A capture error is invisible to every stage after it.

Two different failures

Arithmetic errors announce themselves; capture errors do not.

A formula error produces symptoms — a blank, a negative where none is possible, a total that will not reconcile — and the architecture is dense with cross-checks that surface exactly this class of problem. A capture error produces nothing. A trade logged to the wrong branch, a stop distance recorded as intended rather than as filled, an MAE entered from memory on Saturday: each of these flows through aggregation, expectancy, gate classification and benchmark comparison without triggering a single check, because every one of those stages is verifying that the numbers are consistent with each other, not that they are consistent with what happened.

Why the pipeline cannot help

Consistency checks confirm the arithmetic, and confirm nothing else.

Weekly aggregation converts trades into counts, hit probabilities, average R, risk deployed, fees, adherence, branch contribution, profit factor, RAER and RAPF. Every one of those is a faithful function of the rows it was given. A branch misclassification does not break the function; it moves a real trade into a wrong bucket, and the resulting figures remain internally coherent while describing an operation that does not exist. The Structural Diagnostic Engine will read drift in that branch and report it accurately. The drift is real in the data and absent in the account, and nothing at any stage distinguishes those two situations.

The distinct failure of the second kind

Rows entered from memory are a different problem from rows entered wrongly.

The Trade Capture Document exists specifically as a memory bridge between execution and the Saturday review, and the reason it exists is that recall is not a neutral recording device. It compresses, tidies, and moves entries slightly toward what was intended. An operator reconstructing a week's MAE values on Saturday is not fabricating anything and will nonetheless produce a record measurably kinder than the week was. The error has a direction, which makes it far worse than random noise: noise widens confidence intervals, while directional error shifts every downstream figure the same way and leaves the shape of the data undisturbed.

• Downstream stages verify internal coherence, never correspondence.
• Misclassified rows produce accurate diagnostics of a fictional operation.
• Reconstructed data carries a direction, which is worse than random error.

What this is not

This is a fidelity problem, not the revision problem.

The evidence core also operates an archival rule — rows written at capture time, consumed read-only downstream, corrected through a logged entry rather than an overwrite. That rule protects the record from being quietly improved after the fact, which is a genuine and separate hazard. It offers no protection whatsoever here. A row that was wrong when it was written is preserved perfectly, forever, and inherited by every calculation that touches it. Write-once discipline guarantees that the record is stable; it guarantees nothing at all about whether the record is true.

What actually defends the stage

The defences are all located before the row exists.

Because no downstream check can help, every available protection sits at or before capture. The broker and platform record supplies the fields that should never be typed from memory — fill price, stop distance, duration, fees. The capture document is completed near the event rather than at review. Branch identity is assigned at entry, when the intent is unambiguous, rather than inferred afterwards from how the trade behaved, which is the single most common route to a misclassified row. None of these are sophisticated, and their unsophistication is not a reason to treat them as optional. They are also the only defences available, which is an unusual position for a system this dense with cross-checks and worth stating rather than assuming.

The key idea

The least interesting stage sets the ceiling on every other one.

Journal entry is the most tedious obligation in the system and the one with no visible reward, which is precisely why it decays first when a week gets busy. The consequence is not proportional. A ten per cent decline in capture fidelity does not cost ten per cent of the analytics — it makes every figure downstream unfalsifiable, because the operator can no longer distinguish a finding about their trading from an artefact of their record-keeping. The instruments remain fully functional. What they are measuring quietly stops being the account, and no reading anywhere in the system carries a mark indicating when that substitution occurred.