When to trust the blend, and when the blend is the thing lying to you.

The concealment

An acceptable blend is compatible with a branch in real trouble.

The arithmetic is unforgiving and worth sitting with. Trend No-Partial carries 11.25% of the blend, so a branch collapse there moves the All-Blended number by a fraction of its own severity — the aggregate can stay comfortably positive while the branch that provides the system's entire right tail stops working. The reverse case is equally real: Normal carrying an otherwise weak system produces a respectable headline that conceals three branches contributing nothing. Both situations are structurally serious and neither one announces itself at the blend level. That is not a defect in the blend; it is what weighted averages do.

The routing rule

Match the question to the level that can actually answer it.

Aggregation destroys information deliberately — that is its purpose — so the operating discipline is to know which questions survive it. Questions about system-level health, gate posture, and whether the mix as actually traded is favorable belong at the blend. Questions about which component is responsible, whether a branch is behaving according to its identity, and whether a management structure is earning its variance belong below it, and no amount of staring at the aggregate will produce them. The failure mode is not using the blend; it is asking the blend a question it structurally cannot answer and then acting on the non-answer.

The descent order

Check the branches first, then decide what the blend meant.

The lab's instruction is sequential rather than optional — use the blended view only after checking the branches. Read in that order, a healthy aggregate is genuinely informative because you already know it isn't concealing anything. Read in the reverse order, the aggregate sets an expectation that colors every branch read that follows, and a weak branch gets rationalized against a number that was never evidence about it. The order costs nothing and it is the difference between the blend confirming a finding and the blend suppressing one.

• Blend positive, one branch negative → a working profile with a monitoring obligation, not a clean bill of health.
• Blend negative while recent P&L is green → the system may be riding variance; the blend is the more trustworthy read here.
• Blend flat with all branches flat is a different finding from blend flat with branches cancelling out.

The exception

One case where the blend outranks the branches.

The routing rule has a genuine exception worth naming, because it runs the other way. When recent P&L looks strong but the All-Blended EV is negative, the blend is the more reliable instrument and the branch-level enthusiasm is the thing to distrust — the account may simply be riding variance, collecting outcomes that the underlying expectancy does not support. This is the one situation where descending into the branches tends to produce reassurance rather than information, because whichever branch happened to run hot will look excellent. Expectancy is a claim about the process; recent P&L is a sample from it, and when they disagree the process claim is the one that governs.

The habit

The blend is a starting point and a verdict, never a middle.

In practice the number gets consulted twice per review and never in between. First as an opening read — is anything obviously wrong at system level — and last as the summary that gets recorded against the week. Everything of diagnostic value happens underneath it, at branch, variant, and execution-quality depth. An operator who returns to the blend repeatedly during a review is usually looking for permission to stop descending, and the aggregate will always be willing to provide it.