Black Geometry

Darkness is the corpus's default state. Across 51,453 SDR shots the median frame has 65% of its pixels below L* 20 and 35% below L* 5. Where the highlight study found a minority event, this one is describing where most of cinema actually lives.

The folklore about that region is a pair of adjectives — crushed and lifted — attached to a black point. Neither survives the corpus as a description of a film. Which percentile a film's first per cent sits at is almost entirely a shot property; the toe's shape, measured as how it distributes L* across the lowest thirty percentiles, has an intraclass correlation of 0.003, which is as close to nothing as this study measured.

What does belong to the film is a decision about the floor: whether pixels are allowed to reach absolute zero at all. That has an intraclass correlation of 0.23, four and a half times the equivalent number at the top of the range on the same 41 films. Films commit to a black point. They do not commit to a white one.

And the compositional finding underneath both: a dark frame spends its remaining contrast unevenly and on purpose. Structure per unit of area inside the dark region runs at half the frame's average, while inside detected face boxes it runs at 1.34 times. Darkness is where cinema puts the things it has decided not to spend contrast on.

Corpus. 43 films, 52,527 analysis-eligible shots

Re-measured. 9,856 frames, pixel by pixel

Median frame. 65% below L* 20

Measured resolution floor. ΔE00 0.67 between neighbouring shadow codes

1. corpus — Two thirds of cinema is in shadow: the median frame has 65% of its pixels below L* 20, and a third of all shots have more than 80%.
2. census — The shape of the toe is not a film property. Its intraclass correlation is 0.003 — indistinguishable from nothing. (overturns the folklore)
3. corpus — Whether a film lets pixels reach absolute zero is a film property, at 0.23 — four and a half times the equivalent at the top of the range.
4. census — Contrast is withdrawn from darkness and concentrated on faces: 0.50 of a fair share inside the dark region, 1.34 inside detected faces.
5. census — In the darkest fifth of frames, one in twelve holds a face the file contains and the render does not show.
6. corpus — Subjects separate from their backgrounds more in dark frames, not less: ΔE00 20.9 against 13.8. (overturns the folklore)
7. census — The shadow holds a median of seven distinguishable populations per frame, and 98.8% of frames hold at least two.
8. census — Black separates by level and white separates by colour: 25% of the widest shadow separation is non-lightness against 38% at the top.
9. census — Cut-paper black is rare — 5.2% of frames — and the three most extreme measurements of it are all fully readable pictures.
10. census — Dark shape sorts into five families, and none of them is a film: the same five appear across all 43 titles.
11. census — A curated set of thirty frames displaces a film's shadow statistics by a median 13 percentile points — less than at the top, and still too much. (about the method)

Crushed and lifted describe one number. Everything the eye actually does at the bottom of the range is invisible to it.

Crushed and lifted are the two words the trade has for the bottom of the range, and between them they describe one number: where the black point sits. Everything else the eye does down there — whether a shape reads, whether two dark materials stay distinct, whether a face survives, whether the frame has depth — is invisible to that number and to the histogram it comes from.

This report proposes the term black geometry for what is actually going on: the spatial and tonal organisation of dark regions, measured as arrangement rather than amount. Two frames with 80% of their pixels below L* 20 can be one continuous mass or four hundred fragments; can be a wall of nothing or a face at absolute zero with the eyes intact; can be centred, bordered, bottom-weighted or evenly spread. None of that is in the occupancy figure, and all of it is measurable.

That hypothesis, which this study set out to test, survives the corpus better than the black-point account does — and it turns out to be measurable in a specific way that the phrasing does not anticipate. Selectivity here is not mainly about which tones are kept. It is about where the frame's remaining contrast is put.

A frame has a finite amount of gradient energy, and a dark frame has less of it than a bright one. Measuring how that energy is distributed — between the dark region and the rest, and between faces and everything else — turns an aesthetic claim into an accounting one. The accounts are lopsided, consistently, in the same direction, across every film measured.

The same three tiers as the highlight study, a resolution floor measured from the files, and a readability probe built to be indifferent.

Three tiers of evidence, kept apart, exactly as in the highlight study and for the same reason. The corpus tier is all 52,527 analysis-eligible shots, carrying the pipeline's own measurements — which at this end of the range are unusually good: the shadow zone's internal percentiles, the light-island topology, the centre-against-edge difference, and a local comparison between coherent skin regions and the ring of picture around them. The census tier is 8,537 frames re-measured pixel by pixel, complete for seven films. The curated tier is 1,319 selected frames, which are evidence about themselves and nothing else.

