Composite-only: the scope decision that costs the engine its aim.

The decision

Primitive branches are excluded to reduce noise and formula load.

The SDE's entity universe holds seven — four primitive branches and three composites — and the contradiction engine deliberately uses only the composites. Two reasons are given. Low-frequency noise: a single branch may produce few enough trades in a week that its RAER and RAPF cumulative values move on small samples, and a contradiction derived from two noisy readings is mostly an artefact of sample size. Formula load: the engine computes fourteen functions per entity per week, and extending from three entities to seven more than doubles a calculation grid that already runs weeks vertically against entity blocks horizontally, inside a workbook with a real size constraint. Neither reason is a rationalisation. Both are the sort of constraint that produces a better instrument by being respected.

The cost

A contradiction with no address.

When All Blended reports profit without efficiency, the engine has established that across the whole system, profit quality is outpacing risk conversion. It has not established whether that is true of every branch, of one branch, or of a mix in which two branches pull in opposite directions and one dominates the blend. The finding is real and its location is unknown, which means the operator receives a genuine signal and an unbounded search. This is the honest description of the tradeoff, and it should be stated plainly rather than smoothed over — an engine that finds problems without locating them is more useful than one that finds nothing, and less useful than one that does both.

• A composite contradiction can be produced by one branch or by a shift in the mix between branches.
• The engine cannot distinguish those cases, because it never reads a branch.
• This makes every contradiction the beginning of an investigation rather than the end of one.

The recovery

Three composites give the engine partial aim after all.

The cost is real and it is not total, because the three composites are defined by their exclusions and the differences between them carry locating information. If All Blended contradicts while Normal + Trend does not, the difference between the two entities is Overflow, and Overflow is where the enquiry starts. If Normal + Trend contradicts while Trend Blend reads cleanly, the trend engine is behaving and the finding sits with Normal or with the mix. This is triangulation rather than measurement — it narrows the field without naming the branch — and it recovers a useful portion of what the scope decision gave away. Reading the three composites against each other is the technique the exclusion makes necessary.

Where the branch answer lives

The SDE pipelines run on all seven, and that is the right division of labour.

The exclusion is less severe than it first appears because the branch-level analysis exists elsewhere. The SDE's own pipelines instantiate against all seven entities, so branch-level RAER and RAPF are both available — they simply are not compared to each other by an automated engine at branch granularity. An operator holding a composite contradiction has the branch data to interrogate it manually, which is the appropriate amount of friction for an enquiry that should be deliberate rather than routine. The engine automates the check that should run every week on every composite, and leaves the branch investigation to be opened when a composite check gives cause.

The key idea

Scope decisions should be published, not discovered.

The most damaging version of this design would be an engine that quietly ran on composites while presenting itself as reading the system, leaving an operator to slowly work out why its findings never resolved to a branch. Stating the scope — three composites, four branches excluded, for these two reasons — converts a limitation into a known property of the instrument, and a known limitation can be worked around.