Changing the weights: the slowest decision in the system, deliberately.

The instrument

Scenario_Profiles compares alternates against the active weights.

The question 'was another blend structurally superior?' is answerable rather than rhetorical. Scenario_Profiles takes the same branch populations and reweights them under alternate profiles, producing what the period would have looked like under a different declared mix. This is the honest form of the question, because it holds the trade data fixed and varies only the thing under consideration. It is also the reason weight changes do not need to be argued from intuition — the comparison exists, it runs on real branch outcomes, and it can be consulted before anyone commits to a view.

The persistence rule

One flattering sample is the least qualified evidence there is.

Profile changes require persistence over multiple weeks and months, never one flattering sample — and the rule targets a specific statistical trap. The period that most strongly argues for increasing a branch's weight is, by construction, the period in which that branch performed unusually well. That is the definition of the evidence, and it is also the definition of the sample most likely to be unrepresentative. For the fat-tail branch the problem is acute: Trend No-Partial's realized R over any short window is dominated by whether its outliers happened to land inside the window, which makes short-window evidence close to worthless for weight decisions specifically.

The cadence

Weights move at scheduled review and never mid-week.

Variant attribution runs continuously, but weight changes are held to the review calendar — a deliberate lag whose entire purpose is to keep allocation decisions out of reach of a hot Tuesday or a bruising Thursday. The cost is a little responsiveness. What it buys is that every weight change has a full settled period of evidence behind it, and that no allocation decision was ever made while the operator was inside the emotional aftermath of a specific outcome. A system that permitted mid-week reweighting would find its weights tracking recent P&L, which is precisely the behavior the profile exists to prevent.

• Attribution is continuous; allocation is periodic. The asymmetry is the control.
• A change proposed mid-week is not rejected — it is queued to the review, which is usually enough.
• The active profile is a declared object in the Legend; changing it is an edit with a record, not a drift.

The asymmetry

Reducing a weight and raising one are not mirror decisions.

The governance is intentionally lopsided, matching the wider doctrine that aggressive moves must be harder to qualify than defensive ones. Reducing a branch's weight moves the system toward less variance and can be justified on structural deterioration alone — the SDE lens showing decay is sufficient grounds. Raising a weight moves the system toward more variance and additionally has to clear the persistence bar, the scenario comparison, and a check that the branch's execution quality supports the increase rather than just its headline R. An operator can always become more conservative on evidence that would not license becoming more aggressive.

What a change actually commits to

The profile is a forecast the system will be held to.

Because the All-Blended EV is computed under the active weights, changing them changes the meaning of the system's master health number going forward — and it resets the baseline that subsequent evidence gets measured against. The profile is not a description of what the operator has been doing; it is a declaration of the mix the system expects and will police through quota. Raise Trend No-Partial's weight and the quota floor rises with it, which means the operator has committed to finding those setups, not merely to counting them more heavily when they appear. Weight changes are obligations as much as they are allocations.