Two limits are structural and are measured rather than glossed.

The first is eight bits, which bites harder here than at the top. Below L* 10 a gamma-2.4 encode has about thirteen code values to spend. Every measured frame therefore carries how many distinct luma codes its shadow set actually uses, and the resolution floor is derived from the files: the median nearest-neighbour CIEDE2000 between the distinct quantised colours in a frame's shadow set is 0.67, with a 95th percentile of 0.97. Two shadow populations must clear ΔE00 2 — three times the floor — before this report counts them as two.

The second is the probe used for readability. There is no observer study here; asking whether a human can read a dark frame requires humans. What can be done is to apply a fixed, indifferent machine detector twice to the same frame — once as delivered, once after a single monotone, non-clipping gamma lift — and report the difference. The detector is a 2001-vintage Haar cascade and a poor stand-in for a viewer. Its absolute hit rate means little. The gap between its two runs means something specific: information the file contains that the render does not present.

An earlier version of the probe rescaled each frame between its own 0.5th and 60th percentiles. It opened the shadows and blew everything above the midtones, and on a full pass it destroyed roughly two and a half times as many detections as it recovered. The method was replaced with the fixed monotone curve and that pass was not kept, so those figures are not reproducible from the files shipped here — which is the correct status for a number produced by a method the report does not use.

It is recorded here because the same mistake is easy to make at the desk, and because a probe that destroys information while revealing information cannot measure how much information there was.

• Films: 43 — 1964–2025
• Shots aggregated: 52,527 — analysis-eligible rows
• Frames re-measured: 9,856 — at native resolution
• Complete pixel coverage: 7 films — 8,537 frames
• Quantisation floor: ΔE00 0.67 — measured, median over frames
• Separation threshold: ΔE00 2.0 — ≈3× the floor

Defined here and nowhere else. Every later use links back to this list rather than re-explaining itself in a caption.

Darkness is the corpus's default state: the median frame is 65% below L* 20, and half of all shots put no pixel at absolute zero at all.

The median SDR shot in this corpus has 65% of its pixels below L* 20 and 35% below L* 5. Nearly two thirds of all shots have more than half the frame in shadow, and a third have more than four fifths.

That is the first reason the two ends of the range need different treatment. Highlights are an intermittent event occupying a fraction of a per cent of a frame; shadows are the substrate. A measurement of shadow behaviour is a measurement of most of the picture.

• Median frame below L* 20: 65%
• Median frame below L* 5: 35%
• Shots >50% in shadow: 64.7%
• Shots >80% in shadow: 33.3%
• Shots with >1% at absolute zero: 21.9%
• Shots with no pixel at zero: 52.1%

Half the corpus never puts a single pixel at code zero, and one shot in nine puts more than a tenth of the frame there. That is a bimodal population, not a spectrum, and it is the first sign that the black point is a decision rather than a consequence.

Portrait of a Lady on Fire sits at one extreme with 25% of the typical frame below L* 20 and a first percentile at L* 9.6 — a film that never goes near black, and never had to. Enter the Void sits at the other with 94.7% below L* 20.

Films differ enormously in how often they touch the floor — and the shape of the descent to it turns out to carry almost no film-level information.

Three low percentiles are available for every shot of every film, which gives one descent the whole corpus can draw: where a film puts the first percentile, the tenth and the median of a typical shot, and how much those placements move.

The ordering is legible and the dispersion is the point. On most films the interquartile range of the first percentile spans more than the gap between neighbouring films.

Two films make the case that the toe is not a look. In the Mood for Love puts more than a hundredth of the frame at code zero in 66% of its shots; Oppenheimer does it in 0.1% of its 3,460. Both are widely described as having rich blacks. Only one of them has any.

And the crucial negative: the shape of the descent — how a frame distributes its L* across the lowest thirty percentiles, normalised by the spread above — has an intraclass correlation of 0.003. Whatever a film decides about darkness, it does not decide the shape of its toe.

Darkness sorts into five recognisable shapes, every one of which appears in at least 37 of the 43 films. The shape belongs to the shot, not to the film.

