Why the clock belongs in the log

Time of day matters because it is light direction and colour temperature in disguise: the same window gives different light in the morning and the afternoon. A protocol that fixes the light source without fixing the hour has only fixed half of it.

Why the hour matters

Morning light and afternoon light come from different angles, and angle of light is exactly the variable that decides what a photograph shows in relief - what reads as texture at one hour can read as flat at another, on the same subject, in the same room, through the same window. Colour temperature shifts across the day too, warmer near sunrise and sunset, cooler and flatter in the middle hours on an overcast day. A tool scoring the image cannot separate "the subject changed" from "the light changed"; it only has the pixels in front of it. Image models are measurably sensitive to this kind of shift: Hendrycks and Dietterich (2019) include brightness, which they note "varies with daylight intensity", and lighting-dependent contrast among the 15 common corruptions they use to test how far classifier accuracy degrades.

Why this is worse than it looks

The trap is that a room can look consistent to the eye across a whole afternoon while the light inside it is drifting the entire time. A protocol that fixes the light source in principle - window light, diffuse, side-on - still needs a fixed hour to actually deliver a repeatable result, because "window light" describes a category, not a single condition, and the category covers a wide range across one day. Flash removes the time-of-day problem but introduces its own, which is why the fix here is not "switch to flash," it is "pick a window and stay in it."

The fix

Shoot at the same time, within an hour or so, every session. Note the time in the log alongside device, distance, angle and crop; it costs one extra line. This is one entry in the fuller reference-conditions protocol, not a replacement for it - the other seven variables still need fixing regardless of what the clock says.

Where this shows up elsewhere

The same logic applies wherever a natural process substitutes for a controlled one. Physical measurement has its own version of a condition that looks stable and is not - a tape held slightly differently, at a different point in the day, without either detail written down. And a human judge reading photos taken across a scattered set of hours is reading a set with an uncontrolled light variable in it too, even though nothing about their process changed. Accuracy at the model layer assumes a stable input; time of day is one of the quieter ways that assumption gets broken without anyone deciding to break it.

One line in the log, same window every session. It is the cheapest fix in the whole protocol for the amount of noise it removes.