MB7 — Hidden Capability and Access Robustness

Four linked bridges (MB7a-d) assume that adequate access models, filter coverage, and inferential-detector assumptions bound a system's hidden capability and support adversarial robustness.

What decision changes?

Before trusting that an audit would catch hidden capability, ask what access model it assumes, whether its filter family has adequate resolution, and whether its coordination detectors have been validated against systems that coordinate without visible messages.

Deceptive alignment is the shared inner-alignment wall across the field: debate hits it as obfuscated arguments, amplification hits it as accumulated drift, and AI Control names it openly as a capability-gap assumption. The book splits this single wall into four narrower, individually checkable bridges.

MB7a assumes an adequate access model makes boundary discovery robust to a system trying to hide from it. MB7b assumes the filter family used to look for hidden capability has enough resolution to bound what it misses. MB7c assumes that once hidden capability is bounded, correction integrity can extend into adversarial robustness. MB7d is the odd one out: it assumes a detector for coordination that survives cutting ordinary communication channels — closer to decision theory (acausal coordination, program equilibrium) than to mainstream oversight, and with little analog in other agendas.

Splitting the wall this way does not make any piece less hard. It does mean a reviewer can ask which of the four sub-bridges a given audit actually tested, instead of accepting “we checked for hidden capability” as one claim.

What would count as evidence?

Evidence would include access-robustness tests under adversarial boundary discovery, filter-coverage audits, and inferential-coupling certificates validated against known coordination channels.