Six descriptors, all of them about arrangement rather than amount: how much of the frame is dark, whether the dark is one mass or many pieces, how ragged its outline is, whether it sits at the border or the centre, how high or low its centre of mass is, and how much of the frame's contrast energy falls inside it.

The families were fitted on the census tier alone, balanced across its seven films at 824 frames each, and every other frame was then assigned by nearest centroid. Fitting on everything would let the three Jeunet titles, which contribute a third of all re-measured frames, define what a family is. The check that matters is that the shares come out the same on the frames that did not fit them, and they do — within two points on every family.

What does the same amount of darkness look like arranged two different ways?

Each frame is shown as delivered and again as a black-geometry map: tonal bands painted on, with the largest connected dark mass outlined in white. The two frames carry comparable amounts of darkness arranged in opposite ways, and no occupancy statistic distinguishes them.

Measured on: curated.

A dark frame spends its remaining contrast unevenly and on purpose: half a fair share inside the dark region, a third more than a fair share inside faces.

A frame contains a finite amount of gradient energy and a dark frame contains less of it. The useful question is where that energy goes, and it is answerable directly: compute the gradient magnitude over the whole frame, then measure what share of it falls inside each region, normalised by that region's share of the area. A density of 1.0 means a region carries exactly its fair share of the picture's structure.

The answer is the same on every tier and in every film. Darkness is systematically under-served and faces are systematically over-served.

The face figure needs its caveat stated plainly. A Haar cascade finds frontal faces in reasonable light and misses a great many that a viewer sees instantly; the 3,036 frames it fires on are not a sample of the frames that contain people. What the number establishes is not how much contrast faces get in general, but that where a detector does find a face, that region is reliably above the frame's average structural density — which is the direction the hypothesis predicts and the opposite of what a uniform grade would produce.

The second use of the detector is stronger, because it compares the frame with itself.

Read carefully, this is a measurement of intent rather than of failure. The information is in the delivered file; a viewer with a brighter display or a lifted setup would see some of it; the grade chose not to present it. That is the operation the word crushed is groping at, and it can now be quantified per frame rather than asserted per film.

The lift is monotone but the detector is not monotone-invariant, so it both gains and loses: across the whole corpus it finds a face in 5.3% of frames where the delivered version showed none, and loses one in 2.9% where the delivered version had it. The net is small and the gradient across the darkness bands is the finding, not the level.

It also sets a bound on the folklore's fear. Even in the darkest band, the lift changes the detector's verdict on fewer than one frame in twelve. In the other eleven, either the face was already legible or it is not there in any form the file could give back.

Subjects separate from their backgrounds more as frames get darker, not less. The intuition that darkness merges them is the wrong way round.

The pipeline measures, on every shot with a morphologically coherent skin region, the CIELAB difference between that region and the ring of picture immediately around it. It is a heuristic and not a person detector, and it is the only subject-against-background measurement available on 37,819 shots.

The expected result was that separation collapses as frames get darker. It does the opposite.

The rise is not mysterious once stated: a dark frame is usually dark because the light is directional and local, and directional local light is exactly what produces a large difference between a lit face and an unlit surround. Darkness and subject separation are produced by the same lighting decision.

This is a direct refutation of the intuition that dark images merge their subjects into their backgrounds. Where merging happens it is a property of particular shots, not of dark shots in general.

The film-level extremes are instructive in both directions. The City of Lost Children reaches ΔE00 28.6 with 81% of the frame in shadow — subjects lit hard out of blackness. Let the Right One In sits at 7.6, and its lightness difference is negative: skin is slightly darker than what surrounds it, and what separation exists is carried almost entirely by colour. Both films are legible. They are solving the same problem with different currency.

Almost every frame in the corpus is layered: a median of seven distinguishable populations inside the shadow, three quarters of their separation carried by lightness.

The shadow pixels of each frame were clustered in CIELAB, keeping only populations at least ΔE00 2 apart — three times the measured quantisation floor — each holding at least 3% of the shadow pixels.

Almost every frame in the corpus is layered. 98.8% of frames with any shadow population carry two or more, the median frame carries seven, and 11.6% run into the search's ceiling of twelve and may hold more.

The composition of that separation is where black and white part company. Of the ΔE00 between a frame's two most distant shadow populations, a median of 16.1 — 75% — is lightness. At the top of the range the same measurement gives 62%.

