The engine reads rather than recomputes — except where reading would be wrong.

The consumption boundary

The engine adds a layer of meaning and no new arithmetic about what happened.

The layer map is explicit about direction. The journal is the raw evidence layer and the SDE does not alter it. The weekly summary and scorecard calculate branch statistics, probabilities, expectancy and outcomes, and the SDE uses those outputs as source truth. What the engine contributes is structural interpretation on top — rolling condition, stability, drift, normalisation and meaning — none of which restates a fact the upstream layers already established. The reason for the boundary is the one that runs through the whole workbook: a figure computed in two places has two values the moment either formula is edited, and nothing can arbitrate between them because both are marked as calculated. Reading rather than recomputing means every metric in the system traces to exactly one origin, and a wrong number resolves to one place rather than to an argument.

The exception

A ratio of sums is not the sum of ratios, and the gap is not small.

Composite entities break the simple reading rule, because a composite's efficiency ratio is not a quantity that exists anywhere upstream to be read. The instruction is specific: do not average primitive ratios blindly; composite efficiency is composite net profit divided by composite risk deployed, and composite profit factor follows its own defined risk-adjustment logic rather than a mean. The arithmetic reason is that averaging discards the weights. Four branches with wildly different risk deployed contribute equally to a simple mean and proportionally to a correctly composed ratio, so the mean systematically overweights whichever branch traded least. In a system whose branches are deliberately unequal in frequency — Overflow is supplemental, Normal carries the highest count — that is not a rounding difference. It is a composite figure that describes a portfolio nobody is running.

Which way the error runs

The mistake flatters, which is why it survives review.

The distortion is not random in direction. Low-frequency branches tend to produce more extreme ratio readings in both directions, because a ratio computed over few observations is unstable, and the branch most likely to show a striking efficiency figure is the one with the least risk deployed behind it. A simple mean lifts that striking figure to full voting weight. The result is that the averaged composite is usually more favourable and always more volatile than the composed one, and it moves for reasons that have nothing to do with the system's actual behaviour — a quiet month in a small branch can swing it. Because the number looks plausible, sits in the right cell, and moves in ways that can be narrated after the fact, it is unlikely to be caught by inspection. It is caught by construction or not at all.

The three composites

Each composite is a question, and the exclusions are the content.

The composite entities are not three levels of aggregation but three deliberately different questions, and the correct composition matters differently in each. Normal plus Trend combines the three core branches and excludes Overflow, which reads core plan behaviour without supplemental flow distorting it — and Overflow is precisely the branch whose volume and costs can distort expectancy, so its exclusion is the whole point of that entity existing. Trend Blend isolates the trend engine by combining the two trend branches. All Blended is the total system view and is the only composite that answers what the account as a whole did. Reading them as a hierarchy where the largest is the most authoritative misses the design: All Blended is the most complete and the least diagnostic, because it is the one in which a localised failure is most thoroughly diluted.

In practice

Two questions settle whether a composite figure can be trusted.

The operator-level version of this doctrine is short. First, does this composite figure come from valid upstream composite logic, or was it assembled here? If it was assembled here, was it composed from underlying totals rather than from the branch ratios? Second, for any ratio metric, does the denominator represent the risk actually deployed across the whole composite rather than a count of branches? Both questions are answerable in a few seconds and both are the kind of check that only gets made if it is written down, because the failure mode produces no symptom. There is no cell that turns red, no window that stays blank, and no divergence between two displayed numbers — just one number that has been quietly describing a different portfolio than the one being traded.

• Read what upstream computed; recompose only what upstream never computed.
• Any composite ratio: rebuild from composite totals, never from branch ratios.
• The averaging error flatters and destabilises at once — it will not look wrong.

The key idea

The boundary is about origin, not about effort.

It would be simpler to state the rule as never calculate anything twice, and that version breaks immediately on composites, which have to be calculated somewhere and have no upstream home. The real rule is that every figure has exactly one origin and that origin is the layer where the figure first becomes definable. Branch expectancy is definable at the scorecard, so the engine reads it. Composite efficiency is not definable until the composite is defined, so the engine builds it — once, from totals, in a stated way. Getting that distinction right is what allows a diagnostic layer to sit on top of a calculation layer without the two ever disagreeing about what happened.