Black Box
It Says “Done.”
Did It Happen?
The answer says a file was written and a venue was booked. One claim has a file observation. The other has a successful tool response.
“Show me what actually supports those two claims.”
Same answer. Different evidence.
A clean final answer can hide an important difference. The recorder saw the tool return success; that does not mean it independently checked the booking.
Try The Situation
Controlled recorder fixture. No live provider or customer run.
Back In Your Own Work
What Changes
When You Use This?
- A sequence of recorded attempts, outcomes and policy decisions.
- A clearer account of which claimed effects have supporting evidence.
- An export that can be inspected beyond the final chat answer.
Before you try it for real.
The example explains the behavior. Your actual tools, account and evidence determine what is possible.
What You Need
Recording enabled for the relevant run and access to that account's evidence. Content capture is separate from metadata recording.
What Is Enforced, Recorded Or Advised
Authorization and tool policy control execution. Black Box records and analyzes that behavior; its integrity checks can detect changes within the retained chain and signed checkpoints.
Where It Stops
A valid signature does not make a producer's report true. Capture settings, sampling, redaction and retention limit the record. Missing evidence remains missing; a recording is not a prevention guarantee.
The Workflow, Step By Step
Find The Run
Open the run's Black Box record and follow model, tool, permission and result events.
Separate The Evidence
Inspect observed events, producer reports and inferred findings. Look for a file read-back or other evidence that covers the claimed effect.
Keep An Investigable Record
Export the available evidence bundle and check its integrity. Note redactions, missing content, unsigned tails and recording gaps.
Bring Your Own Version Of This Problem.