The incident archive
Fighter passed its geometry checks and still fell short of the product bar. The factory had measured whether the interface fitted. It had not measured whether the interface worked as a composition.
Method: the incident archive publishes reviewed lessons from project work. Private repository records are not reproduced.
The incident archive contains one published incident. The pattern register below collects risks in the reviewed project work. These are controls to keep, not six more incident reports.
AI and factory failure patterns
The examples come from recurring project risks. Each one needs a control that survives the next run.
Leave unknowns visible. Require a source for each factual claim and keep expansion behind the parity gate.
Tests can miss hierarchy, pacing and play quality. Inspect the screen in the real flow and intended viewport.
Bind a result to the exact source or release, then repeat the check after a change.
Record what the denominator includes and when it changes. Name the work that remains.
Software checks cannot prove the device, environment or human path. Observe each physical boundary directly.
Label promotional art, show the product state and next proof, and remove claims the source cannot support.