Too wide leaves a measurement. Too tight leaves a small loss that looks like bad luck.

The two costs

Both are real, both are ongoing, and only one of them is denominated in a field you look at.

Set the envelope too wide and every trade that retraces before running gives back more than it needed to; the position survives but the outcome is smaller, and giveback records the difference in R. Set it too tight and a share of trades that would have run are removed by movement that did not invalidate anything. The cost is the entire remainder of a move that never appears in the log, because the position was gone before the move happened. Nothing measures a trade's counterfactual. The trade closes at a small loss and joins the population of ordinary losing trades, where it is completely invisible.

The drift this produces

An operator optimising honestly against visible evidence will tighten, quarter after quarter.

This is the practical consequence and it is worth stating as a prediction rather than a warning. Each review shows giveback as a number that could be lower. Tightening the coefficient lowers it, visibly and immediately. The premature stop-outs that the tightening causes arrive as ordinary losses, spread across the sample, attributable to nothing. So every individual adjustment is supported by the evidence on the page, and the sequence of them walks the setting past the optimum and keeps going. Nobody made a mistake at any step. The measurement apparatus was asymmetric and the setting followed it.

The same shape one layer up

This is the capture ceiling, arriving at the coefficient instead of at the exit rule.

The pattern will be familiar to anyone who has read the execution lab's argument about capture efficiency, and it is the same mechanism rather than an analogy. There, driving capture toward its maximum means exiting at every pause, which raises a measured ratio while removing the right tail. Here, tightening the envelope lowers a measured giveback while removing the trades that would have reached the tail. Both are cases of a metric with a monotonic direction being optimised past the outcome it was introduced to protect. The coefficient version is worse only because it operates before the trade rather than during it, so there is even less trace.

What detects the invisible half

The noise stop-out share and the far-rung hit rates, neither of which is a volatility field.

Since the cost cannot be measured directly, detection has to come from adjacent evidence. The stop-out autopsy is the first instrument: a rising share of noise stop-outs in a band means the envelope is failing to contain ordinary movement, and that is exactly the invisible failure made countable. The second is the checkpoint ladder — if the share of trades reaching the far rungs is falling while entries and structure look unchanged, positions are being removed before they get there. Neither field lives on a volatility surface, which is precisely why an operator working only from the calibration screen will not see the drift.

• Noise stop-out share per band is the direct detector; it lives in the feedback loop.
• Falling far-rung hit rates with unchanged entries is the corroborating signal.
• Giveback falling while both of those deteriorate is the signature of over-tightening.

The stated target

A band needs a target, because against a direction it will be optimised forever.

The structural fix is the same one the capture argument reaches: state what the band is supposed to achieve rather than which way it should move. A coefficient band with a target noise stop-out share and a target giveback range can be read as above, below or at specification, and all three readings are meaningful. A band evaluated against the instruction to reduce giveback has only one possible verdict every quarter, forever, and the setting will keep travelling in the one direction the evidence can see.

The key idea

Check what your measurements cannot see before trusting the direction they point.

The general lesson is about measurement design rather than about volatility. Whenever two failure modes bracket an optimum and only one of them produces a field, the visible one will dominate every review and the setting will drift toward the invisible failure indefinitely. The defence is not vigilance, because each step is individually justified. It is to identify the unmeasured direction in advance, find a proxy that does register, and put the proxy in the same review as the metric it is supposed to balance.