What the system
can do.

Govern authority

Classify and bound permission

Human-only, sealed, safety-only and bounded machine authority are represented as different operating states across 63 surfaces.

Dated evidence

63 classified authority surfaces

Claim boundary

Classification does not imply that every surface was exercised or attributed in the latest window.

Reproduce decisions

Re-run meaningful choices

Recorded inputs are used to re-execute selected decisions away from the live process and compare the resulting action.

Dated evidence

821 / 821 meaningful decisions matched

Claim boundary

The result is system-, window- and environment-dependent rather than a fleet-wide equivalence proof.

Preserve veto evidence

Keep refusals inspectable

Rejected actions remain part of the evidence surface. A veto is treated as a decision that should be reproducible, not as missing output.

Dated evidence

115 / 115 covered vetoes matched

Claim boundary

Replay covers recorded vetoes inside the selected observation window.

Observe without actuating

Separate visibility from control

The primary observer is designed to inspect the fleet without inheriting a trading actuator. Fallback preserves visibility under a weaker claim.

Dated evidence

Primary observer + declared fallback mode

Claim boundary

The latest cycle used the fallback and therefore does not support observer independence.

Verify effects over time

Distinguish adoption from closure

A remediation is not credited when code merely lands. Its preregistered effect must appear and persist in the measured system.

Dated evidence

14 adopted · 2 effect-verified · 2 reopened

Claim boundary

Unverified and reopened work remains outside the closure count.