The system measures whether your discretion helps or costs.

The question nobody asks

Discretion is permitted everywhere and audited almost nowhere.

Systems that allow manual intervention almost always treat it as a safety valve whose value is self-evident — the operator sees something the rules cannot, acts, and the action is presumed helpful because it was made by a person with information. The presumption is testable and rarely tested. Each override is a decision with a counterfactual attached: what the governed path would have produced had the rule been followed. Accumulate those comparisons and discretion stops being a philosophy and becomes a measured contribution, positive or negative, with a magnitude.

The two halves

Alpha and damage are tracked separately, and both are always present.

The table does not produce a single verdict on discretion, and the refusal to average is the point. Some overrides help — genuine exceptions where a rule's general case fitted the specific situation badly. Others hurt, and they are usually the ones taken under pressure to recover or to press an advantage. Reporting only the net would let a run of costly interventions hide behind a few excellent ones, and would obscure the pattern that actually matters: whether the helpful overrides share identifiable conditions that could be written into a rule, and whether the harmful ones share a mood.

The frequency signal

Contribution and usage rate move in opposite directions.

The characteristic finding is that override value is inversely related to override frequency, and the mechanism is not mysterious. An override used a handful of times a year is being reserved for genuinely unusual situations, which is the condition under which operator information plausibly exceeds the rule's. Used weekly, it is no longer selecting for unusual situations — it has become a parallel risk policy, applied on judgement, with no evidence base and no benchmark. The instrumentation elsewhere in the stack agrees: the override stress table exists explicitly to stop full-tier forcing from becoming habitual, and the standard's guidance is that it be used only with an explicit written reason while smart capacity remains the default.

The companion measurement

The system also audits itself, on the same terms.

The framework applies the identical test to its own automation. A separate table compares performance with the throttle engaged against performance without it, asking directly whether adaptive risk management adds value or merely restricts. That symmetry is what makes the override audit fair rather than moralising: the machinery is not exempt from the question it puts to the operator. Both instruments can return unwelcome answers, and a framework willing to compute whether its own governance layer earns its keep has standing to ask whether discretionary intervention does.

What a negative reading means

Not that the operator is bad — that the rule was better than expected.

A negative override contribution is easy to read as a personal verdict and is more useful read as information about the rule. It says the governed path, followed mechanically, would have outperformed the intervened one over this sample — which is the strongest available evidence that the rule is capturing something the operator's in-the-moment judgement is not. The correct response is neither self-criticism nor rule-loosening. It is narrowing the conditions under which override is considered, and where a class of overrides did show positive contribution, testing that class as a candidate rule through the sandbox rather than continuing to apply it by feel.

• Overrides are compared against the counterfactual of having followed the rule.
• Alpha and damage are reported separately so a few good calls cannot mask many costly ones.
• A helpful override pattern is a rule candidate, not a licence to keep improvising.

The key idea

The exception clause is the part of a system most likely to eat it.

Governance frameworks rarely fail at their rules. They fail at the discretion permitted around the rules, which expands quietly because each individual use feels justified and nothing counts them. Measuring override contribution against its counterfactual is what converts that expansion from a matter of self-perception into a number that can be looked at once a month. Most operators have never seen the number for their own trading. Having it, and being willing to read it, is a meaningful part of what the simulation layer is for.