Depth in the dark is therefore built the way the intuition says it is: background wall, hair, jacket, face edge, each a little lighter than the last, inside a range a display can barely resolve. What the corpus adds is that the range really is that narrow and the layering really is that fine, and that it survives an eight-bit encode by a comfortable margin over the measured floor.

Two films dissent. Zatoichi carries 59% of its shadow separation in colour and Sisu 48%: their darks are distinguished by hue as much as by level. Both are films with strong, cold ambient light, and both are exceptions that the corpus median would hide.

Cut-paper black is rare, and the three most extreme measurements of it in the corpus are all fully readable pictures — which changes what the measure is for.

The failure mode the trade names is cut-paper black: several objects collapsing into one identical value, so that the frame reads as flat shapes cut out and laid on each other. It is measurable — the share of the dark region with almost no local variation, against the number of effective tonal levels the region still holds — and it is rare.

Only 8.3% of frames have dark regions more than 80% locally flat, and only 5.2% are flat and down to fewer than twelve effective levels. What is more instructive is what happens when you go and look at the frames that measure most extreme.

What do the three flattest blacks in the corpus actually look like?

All three measure as catastrophic collapse and none of them is. The measurement is describing the pixels that are dark; the pictures are being carried by the pixels that are not. A flatness statistic computed over the dark region says nothing about the frame, which is the single most important caveat in this report.

Measured on: curated.

That is not a reason to discard the measure, but it is a reason to state what it is for. Flatness of the dark region is a description of one region. It becomes a diagnosis only in company: with the share of the frame that region occupies, with the contrast energy left elsewhere, and with a look at the picture.

The one relationship that does hold is with the floor. Frames that put more of the picture at absolute zero have flatter dark regions (ρ +0.38) and fewer surviving levels (ρ −0.58). At the top of the range the equivalent correlation runs the other way — frames that clip to white have more highlight texture. Clipping is a different event at the two ends, and the folklore is roughly right about one of them.

A film has a black point and a shadow density. It does not have a toe: the shape of the descent scores 0.003, and everything spatial scores barely more.

The same variance decomposition as the highlight study: for each measurement, the share of its total variance that belongs to the film rather than the shot, with a film-level bootstrap interval.

One measurement stands clear of the rest, and it is not one of the ones the vocabulary has a word for.

Everything spatial sits at the bottom of both charts. Fragmentation, raggedness, border weighting, the height of the dark centre of mass, contrast density inside the dark: all between 0.03 and 0.13. Darkness has shape, and the shape is chosen shot by shot.

Everything about the floor and about what survives above it sits at the top: whether pixels reach zero, how many effective levels the dark region keeps, how flat it is, how much of it carries any colour at all. Those hold still across a running order.

Which yields a compact and checkable claim. A film has a black point and a shadow density. It does not have a toe and it does not have a black geometry.

Three kinds of temporal structure, and a fourth that looks like structure and is not.

If shadow structure is a shot property it should still be an organised one. Plotting the band from the first to the tenth percentile along each film's running order, filled with the measured hue of its shadow zone, shows what kind of organisation it is.

Three patterns are visible and they are different things. Oppenheimer holds a floor a measurable distance above zero for its whole length — a rule the film keeps. In the Mood for Love sits on zero almost continuously, and its variation is in how much of the frame goes there rather than how far down it goes. Kwaidan and Blade Runner 2049 move in blocks, the floor stepping between sequences.

First Cow is the interesting case: its band is low and narrow throughout, and the film's darkness is almost entirely a matter of occupancy rather than depth. It is dark because most of the frame is dark, not because the bottom is unusually low.

Neutralisation on the way down is close to universal, and much stronger than the equivalent at the top. The films that dissent are the ones worth naming.

Each film draws one line from the mean chromaticity of its midtone zone to that of its shadow zone. The dominant behaviour is neutralisation, and it is much stronger than the equivalent at the top of the range.

The pattern is close to universal and the reason is partly arithmetic: chroma in CIELAB shrinks with lightness, and a heavily saturated midtone has further to fall. But the spread around that trend is real and is where the strategies are. Zatoichi and Sisu keep enough chroma at the bottom that 59% and 48% of their shadow separation is carried by colour; Let the Right One In turns almost 150° on the way down, its shadows landing on a completely different hue from its midtones.

