Architectures for Auditable Foreclosure asks how a physical anomaly detector can make its own blindness visible.
The document withdraws the earlier phrase “non-foreclosing classifier.” Classification necessarily maps distinct inputs into fewer output states. The achievable target is not absence of foreclosure but a system in which information loss can be measured, reviewed, and contested.
Five features compose the architecture. A separately calibrated abstention or noncoverage channel allows the system to decline ordinary classification within declared evidence limits. Cross-representation disagreement preservation saves events whose anomaly rankings diverge across object, image, and detector-channel representations. A prospective replay bank preserves anchor events and compatible emulators across generations. A per-stage retention map records what every gate discards. Audited noncoverage estimation becomes a first-class output rather than a residual uncertainty.
The document proposes three deployment levels: an offline and emulation study achievable near term; a replay-bank commitment across future runs; and a longer-term three-tier system integrating plural representations, preservation streams, and audited uncertainty.
Its closing requirement is architectural confession. An instrument does not become open-world by adding an “unknown” label. It becomes more responsible when it states what its representations cannot preserve and creates routes for disagreement to survive.