What an audit log for robot decisions should contain

Most robot cells keep a log of what happened. Few keep a log of why. When an auditor, a customer or your own quality team asks about one item handled six weeks ago, “the robot placed it in bin B” is not an answer. This is the record Sixth Sense writes for every decision, and the reasoning behind each field.

One record per decision

• Time and place. Timestamp, cell, lane, camera. Enough to find the moment on any other system.
• The item. The identifier, if one was read: batch, serial, order line. If none was read, the record says so, because that is often the reason for the decision.
• What was seen. The structured facts perception produced: item class, state, position, and the image reference. Not the model’s opinion. The facts.
• What was allowed. The list of actions the rule engine permitted, and the rule versions that produced it.
• What was chosen. The action, the confidence, the threshold in force and the model version.
• What happened. Acted, asked, or blocked. If asked, who answered, what they chose, and how long it took. If blocked, which rule blocked it.

Six groups, one line per decision in the export. Everything a later question needs is in that line or one reference away.

Why the facts and the choice are separate

A log that stores only the decision cannot tell you whether the model was wrong or the camera was. Storing the facts separately lets you replay a decision: given what was seen, would the current rules and model choose differently? That replay is how rule changes are tested and how calibration is checked, and it only works if the input is preserved.

What the log is not

It is not a video archive. Images are kept by reference, for a retention period you set, and they never leave the site. It is not a performance ranking of operators. Who answered is recorded because accountability requires it, and access to that field can be limited to the roles that need it.

Export and ownership

The log is exported as CSV and JSON, on demand or on a schedule, into a folder or a system you own. The format is documented and stable. If you stop using Sixth Sense, the log stays readable without it. A vendor who cannot promise that is asking you to trust a black box with your own records.

Three questions to test a log with

Before accepting any decision log as adequate, try to answer these from the export alone.

1. For this item, what did the system see, what could it have done, and why did it do this?
2. Which rule version and which model version were active at that moment, and who confirmed them?
3. How many items in this batch were handled by a person rather than the robot, and how long did those take?

If any answer requires opening a second tool or asking the vendor, the log is incomplete.