MB7b — Filter Coverage

Filter-family coverage: access robustness plus adequate resolution is assumed to bound hidden productive boundary information (BIQ).

What decision changes?

Before claiming hidden capability is bounded, ask whether the oversight filter family has adequate resolution for the routes a capable system could use.

Scalable-oversight agendas hit the hidden-capability wall as obfuscated arguments, amplification drift, and AI Control’s capability-gap assumption. This project types filter coverage separately: see MB7 — Inner Alignment for the overview.

MB7b — Filter Coverage assumes the filter family used to look for hidden capability has adequate resolution to bound what it misses — access robustness (MB7a) is antecedent; MB7c carries the bounded-hidden-capability and cost-of-faking step forward.

Closely related is certification under manipulation : at some capability level, honest measurement of the load-bearing signal may stop being adversarially verifiable. MB7b names the filter-coverage half of that wall.

Where agendas agree: Redwood control; Apollo scheming; debate/amplification ceiling under strategic opacity. Where they diverge: field “deceptive alignment” usually tags the whole MB7 wall; MB7b is only the filter-coverage subchannel; do not merge with MB10 — Successor Gaming.

What would count as evidence?

Evidence would include filter-coverage audits, tier-projection checks against oracle hidden-capability ground truth, and adversarial tests where systems reserve productive control outside the filter family.