A pair's home band is not permanent, and a pair managed with last year's coefficient pays last year's stop-out rate.

The bands

Tighter buys precision and pays in premature exits; wider buys survival and pays in giveback.

The spectrum runs from around 1.5 upward, and each band is a position on a single trade-off rather than a quality ranking. The tighter end suits clean, orderly expansion where noise is small relative to the move being captured — precision is cheap there because ordinary movement rarely reaches the stop. The wider end buys survival room in high-intensity conditions and pays for it in profit surrendered on every trade that retraces. An instrument's home band is the position on that spectrum where its typical noise is contained without buying room it does not use.

How migration happens

Slowly, without an event, and usually without the operator forming any view about it.

Migration rarely announces itself. There is no day on which a pair changes band; there is a quarter over which its typical range relative to its price drifts, and by the end of it the stored assignment is sized for behaviour the instrument has stopped exhibiting. Because the drift is slow it never triggers the live volatility read — the ratio compares against a baseline that is itself drifting, so the instrument keeps reading as normal for itself throughout. Both the stored band and the live ratio can be internally consistent and jointly wrong, which is what makes this failure specific rather than a special case of staleness.

The two directions

Migrating up costs stop-outs. Migrating down costs giveback on every trade.

An instrument that has become more volatile than its assignment is being managed with an envelope that ordinary movement now reaches, so its noise stop-out share climbs — the visible symptom, if the autopsy classification is being kept. An instrument that has become calmer than its assignment is being managed with room it no longer needs, which costs a slice of every retracement and shows up as elevated giveback for that pair specifically. Both are detectable, both are cheap to check, and neither is detectable from the volatility panel alone, because the panel's ratios are computed against the same drifting baseline.

• Upward migration shows in noise stop-out share; downward shows in per-pair giveback.
• The live ratio cannot detect either, because its denominator drifts with the instrument.
• Both symptoms are pair-specific, which is what makes them findable.

The review that catches it

Realised noise against assigned band, per instrument, once a quarter.

The check is deliberately coarse and infrequent, matching the timescale of the thing it is looking for. Once a quarter, for each instrument traded meaningfully, compare the realised noise behaviour — stop-out classification share and giveback — against what the assigned band should be producing. Instruments sitting outside their expected profile get flagged for reassignment, which is a deliberate edit with a date attached rather than a drift. Running the check monthly would produce noise and invite reassignments on samples too thin to support them; running it annually means carrying a wrong assignment for most of a year.

Why reassignment is an event

Changing a band resets the evidence for that instrument, so it has to be dated.

A reassignment is not a small adjustment. Every trade on that instrument after the change is managed under a different parameter, which means the instrument's trade history now spans two settings and cannot be pooled for calibration purposes without a segment boundary. This is the same discipline the weekly scorecard's Legend constants require, arriving in the volatility layer: record the date, record the reason, and read the instrument's evidence as segmented at that date rather than continuous across it. An undated reassignment makes the previous quarters of evidence quietly unusable.

The key idea

A stored parameter describing a changing thing needs an expiry review, or it becomes a fossil.

The calibration's stability is a feature and it carries a maintenance obligation. Every value that is stored because it is slow — the band assignment, the baseline, the branch weights, the expected excursion signatures — describes something that moves on a timescale longer than the review cadence but shorter than forever. Without a scheduled check, each of them keeps returning a confident answer about a state that has passed. The check does not need to be frequent. It needs to exist, and it needs to be on a calendar rather than triggered by the operator noticing something.