The variables that will not hold still

Three conditions will not hold still however careful the protocol: ambient light, the ageing of a device, and the subject's own state. The honest response is not denial but two moves - take more samples, and report a wider range - after standardising everything that can be fixed.

Ambient light, even when you fix the source

Choosing window light on an overcast day, at a fixed hour, is the most reproducible light most people can arrange - and it still is not identical from one week to the next. Cloud cover varies, season changes the sun's angle at the same clock time, and a stretch of unusually bright or dim weather will shift what "overcast" delivers. Time of day narrows this a great deal but does not close it; weather sits underneath the hour and it does not take instructions. Light shifts are not cosmetic to a model either: Afifi and Brown (2019) showed that colour casts from incorrectly applied white balance significantly degrade deep network performance on image classification and segmentation.

Device aging

A phone camera is not a fixed instrument over its lifetime. Lenses pick up dust, sensors degrade slightly, and manufacturers periodically push processing updates to the camera app itself, changing sharpening or tone mapping without changing the hardware at all. Using the same device every time removes the model-to-model variation, which is the larger source of noise, but it does not remove this smaller and slower one sitting inside a single device's own lifespan.

Subject state

State drifts for reasons that have nothing to do with the camera, the light, or the protocol, and no amount of holding the rest of the setup constant reaches it. This is the variable a protocol is least equipped to fix, because it lives entirely outside the photographic conditions the rest of this protocol controls.

The two honest responses

There is no third option that makes these variables go away, so there are only two moves worth making.

Take more samples. A single result under conditions that cannot be perfectly held is one draw from a distribution that has some width to it whether you like it or not. Several results taken close together average out a meaningful share of that width, in a way that no amount of care on a single shot can.

Widen the range you claim. When the underlying conditions carry irreducible variation, the honest description of a result is a range, not a point. An error bar exists for exactly this reason - it is the part of the result the display leaves out, and estimating your own from repeated shots is more useful than assuming it away. Turning one result into a defensible range is the practical version of this - a worked method for going from a handful of repeats to something you can actually state a confidence about.

Distinguishing irreducible noise from a fixable variable

Before writing a variable off as unfixable, check whether it is genuinely outside your control or just currently unfixed. Weather and device aging are irreducible in the sense that no amount of care removes them; a crop done by eye instead of against a fixed reference is not irreducible, it is simply undone discipline dressed up as bad luck. Metrology draws the same line: the JCGM international vocabulary of metrology defines reproducibility conditions as ones that can include different locations, operators and measuring systems, and notes that a specification should give "the conditions changed and unchanged, to the extent practical." The test is whether a documented method exists that would close the gap - if one does, the variable belongs on the fixable list, not this one, and the honest move is to adopt the method rather than budget for the noise.

What this is not permission to do

This is not licence to skip the parts of the protocol you can fix. The full list of what a protocol does control - device choice, distance, angle, crop, background - stays mandatory; the point of nailing those down is that it shrinks the pool of variation down to the few things genuinely left over, which is what makes the sample-and-widen approach affordable in the first place. Treat everything as unstandardisable and you need dozens of samples to say anything; standardise what you can and the residual noise from weather, device aging and state is small enough that a handful of repeats covers it.

Where this shows up elsewhere

Physical measurement carries an equivalent admission. Even a documented method has a stated tolerance rather than a claim of exactness, because no measuring process removes all variation either. A model's read of an image has its own irreducible component on top of yours, which is one more reason a single result was never going to be the whole story. A human reviewer forms an impression that also varies session to session for reasons that have nothing to do with your protocol - mood, attention, the last submission they looked at - and the same fix applies: more than one data point, and an honest account of the spread. Rate Cock displaying a breakdown rather than one number is, among other things, a way of surfacing some of this variation instead of collapsing it into a single figure that looks more certain than it is.