Raw stop-out rate cannot tell you whether the envelope was too tight. The classification can.

The two categories

The thesis failed, or the envelope was too small for ordinary movement.

A structural stop-out is one where the market did what the trade was betting against: the level broke, the trend resumed the other way, the setup was invalidated in the terms the plan defines. The coefficient is irrelevant to it — a wider stop would have produced a larger loss on the same failed thesis. A noise stop-out is one where the market did not invalidate anything and the position was removed by ordinary movement, with the structure still intact and frequently with the original thesis resolving favourably afterwards. That is the only category the coefficient can fix, and the whole loop depends on being able to count it.

Why it must be done at the time

The classification depends on chart context that stops being available almost immediately.

The judgement requires seeing what the market did around the stop — whether the level actually broke, whether the move continued, whether structure was intact. That context is available for a short window and then requires reconstruction, and reconstruction is where hindsight enters. A trade reviewed at the weekend gets classified partly on what happened afterwards, which contaminates the category with information the decision did not have. Worse, the direction of contamination is predictable: trades that would have worked get remembered as noise stop-outs, because the knowledge that the thesis was right makes the exit feel premature.

What the loop does with the counts

Noise share per coefficient band is the number that moves a calibration.

Once classified, the aggregation is straightforward and the output is specific. Group the stopped trades by the coefficient band they were managed under and compute what share of each band's stop-outs were noise. A band with a low and stable noise share is correctly sized. A band whose noise share is climbing is being asked to contain movement it was not sized for, and that is a calibration finding with a named band attached rather than a general sense that stops feel tight. It is also checkable against the opposite failure, because a band with almost no noise stop-outs may be wider than it needs to be and paying for it in giveback.

• The unit of the finding is a band, not the book — which is what makes it actionable.
• Zero noise stop-outs in a band is a signal too, pointing the other way.
• The comparison is across bands and across quarters, never across single trades.

The honest limit

The classification is a judgement, and it is the softest input in the whole calibration.

This should be stated plainly rather than glossed. Whether a stop-out was structural or noise is not computed from a formula; it is decided by an operator reading a chart, and two competent people will disagree on some fraction of cases. The loop is built on a subjective input, and its outputs inherit that softness. What makes it usable anyway is that the judgement is made per trade, in the moment, without knowing the aggregate it will feed — so individual errors are unlikely to point in a consistent direction, and the band-level share is far more stable than any single classification.

Where it lands

The classification travels with the coefficient, the zone and the ratio into the trade record.

The autopsy is only useful because it sits alongside the fields captured at entry. A stop-out classified as noise, on a trade managed at a known coefficient, in a known volatility zone, at a known ratio to baseline, is a complete observation — it says that this band failed to contain ordinary movement under these specific conditions. Without the entry fields the classification is an anecdote about a stop being tight. The capture discipline at entry and the autopsy discipline at exit are two halves of one record, and either alone produces very little.

The key idea

A loop is only as good as the one field that cannot be computed.

Almost everything in the calibration is derived — the ratio, the zone, the intensity, the distance. All of it is arithmetic on data the system already holds. The single input that has to be supplied by a human, at a specific moment, from context that expires, is the one the entire loop's usefulness rests on. That pattern recurs across the system, and the practical lesson is consistent: find the field that cannot be derived, and treat it as the discipline that actually matters, because everything else will keep computing correctly whether or not it is there.