The usage audit: EV share, the Standard ratio, and overuse detection.

The count read

Variant distribution — the first overuse detector.

The simplest rollup is the most protective: variant count by week and branch, which the manual assigns a pointed purpose — showing variant distribution and whether the operator is overusing a variant. Selection doctrine implies a shape: Standard should dominate as the default, the aggressive variants should be occasional and evidence-gated, the defensive ones situational. A distribution drifting off that shape is a finding before any performance number is consulted — Time-Aggressive climbing toward a plurality means 'exceptional conditions' are being diagnosed weekly, which means they aren't exceptional diagnoses anymore; Time-Conservative dominating a trending quarter means the fat tail is being systematically declined. The failure-modes table names both drifts explicitly, and the count read is where they surface first.

The share read

EV share and total R — where the period's result actually came from.

The next layer weighs the counts: total R by variant and the variant EV share — each structure's realized R as a fraction of the branch or system total. This is the read that separates activity from contribution, and its interesting findings are the disproportions. A variant with 8% of usage driving 40% of realized R is either a genuinely superior structure for current conditions or a small-sample outlier wearing one's costume — the statistical battery's habits apply directly. A variant with heavy usage and negligible share is consuming selection slots, attention, and risk for nothing, which is a quieter problem than losing but the same decision error. And a period where one variant's share is effectively the whole result is concentration worth knowing about, whatever the total says: the week worked because one bet worked, and the record should say so.

The control-group ratio

Every variant versus its own branch's Standard, same period.

The audit's sharpest instrument is the variant-versus-Standard ratio: a variant's average R divided by the same branch's Standard average R over the same window — the manual's designated comparison of experimental management to baseline. The construction quietly handles the confounders that wreck naive variant comparisons: same branch (so exit-architecture differences don't masquerade as skill), same period (so regime doesn't), and a ratio (so absolute market generosity cancels). What remains is close to the question that matters: did deviating from baseline, in this direction, pay? Persistent ratios above one earn a variant wider — still gated — use; persistent ratios below one revoke the claim that the deviation is worth its complexity; and ratios that swing with conditions are themselves the finding, mapping which environments each deviation actually serves.

• Net-of-fees ratios are the honest ones — aggressive variants' longer holds carry costs the gross ratio hides.
• The ratio is a per-branch verdict: Time-Aggressive can earn its keep on Trend Partial while failing it on Normal.
• Sample patience scales with rarity: the quota-limited structures get directional notes for quarters before they get ratios worth acting on.

The selection-behavior mirror

The audit's real subject is the operator.

Read together, the three instruments audit something no other tab does: the selection process itself. The counts reveal what the operator's revealed preferences are, as opposed to the doctrine's stated ones. The shares reveal whether those preferences are earning. The ratios reveal whether each deviation from default justifies existing. And the variant-note requirement — a written reason for every exception selection — supplies the qualitative record the numbers get checked against: a quarter of Time-Aggressive notes citing 'clean expansion' can be laid beside the 1.3R hit rates those weeks actually produced. This is the discretionary-quant bargain enforced at its most granular: the operator chooses freely, and the choosing is measured exactly as seriously as the trades.

The key idea

Unaudited discretion drifts; audited discretion learns.

Variant selection is the most frequently exercised discretion in the system — a choice per trade, sixteen times a week — which makes it the fastest channel for drift and the richest channel for learning, depending entirely on whether it's measured. The usage audit makes it measured: distribution against doctrine, contribution against activity, deviation against baseline, stated reasons against realized rungs. What the operator gets back is rare and valuable — a quantified portrait of their own management judgment, updated monthly, that turns 'I feel like the aggressive variants have been working' into a number with a sample size. That's the module's closing gift: it doesn't just attribute the system's results. It attributes yours.