Never average a ratio: the rule that keeps composites honest.

The rule

Composite ratios come from composite totals.

The manual is explicit: composite values should come from valid upstream composite logic, or be recomputed from the correct underlying branch totals, and for ratio metrics such as RAER and RAPF the primitive ratios must not be averaged blindly. Composite RAER is composite net PnL divided by composite risk deployed. Composite RAPF follows its own defined profit-factor and risk-adjustment logic against composite inputs. The distinction matters because a ratio carries a denominator, and the denominator is a weight. Averaging four ratios treats four branches as equally important regardless of how much risk each one actually deployed, which is exactly the information the ratio was constructed to carry.

Why the error runs one way

The distortion is not random — it usually flatters.

If the error were symmetrical it would be a nuisance. It is not. In practice the branches deploying the most risk are the ones whose efficiency matters most and, when a system is under structural stress, they are frequently the ones whose efficiency has degraded — a high-variance branch running its full allocation through a difficult period. A small branch running modest risk with a strong ratio contributes a large upward pull to a simple average and a small one to a properly weighted composite. The averaged figure therefore tends to report a healthier composite than the account actually experienced, and it does so most reliably in exactly the conditions where an accurate reading matters. A metric that is most wrong when it is most needed is worse than no metric.

• Weighting by risk deployed is not an adjustment to the ratio. It is what the ratio means.
• The bias direction is toward optimism, because efficient small branches are over-represented in a flat average.
• The distortion grows with the spread of risk deployment across branches, which widens under stress.

Where it enters

The error arrives through convenience, not through ignorance.

Nobody sets out to average ratios. It happens because the branch ratios are already computed and sitting in adjacent cells, and forming a composite from them takes one formula while recomputing from totals requires reaching back to composite net PnL and composite risk deployed. The shortcut is available, plausible, and produces a number in the expected range. This is why the rule is stated as a formula-integrity rule rather than left to judgement: it is not the kind of mistake an operator catches by looking at the output, because the output looks entirely normal. It has to be prevented at construction, and any composite ratio whose provenance is unclear should be treated as suspect until traced.

The downstream stake

The Contradiction Engine reads composites exclusively.

This rule has a consequence beyond tidiness. The RAER/RAPF Contradiction Engine is deliberately composite-only — it runs on Normal + Trend, Trend Blend, and All Blended, and excludes primitive branches entirely. Every contradiction state it publishes therefore rests on composite RAER and composite RAPF being correctly formed. If either was assembled by averaging, the engine is comparing two numbers that do not mean what their labels claim, and the state it produces — alignment, under-monetisation, profit without efficiency, dual weakness — is not a diagnosis of the system but an artefact of the arithmetic. The composite ratio rule is not a detail inside the SDE. It is a precondition for an entire downstream engine.

The key idea

Ratios do not add, and a system that forgets it stops measuring itself.

The arithmetic here is elementary and the consequence is not. A composite is meant to describe what the blended machine actually did, which means it has to be weighted by how much each branch actually did. Averaging replaces that with a description of a hypothetical system in which every branch mattered equally, and no such system was ever traded.