The Continuity Check Runs On 5 Tools

No off-the-shelf cinematography analyser exists. A studio can rent a colour suite, buy a shot detector or license a classifier, and still not have 1 product that runs the continuity check from the first cut to the last flag. So the check is assembled, and it is assembled from 5 tool classes. The 5 are shot detection, shot-type classification, lighting and colour consistency, exposure detection, and the prompt-level kits that run before a single frame is rendered. 4 of the 5 are software. The 5th is the 1 the edit actually leans on. 1 probability sits at the front of the build. The probability that a tool class left unstated is a tool class left unowned is high, because an unnamed step in a pipeline is exactly where the work falls between 2 roles and 0 people notice until a scene does not cut. So naming the 5 is not administration. It is the difference between a check and a habit, and a habit is the thing that gets skipped on the shoot day when there are 40 other decisions to make. The other reason to name them is cost. 3 of the 5 classes can be covered by 1 built skill with 0 model weights. 2 of them cannot be covered by anything that ships today, and knowing which 2 is what stops a studio from buying software that will never do the job it was bought for.

Shot detection comes first because every reading taken after it inherits its errors. The neural shot-boundary detector resolves cuts, dissolves and wipes, which is 3 transition families rather than 1, and the reason the extra 2 matter is that a dissolve has no hard edge for a frame difference to catch. Underneath it runs a weightless content-aware scene detector as the fast fallback. It costs no weights, no GPU time worth counting and about 1 pass over the file, and it is right often enough on clean cuts that the heavier model is not the default. 1 frame per 24 is the reason the fast path holds up. At 24 frames per second a cut is a discontinuity, and a discontinuity survives being sampled, because the content on either side of it differs by more than a sample. A dissolve at the same 24 frames per second is a gradient instead, and gradients are what a frame difference is blind to. 1 probability sits under the gate. The probability that a boundary set with 1 missed dissolve corrupts the shot list downstream is high. The probability that anyone traces the corrupt row back to the missed dissolve rather than blaming the classifier is low, which is why the boundary is checked before the analysis rather than after it. The order is the whole argument. Check the cut, then name the shot, then measure the light.

Shot-type classification is the 2nd class, and it is the 1 that turns a boundary into a description. A classifier trained on about 400,000 frames recognises 6 shot types. A lighter 5-class model names 5 of them: long shot, full shot, medium shot, close shot and extreme close shot. 5 is not a rounding of 6. The difference is 1 band, and the band sits at the tight end, which is where a scene does its emotional work. That is why the heavier model is kept for the beats that lean on a close shot and the light model is allowed to carry the wide coverage where a miss costs less. There is a second reason to keep both. 400,000 frames of training is a large number for a classifier to have seen, and a large number is exactly what makes it confident on the 1 shot type it saw least. A 5-class model that is honest about its 5 is worth more on a wide shot than a 6-class model that is guessing at its 6th. 1 probability applies here. The probability that a classifier mislabels a shot it has rarely seen is high, and the probability that nobody checks is higher still, because a mislabelled medium shot looks like a working shot list rather than an error. 2 models, 1 job, and a rule about which 1 gets the close coverage.

Colour consistency is the 3rd class, and it is the 1 that produces the only output in the whole check that a director can argue with in a room. A colour matcher maps the distribution of 1 shot onto another, on a histogram or on MVG, and the residue between the 2 becomes a drift score. The drift score is the point. It converts this shot looks different into a number that can carry a threshold, and a threshold is what lets 40 shots be reviewed in the time 1 of them used to take. Exposure detection is the 4th class: low-light, bright, flat and optimal, run natively inside the render stack so it adds no separate pass and no separate tool. 4 states is a small vocabulary, and it is enough, because a frame that is flat is flat regardless of how it got there. Read per shot, 3 numbers cover the rest of it: luminance, warm-cool balance and RGB. The studio recommendation adds 1 more step rather than 1 more tool - a z-score against the run, so the flag points at the shot that sits furthest from its own neighbours instead of at the shot that is merely bright. 1 probability closes the section. The probability that a drift score left unthresholded gets ignored is high, because an unthresholded number is an opinion. The probability that a flagged shot gets a second look is also high, and that is the entire return on the measurement.

There is no product that runs the continuity check. There are 5 tool classes and 1 decision about which of them to build.

The 5th tool class is not software. It is the storyboard kit, and it exists because 3 rules are not automated anywhere: the 30 degree rule, the 180 degree rule and screen-direction enforcement. Those 3 are not edge cases. They are the rules that decide whether 12 shots assemble into 1 scene or into a set of 12 unrelated frames, and no classifier, matcher or detector in the toolchain enforces any of them. The reason is structural: they are statements about intent across a sequence, not properties of any 1 frame, and a frame-level model has nothing to read them off. So they live at prompt level, in 4 pieces of kit: a camera-movement continuity matrix, a screen-direction checklist, a reverse-angle kit and a shot-continuity map. The matrix is the one that earns its place, because it makes the axis of each move explicit before the move is written into a prompt, and an axis written down is an axis that can be reused by accident or changed on purpose. Character, wardrobe and ethnicity continuity has no clean off-the-shelf answer either. The pragmatic fix is a similarity check: embed every shot, then flag the outliers, which catches a changed collar or a shifted skin tone without asking a model to understand either. 1 probability sits over the class. The probability that a studio automates the 4 classes it can and quietly skips the 1 it cannot is high, and the cost of that skip is not visible in any single frame. It shows up once, at the end, when 12 correct shots refuse to become 1 scene. 4 of the 5 classes run themselves. The 5th is the reason a studio still needs a pair of eyes on the sequence.

The recommendation that came out of the review is 1 built skill rather than 5 subscriptions: cut detection through ffmpeg scene difference with no model weights, per-shot luminance, warm-cool and RGB readings, and a z-score drift flag over the run. That covers 2 of the 5 classes end to end and gives the other 3 a place to write their numbers. It also costs $0 per shot and 0 GPU hours, which matters more than it sounds, because the reason a continuity check gets skipped is almost never the difficulty. It is the 40 minutes. The same order shows up in the work the tools sit beside. Companies built across 12 countries. A EUR 75 million industrial group restructured. 210 energy systems deployed across Africa and Asia. 200 wood-gasification machines built across the UK and Europe. 16 years of field work from rural Nigeria to post-conflict Serbia, Kazan and Fukushima. A book of 393 pages published on 8 April 2026. 25,000 applications arrive for 25 places in a cohort, and 210 alumni across 19 countries now run the same order in their own work. In every 1 of those the sequence was the same as it is here: name the classes, automate the 4 that can be automated, keep the 1 that cannot under a human eye, and let the numbers point at the shots that need it. 1 last probability. The probability that a continuity check built from 5 named classes catches more than a check built from habit is high. The probability that the 1 unautomated class is the 1 that decides whether the scene reads as a scene is close to certain. So the next step is small and it is yours. Take the 1 quality check you still run by eye on every project, and write down which of the 5 classes it belongs to. Which 1 of your checks has never been named?