Change the weights today and last quarter's grades change with them.

The mechanism

The stored weights are inputs to a formula, and formulas do not remember when they were first evaluated.

Each week's row holds its own branch probabilities, and those are historical facts that stay put. The blended figure is not stored; it is computed on demand from those probabilities and the current weight profile. Change the profile and every blended figure in the workbook recomputes, including every one that has already been read, acted on and recorded in a review note. The weekly status recomputes with it, so the count of green and yellow and red weeks in the monthly rollup can shift without a single probability having been touched.

Why it is worth caring about

A shifting baseline makes deterioration and improvement equally unreadable.

The purpose of a weekly series is comparison — this week against recent weeks, this month against last. That comparison requires the measuring instrument to be constant over the range being compared. If the weights were adjusted six weeks ago, then the series spans two instruments, and a decline across it might be the trading getting worse or might be the model getting stricter. Nothing in the workbook records which. The most valuable thing a long weekly series contains is trend, and an undated assumption change is the one edit that can remove that value entirely while leaving every cell populated and plausible.

The legitimate reason to edit

Assumptions should change when the strategy changes, and that is a deliberate act with a date attached.

None of this means the profile is frozen. The weights encode the intended branch mix, and when that intent genuinely changes — a deliberate reweighting toward or away from a branch, a revised view of what a runner is worth after protection — the constants must follow or the workbook is measuring a strategy nobody is running. The distinction is between a considered model update and a casual adjustment, and the safeguards are the same ones the promotion workflow applies everywhere else: the change is proposed, tested against evidence rather than intuition, applied deliberately, and recorded with the date it took effect so the series can be read in two segments rather than one misleading one.

• Record the date and the reason at the moment of the edit; reconstructing it later is guesswork.
• Treat the series as segmented at that date rather than continuous across it.
• Verify the weight sum immediately after any edit — a broken profile still grades.

The tempting edit

The most dangerous adjustment is the one made while looking at a disappointing month.

The specific failure this guards against is narrow and recognisable. A month grades poorly, the operator inspects the branches, notices that the underweighted branch was the one performing, and adjusts the weights toward it. The month's grade improves immediately, because the recomputation runs over the same data. Nothing was learned and something was lost: the profile now encodes a mix chosen by a recent sample rather than by strategy, and the record of the poor month has been overwritten by an assumption change made in response to it. If the reweighting is genuinely correct, it is still correct next month after it has been tested somewhere other than the data that suggested it.

Where the boundary is drawn

Evidence lives on the weekly rows; assumptions live on the Legend tab; nothing lives in both.

The tab separation is a permission structure rather than an organisational preference. Weekly probability cells are safe to edit because they are evidence and evidence gets entered. Legend constants are edited only when assumptions are intentionally being changed, and that phrasing is the actual rule — the tab is not protected, it is governed. Everything else in the workbook is formula and should be left alone. An operator who knows which of the three categories a cell belongs to will not make this class of mistake, and one who does not will eventually make it while doing something that felt like ordinary maintenance.

The key idea

Separate what you observed from what you assumed, and date every change to the second.

This is a general property of any workbook that mixes recorded evidence with model parameters, and the weekly scorecard is where it bites hardest because the series is long and the parameters are few. The observations are safe: a probability entered eleven weeks ago says what it said. The assumptions are shared, live, and retroactive. Keeping them on their own surface, editing them rarely, and dating them when you do is the entire discipline, and it costs one line in a log against the alternative of a history that quietly no longer means what it meant.