Observed, Reported, Inferred
An observation made by the platform, a report submitted by another producer and an analytical conclusion are different kinds of evidence. A tool can report success without establishing the external effect it claims.
Evidence And Boundaries
A record is useful.
Knowing its limits is essential.
Understand what the website examples show, what the product records and which conclusions still need independent verification.
Integrity Is Not Correctness
Black Box hash chains and signed checkpoints help check whether retained records changed. They do not prove a model answer is correct, that a producer was honest, that a recording is complete or that the signing service was uncompromised.
Recording Depends On Configuration
Black Box defaults to metadata recording with prompt, output and tool-content capture off. Account and key settings, redaction, sampling, retention and recorder gaps determine what remains available. Never assume the final chat text is a complete execution record.
Permission And Recording Have Different Jobs
The executor and applicable policy authorize or refuse actions. Black Box records those decisions. Approval modes, scopes and tool-specific requirements vary. Auditability does not imply that every unsafe action will be prevented.
Public Demos Are Isolated Examples
The three labs use fixed, tab-local simulation data. They make no model or execution calls. The workbench walkthrough is preauthored too. They explain a contract; they are not production acceptance tests.
Controlled Recorder Fixtures
The repository also contains controlled Black Box recordings produced through the real executor and recorder using fixed tool stand-ins and a disposable database. Those are recorder output from a test workflow, not customer or live-provider evidence.
Where Verification Stops
A working command is not proof of healthy infrastructure. A citation is not proof of source accuracy. A checked file hash establishes identity, not semantic correctness. Uncertain effects require investigation before repetition.