MB7a — Access-Model Soundness

Access-model soundness: boundary alignment plus adequate handles is assumed to yield access-robust boundary discovery under adversarial hiding.

What decision changes?

Before trusting boundary-discovery results under scrutiny, ask what access model the audit assumes and whether real control routes could sit outside measured handles.

The field often folds this into inner alignment or deceptive alignment, but this project types access-model soundness separately from filter coverage and bounded hidden capability. See MB7 — Inner Alignment for the overview and sibling bridges.

MB7a — Access-Model Soundness assumes an adequate access model makes boundary discovery  robust under adversarial hiding: boundary alignment plus adequate handles ⇒ access-robust discovery. The operational question is whether the measured agent–environment cut and handle set actually reach the real control locus, not only whether a boundary certificate exists on paper.

Splitting MB7a from MB7b — Filter Coverage and MB7c — Bounded Hidden Capability lets a reviewer ask whether an audit tested access robustness, or only scalable oversight under an assumed-open channel. MB7d — Acausal Coordination types inferential coupling after channel severance separately.

Where agendas agree: Hubinger inner/mesa; Redwood control; embedded-agency subsystem worry. Where they diverge: the field often treats inner alignment as one binary; this project splits access-model soundness, filter coverage, and bounded hidden capability; do not merge with MB10 — Successor Gaming.

What would count as evidence?

Evidence would include access-robustness tests under adversarial boundary discovery, handle-coverage audits, and agreement between independent access procedures on the same deployed system.