The five-layer stack: from what happened to why to whether it lasts.

Layer one

CP3 main EV — the truth engine attribution decomposes.

The stack's foundation is deliberately not inside the attribution module at all: Layer 1 is the CP3 main output — branch EV, blended EV, weekly and monthly EV, gate and brake state, the dashboard. The manual's phrasing preserves the hierarchy: this is still the main truth engine. Attribution never competes with it, never recomputes it, and never overrides it; everything above Layer 1 is decomposition of a truth established elsewhere. The design protects both directions — the production record keeps one authoritative source, and the attribution layers are free to slice aggressively because nothing they produce can contradict the record, only explain it.

Layers two and three

Where the R came from, then why it behaved that way.

Layer 2 is variant attribution proper: variant counts, average R by variant, total R by variant, EV share, and the variant-versus-Standard ratio — the accounting of which management structures actually produced the period's result. It answers where, and deliberately stops there. Layer 3 is the explanation underneath: the checkpoint hit-rate battery — 0.75R, 1.25R, 1.3R, 1.4R, 1.6R, 1.8R, 2R-plus flags — that tells you why the Layer 2 numbers came out as they did. The separation is what makes findings actionable. 'Time-Aggressive underperformed Standard this quarter' is a Layer 2 fact with a dozen possible causes; 'because its 1.3R checkpoint hit rate collapsed while 1R held' is a Layer 3 explanation with one — the delayed monetization point stopped being reached, which indicts the conditions the variant was selected in, not the variant's design.

Layer four

Execution quality — whether the variant's result is sustainable.

A variant can show strong attributed R built on unsustainable mechanics, and Layer 4 exists to catch it: MAE, MFE, capture efficiency, giveback, fees, duration, and adherence, sliced by variant. This is where a flattering Layer 2 read gets its physical. An Exposure-Aggressive line printing excellent average R while its giveback trends up and its capture decays is monetizing conditions that are deteriorating underneath it; a Time-Conservative line with modest R but pristine capture and minimal fee drag is a structure performing exactly as designed. Layer 4's question — is the variant quality sustainable? — is the difference between attributing a result and endorsing one, and it's deliberately positioned before the structural layer because mechanics decay faster than regimes shift.

• Net-of-fees R by variant lives here: a variant that wins gross and loses net isn't earning its costs, whatever Layer 2 says.
• Capture and giveback by variant audit whether the structure captures opportunity or leaks it — the manual's own framing.
• Adherence by variant catches the subtle failure: a variant that only wins when its rules are bent isn't the variant winning.

Layer five

Structural context — the layer that decides what it all means.

The stack tops out in the SDE and regime review: the longer-horizon read that decides whether variant behavior is improving, decaying, or unstable. Everything below it is period accounting; Layer 5 is trajectory. Its role in the module is deliberately judicial rather than computational — it consumes the four layers of evidence and returns the only verdicts that should drive change: this variant's conditions have genuinely shifted (a scenario candidate), this variant's execution is slipping (a discipline finding), this variant is fine and the quarter was weather (the most common verdict, and the one the stack exists to make safe to reach). Without Layer 5, attribution findings arrive context-free and every anomaly looks like a mandate. With it, the stack's output is proportionate: most reviews end in documentation, not surgery.

The key idea

Depth discipline is what separates attribution from narrative.

The five layers encode a routing rule: every question has a depth, and answers fetched from the wrong depth are stories. Source questions stop at Layer 2; mechanism questions need Layer 3; endorsement needs Layer 4; meaning needs Layer 5; and nothing anywhere overrides Layer 1. An operator who honors the routing gets findings that survive scrutiny — where, why, whether it lasts, what it means, in that order. One who skips depths gets the classic attribution failure: a quarter's numbers, a plausible story connecting them, and a rule change the next quarter regrets.