The slow failure of a protocol

Protocol drift is the slow, unnoticed change in your standard conditions - a little closer, a lamp nudged, a looser crop. Each session looks fine against the last, but over a year the change is obvious. A written log, checked against the room rather than memory, is the defence.

How it happens

You stand slightly closer than last time because the room felt more crowded, or the lamp gets nudged during cleaning and never gets nudged back, or the crop gets a little looser because tighter felt awkward that day. None of these single changes would show up if you compared this session only to the last one. Compared to the session from a year ago, the drift is often obvious - a different distance, a different light angle, a frame that includes noticeably more background than the one it started as. Camera distance alone accounts for a large share of unrelated score movement, which is exactly why a small, repeated drift in distance compounds into something significant well before you notice it happening.

Why it is invisible in the moment

Human memory is bad at holding a precise distance, angle or crop stable across months without an external reference. You remember the intent - "level, arm's length, plain background" - and each session you reconstruct that intent from feel rather than from measurement, and feel drifts. The same problem exists with a documented measuring method, which is why serious methods specify a fixed reference point rather than relying on the person repeating a gesture from memory.

The defence is the log

The fix is not better memory or more discipline in the moment. It is a written record from the reference protocol that gets checked against the current session, not just filled in and filed away. A log with a mark on the floor for distance, a note for the lamp's position, and a fixed crop ratio gives you something external to check against, which is the only thing that catches an inch of drift that feels, from the inside, like no change at all. Laboratories use the same idea with a physical object: the NIST/SEMATECH Engineering Statistics Handbook says "bias and variability are controlled by monitoring measurements on a check standard over time," and a persistent shift in those readings flags a changed procedure. A more discrete set of events retires a baseline outright - a new phone, a tool update - and drift is the other failure mode: nothing discrete happened, the conditions just wandered.

Where else this applies

A model's read of an image has no way to distinguish drift in your conditions from a genuine change in the subject; it only ever sees the current frame. A human reviewer comparing a photo from this month to one from a year ago is subject to the same blind spot from the other side, and cannot correct for a drift they were never told about. Rate Cock showing conditions alongside a public entry, rather than just the number, is one of the few places where drift becomes checkable by someone other than the person it happened to.

Check the log against the room, not against memory, and drift stops being invisible.