The engine refuses to run on individual branches, and the refusal is the design.

The scope

Three entities in, four out, and the exclusion is not about importance.

The engine runs on the core plan blend, on the trend engine in isolation, and on the total system. Normal, Trend Partial, Trend No-Partial and Overflow — the four branches that carry every trade the system takes — are excluded from it entirely. This is not a statement that branch-level behaviour is uninteresting; branch diagnostics are the whole subject of the structural pipelines and the branch risk engine, both of which run happily on primitives. It is a statement about what this particular test needs in order to mean anything. The contradiction test asks whether two quality ratios agree, and agreement between two noisy estimates is not a detectable property. Running it on a primitive would produce output, and the output would be a measurement of sample size.

Why thin samples break this test specifically

Comparing two unstable estimates produces instability squared.

Most diagnostics degrade gracefully on small samples — the estimate is noisy, the confidence is low, and the reading is treated with appropriate caution. A divergence test degrades differently, because it is a comparison between two quantities that are each independently noisy. Efficiency and profit quality are both ratios computed from the same small set of trades, and each will wander on its own; the difference between two wandering series wanders more than either. The engine then classifies that difference into a named state with an associated severity and a routed review. So a primitive branch with eleven trades in a quarter would not merely produce a weak signal — it would produce a stream of confidently named contradiction states, changing week to week, every one of which is an artefact. Noise that gets classified is worse than noise that stays noise.

What the three composites ask

The exclusions inside each composite carry as much meaning as the inclusions.

The three entities are not a hierarchy of aggregation but three separate questions. The core plan blend combines Normal with both trend branches and deliberately leaves out Overflow, so it reads the behaviour of the plan proper without supplemental flow — which matters because Overflow's volume and costs are the most capable of distorting an efficiency reading. The trend blend isolates the two trend branches, which lets the trend engine be tested for internal coherence separately from the static core that behaves quite differently. All Blended is the total system and is the only one that answers what the account as a whole did. Running the test on all three and comparing the results localises a divergence: a contradiction visible in the total but absent from the core plan blend is a statement about Overflow, arrived at without ever testing Overflow directly.

Localisation without measurement

A branch can be implicated by the composites it does and does not appear in.

This is the design's most useful and least obvious property. Because the three composites overlap in a known pattern, the difference between their readings carries branch-level information that none of them measures directly. A divergence present in the trend blend and in the total but absent from the core plan read points somewhere quite different from one present in the core plan read and absent from the trend blend. The operator is doing inference rather than measurement, and the inference is sound precisely because each composite is large enough to be trusted individually. This is how the engine honours both constraints at once: it never computes a statistic on a sample too small to support it, and it still ends up pointing at particular branches. The pointing is a conclusion drawn across three reliable readings rather than a fourth unreliable one.

The load argument

The second reason is real and secondary.

Reducing formula load is given alongside noise reduction, and it deserves an honest weighting rather than being quietly promoted or dismissed. Each entity block computes a substantial set of functions per week, so running seven entities instead of three would more than double the grid, in a workbook that already carries five six-table pipelines and a full cockpit. That is a genuine engineering constraint. It is not the governing reason, and it should not be presented as one, because if the primitive readings were diagnostically sound the right response would be to pay the cost. The correct statement is that the primitives were excluded because the test would not be valid on them, and the saved computation is a welcome consequence rather than the justification.

• A divergence between two thin-sample ratios measures the sample, not the machine.
• Classified noise is worse than unclassified noise — it arrives wearing a diagnosis.
• Three overlapping composites localise a branch without ever testing one.

The key idea

Choosing what an instrument will not measure is part of building it.

The instinct when constructing a diagnostic is to extend it everywhere it will technically run, on the grounds that more coverage is more information. This engine does the opposite and is stronger for it. Restricting the test to entities where it is statistically defensible means every state it reports can be believed, which in turn means the summary counters that accumulate those states are counting something real. An engine that ran on all seven would produce four extra columns of output, and the presence of those columns would degrade the three that were sound, because nothing on the surface would distinguish the trustworthy readings from the artefacts.