Institutional adaptation × Alignment/assurance

The warning works; the building can't move. Audit logs pile up as archives of known, unremedied risk — and worse, they become discovery material: once a court can subpoena the log proving the institution knew, corrigibility itself becomes a liability, so slow institutions rationally prefer plausibly-deniable opacity instead. Assurance decays into a compliance exhibit while vendors patch faster than approval processes can move. The evidence and rollback capacity do survive, though, for whoever eventually reforms the institution — but this corner doesn't drift gently: it stores pressure and snaps, often governed by whoever built the unofficial workaround in the meantime.

Its naive self-description — failures become patches — undersells its own danger: seamless correction is a steering wheel, and in a fast-moving institution lots of hands reach for it. "Corrigible to whom" becomes the central political fight, and the corner that does exactly what its controller intends is equally the best case for good governance and for competent tyranny. Reorganizing around what the audit can see also deletes whatever resisted formalization — discretion, mercy, informal slack — producing confident, high-assurance alignment to an impoverished version of what people actually wanted. This corner's risks are generated by its own virtues.

Nobody understands the system, so everyone writes a procedure instead. Institutions wrap brittle, uninspectable tools in committees, bans, and vendor assurances that substitute reputation for understanding — while real usage routes around the ban into unaudited personal and shadow use, so official adoption stats understate real exposure. It's the default failure mode and the one closest to precedent (Horizon, legacy COBOL, insurance-exclusion boilerplate): inertia functions as accidental containment, a boring, grinding dystopia rather than an explosive one — until a crisis forces sudden reliance on the very systems nobody characterized.

The institution can turn, but can't see the road — and turning fast is the actual danger, not a mitigant. Institutions restructure around an opaque system so thoroughly that the pre-AI competencies needed to run without it decay past recovery; the off-switch survives, but nobody remembers how to use it and the people who did have been reorganized away. Each reorganization also resets the track record, so trust can't accumulate no matter how long the system runs, and fast recovery from frequent small failures builds false confidence right up to the tail failure that exceeds it. Enough visible fires do eventually force real assurance — but tools built that way, under pressure, tend to be shaped for liability shielding rather than genuine inspection.