The governance layer is required to justify itself, on the same evidence as everything else.

The claim under test

Governance is not free, and its cost is easy to state.

Adaptive deployment gives things up in exchange for survival. It caps tiers in defensive gates, which slows recovery precisely when recovery would be most valuable. It compresses fresh deployment when open exposure is carried, which reduces throughput. It forgoes the upper tiers in exactly the situations where an unconstrained account would press hardest. Every one of those is a measurable subtraction from returns in the futures where nothing went badly wrong, and the framework's entire position is that the subtraction is worth what it buys. That is a claim, and claims of that size get evidence.

The comparison

Same paths, same evidence, one mechanism removed.

The test is a controlled one in the sense the scenario doctrine requires: identical assumptions, identical evidence, with adaptive deployment as the single variable. What emerges is not a single number but a pair of profiles that usually differ in a characteristic way — the ungoverned configuration showing stronger results in its favourable half and materially worse outcomes in its adverse tail, particularly in lock rate. Reading only the headline return would make governance look like a cost. Reading the distribution shows what the cost purchased, which is why the survival statistics belong beside the growth ones rather than in an appendix.

The efficiency reading

Value-add has two failure directions, and both are instrumented.

Alongside the with-versus-without comparison sits a measurement of how well deployed risk actually converts — return per unit of deployed pool and per unit of active risk, with explicit penalties for under-deployment and over-deployment, resolving into a value-add band. The two penalties are what make the reading honest. A throttle that restricts unnecessarily is failing just as surely as one that permits too much, and a framework that only instrumented the second would be measuring its own caution as a virtue by construction. Restricting is the mechanism's characteristic failure mode, so instrumenting it is where the audit earns its credibility.

Why the audit matters more than the answer

The willingness to compute it is what makes the operator audit fair.

The framework asks the operator to accept an uncomfortable measurement: whether their discretionary overrides help or cost, against the counterfactual of having followed the rule. That request has moral standing only if the machinery submits to the equivalent question, and the with-and-without comparison is that submission. It also disarms the most common objection to governed systems, which is that the constraints are asserted rather than demonstrated. They are demonstrable, the demonstration ships in the workbook, and it could in principle come back negative.

What a poor reading routes to

A candidate change through the gauntlet — never a disengagement.

If the value-add reading is weak, the finding is that this throttle configuration is not converting well, and it enters the research process like any other finding: a scenario contract with stated expected behaviour, a sandbox run, a persistence window, a documented verdict, a shadow period. What it does not authorise is operating without the mechanism while the question is open. The distinction is the same one that governs every other adverse reading in the system — an instrument reporting that a component underperforms is a reason to test a replacement, not a reason to remove the component and trade unprotected during the investigation.

• The comparison holds evidence and paths constant and removes only the mechanism.
• Under-deployment carries an explicit penalty, so the audit can catch excessive caution.
• A weak reading is a candidate change with a full paper trail, never a switch to flip.

The key idea

A framework that cannot fail its own test is not making a claim.

Risk governance is usually defended by argument — the reasoning about recovery arithmetic and sequence risk is sound, and the conclusion is treated as following. The simulation layer takes the less comfortable route of computing it, on the same paths, with the mechanism switched off, and reporting what the difference actually was. That the answer has consistently supported adaptive deployment is worth something. That the framework built the instrument capable of contradicting it is worth more, and it is the same property that makes every other measurement in the stack credible.