention nobody chose.
Does the model version really need to be in every row?
Yes. Without it, a review across six months is comparing decisions made by different systems and calling the difference a trend.
Keep reading
All articles
Platform
Does a model trained on Terra hold up on your cameras?
The gap between a generated scene and a wet lens at four in the morning, measured rather than argued about, and what closes it.
Read article
Deployment
What a hundred cameras cost your uplink
Inference at the edge, clips to the centre. A worked bandwidth budget for a single site, and the two settings that halve it without losing evidence.
Read article
Field notes
The labelling bill nobody budgets for
Annotation is the line item that arrives after the model is approved. Where the hours actually go, and which of them a generator can take back.
Read article
