The condition of the subject as a protocol variable

Subject state moves a score as readily as distance or light, because it changes what is actually in the frame - and unlike a camera setting, no tool can correct for it. Distance, angle and light get the attention because they are set before the shutter; state is not.

What "state" means here

Without getting explicit about it, a subject can be in a noticeably different condition from one session to the next, for reasons that have nothing to do with camera setup at all - time of day, temperature, recent activity, arousal, or simply where in a natural range of variation a given moment happens to fall. None of that is a flaw in the protocol. It is closer to photographing the same landscape in different weather: the scene itself changed between shots, not the camera pointed at it.

A rating tool has no way to detect this. It receives an image and scores what is in the image, the same way it would score any other input, because it is scoring the submission it was given, not the underlying reality the submission is meant to represent across sessions.

Why the tool cannot correct for it

Camera variables like distance and angle are, in principle, correctable - a tool could in theory detect and adjust for a foreshortened angle, even if most do not bother. Subject state is not correctable in the same sense, because the tool has no reference for what "resting state" or "typical state" looks like for a specific individual. It has never seen a first, neutral photo to calibrate against. Every image is evaluated on its own, so a state difference between two submissions reads as a real difference in the subject, not as noise around a fixed baseline - and there is no processing step downstream of capture that could catch it, the same way normalisation maps a raw score onto the visible scale but does nothing about what the raw score was based on in the first place.

Why mixing states breaks a comparison

A two-photo series where the state genuinely differed between shots is not one series measuring change over time. It is two independent, unrelated data points that happen to share a subject, and any movement between their scores is uninterpretable - it could be the state, could be something else entirely, and there is no way from the numbers alone to separate the two. This is the same failure as changing two things between one comparison and the next: a result you cannot attribute to a single cause is a result you cannot use. It also sits underneath mixed-condition sets more broadly, where state is one of several conditions that has to hold for a set to describe one thing rather than several.

What to do about it

Fix the state the same way you would fix distance or light: choose a consistent point to shoot at, session to session, and note it in the same log entry as everything else. "Same time of day, similar recent activity" is enough of a note to flag a session that clearly does not match, without needing anything more specific than that. This is a photo-side discipline, not a claim about the subject itself - measurement in the physical sense belongs to a tape and a method, and a photo-based rating tool was never going to hold that kind of reference regardless of state. It is also a reason two people comparing scores against each other run into more noise than they expect: state varies between people in ways a single photograph cannot show, which is one more variable sitting underneath a comparison a human judge might make more allowance for than a fixed model does. A tool that publishes a breakdown, the way Rate Cock shows per-axis results on public entries, at least gives a chance of noticing that the axes most sensitive to state moved together while the rest held still - one more reason a rubric earns its place over a single number.

What a series with mixed states actually shows

None of this means the sessions were wasted, only that they should not be read as one continuous line. A set of results taken across genuinely different states is still real data - it is just data about the range you cover across states, not data about a trend over time. Reading it as a trend is where the mistake happens: a dip or a rise gets attributed to some change in setup or effort, when the more likely explanation is that the state simply differed between the two sessions being compared. Epidemiologists have a name for one version of this trap: Barnett, van der Pols and Dobson (2005) describe regression to the mean as a phenomenon that "can make natural variation in repeated data look like real change". The honest read is to group results by state where you can tell them apart, and only compare within a group.

This is a narrower discipline than logging distance or light, because state is harder to pin down in a single word and easier to forget to note at all. A rough flag - typical, atypical - next to a result costs almost nothing to write and is usually enough to separate a real comparison from one that was never valid in the first place. What a model is actually doing with the pixels it receives has nothing to say about state either - it has no concept of a subject's history, only the single frame in front of it, which is the root of why this variable is invisible to the tool by construction rather than by oversight.