The two ways a diagnostic layer gets misread — and they are opposites.

Failure one

Over-trust: reading intermediate diagnostics as instructions.

The SDE's outputs look decisive. A z-scored drift reading with a stated conviction level and a final interpretation line has the surface texture of a directive, and an operator who has stopped distinguishing diagnosis from prescription will start trading it — increasing size when the structural read is constructive, standing down when it is not, effectively promoting the engine into an allocation system it was never built to be. The specific danger is the word intermediate. The SDE's pipeline exposes many values that are inputs to a conclusion rather than the conclusion itself: a single table's rolling value, a stability flag, a drift before its z-score has established whether it matters. Each of those looks like a finding. Acting on them individually means acting on a half-finished computation, which is worse than acting on no computation at all.

Failure two

Under-use: filing the diagnosis and changing nothing.

The second failure is quieter and far more common. The engine runs, the readings are produced, the operator glances at them, and the week proceeds exactly as it would have without them. Nothing is broken; nothing is used. This is what the manual means by a passive report — a component that produces information which never reaches a decision. It usually arrives with a plausible justification: the reading is only advisory, the gate is the real authority, so the diagnosis is interesting rather than actionable. Every clause of that is true and the conclusion does not follow. Advisory does not mean optional. The whole reason the layer exists is that raw weekly PnL, isolated EV, and a hit-rate reading are not enough to see a system weakening beneath a profitable surface, and an unread diagnosis restores exactly the blindness the layer was built to remove.

Why they are mirror images

Both errors come from collapsing posture and permission into one thing.

The over-truster believes a diagnosis that influences posture must therefore grant permission, and so treats a constructive read as authorisation. The under-user believes a diagnosis that cannot grant permission must therefore not influence posture, and so treats every read as decorative. They share a premise — that influence and authority are the same substance — and disagree only about which way to resolve it. This is why an operator who catches themselves over-trusting frequently overcorrects into indifference within a month, and why the correction has to be aimed at the premise rather than at the behaviour. The layer influences what you attend to and how confidently you operate inside a fixed envelope. It never moves the envelope. Both halves of that sentence are load-bearing.

• Over-trust converts a reading into a size change. That is the gate's job, and the SDE has no drawdown input.
• Under-use converts a reading into a filing action. That is a report's job, and the SDE is not a report.
• The correct output of a structural diagnosis is usually a changed review priority, not a changed position.

The practical test

Ask what the reading changed, and whether it was allowed to.

There is a two-part check that catches both failures without requiring any theory. First: after reading the diagnosis, name one thing you will do differently this week. If the honest answer is nothing, and it is nothing again the following week, the layer has become a report and the review has become ceremony. Second: check that the thing you named is inside the layer's remit — a review target, a scrutiny level, a question to open, a metric family to examine first. If what you named is a size change, a gate override, or a branch shutdown, the reading has been promoted past its rung and the correction belongs before the trade, not after it.

The key idea

The useful posture is narrow, and both ditches are wide.

There is a lot of room on either side of correct here, which is why the manual bothers to name the failures rather than simply stating the rule. An engine that gets obeyed is dangerous and an engine that gets filed is wasted, and the interval between them is not intuitive — it has to be held deliberately. What sits in that interval is unglamorous and genuinely valuable: a reading that changes where the operator looks first, and nothing else.