Roma appears in the data with C* 0.0 at both ends. It is monochrome, and it is a useful control: a film with no chroma anywhere still layers its shadows into a median of several distinguishable populations, entirely by level.

Four asymmetries between the two ends of the range, all measured the same way on the same rows.

This study and the highlight study measured the same 43 films, the same 52,527 shots and the same 9,856 frames, using the same method at opposite ends of the tonal range. Putting the four comparable measurements on one axis is the clearest single statement either report can make.

Layering against crowding

98.8% of frames hold two or more distinguishable shadow populations, at a median of seven; 84% hold two or more whites, and 57% of frames hold no white population at all. The bottom of the range has room and the top does not — which is what a ceiling at L* 100 and a floor at L* 0 with a gamma-2.4 encode between them actually implies.

Level against colour

75% of the widest shadow separation is lightness; at the top the figure is 62%. Highlights, crowded against the ceiling, are forced to use hue to stay apart. Shadows, with more room below them, do not have to.

Two floors, one ceiling

Whether a film lets pixels reach zero has an intraclass correlation of 0.23; whether it lets a channel reach 255 has 0.05. A black point is a decision a production makes and keeps. A white point is whatever the brightest thing in the shot turned out to be.

Where the colour signature lives

Highlight chroma is the strongest film-level signature in either study at 0.42. Shadow chroma is 0.19. A film's colour identity lives in its bright zone, not its dark one — which is the reverse of where most grading conversation about a look tends to go.

One film puts a whole readable picture inside the bottom tenth of the range; another compresses both ends hard and spends everything in the middle.

Chosen because they are informative about the thesis rather than because they are extreme.

First Cow puts 93.7% of the typical frame below L* 20 and 64% below L* 5. It is the darkest film in the corpus by occupancy, and it clips to zero in only 12% of its shots — the darkness is achieved by putting the whole picture low rather than by pushing the bottom down.

Shot 182 is the case in one frame: two pairs of hands sewing by firelight, with 99% of the picture below L* 10. Fifteen effective levels carry the whole of it — cloth weave, a thread, the edge of a sleeve, the modelling on the knuckles. The frame is entirely legible and lives inside a range the display can barely address.

Its map shows why. The largest connected dark mass covers most of the frame and the hands sit inside it as a separate population; almost all the frame's contrast energy is in the hands and the cuffs, and the rest of the picture is being spent on nothing at all.

Can a whole readable picture live inside the bottom tenth of the range?

The lifted picture is not a better version of the frame. It is a picture of what the grade decided not to show, which is the quantity this report calls the shadow information budget.

Measured on: curated.

Sicario never reaches L* 90 at the ninety-ninth percentile in any of its 1,205 shots — that is the highlight study's most extreme single finding. At the other end it is equally committed in the opposite direction: 57% of its shots put more than a hundredth of the frame at absolute zero, against a corpus rate of 22%.

Shot 997 shows both decisions at once. Seventy-eight per cent of the frame is at code zero. What is left is a face, two out-of-focus practicals, and nothing else — and the face has eyes, brows, a beard and the modelling on one cheek. The film compresses both ends hard and spends everything it has in the middle.

This is the pairing that made both studies worth doing together. A film described as having beautiful light turns out to have almost no highlights and an aggressive floor, and to be doing all of its work in the band between them.

What is the difference between too dark to see and quantised to zero?

The frame as delivered, its map, and the same frame through the readability probe's fixed monotone curve. The lift separates two things the word crushed runs together: material rendered too dark to see, which comes back, and material quantised to zero, which does not. Only the second is irreversible, and this frame contains a great deal of it — on purpose.

Measured on: curated.

Two conclusions survived the frames most likely to break them, one had to be weakened, one was abandoned, and the study's largest single number is its least reliable.

Each conclusion was tested against the frames most likely to break it. Three held, one had to be weakened, and one was abandoned.

Held: flatness is not collapse

The three flattest dark regions in the corpus are a pair of eyes, a star field and a face at 78% zero. All three are fully readable. Searching specifically for beautiful flat blacks and ugly differentiated ones produced plenty of both.

Held: dark frames separate subjects more

The counterexample would be a film whose subjects merge as it darkens. Let the Right One In comes closest, at ΔE00 7.6 with negative lightness separation — and it separates by colour instead, and remains legible. No film in the corpus merges subject and background as a policy.

Weakened: films have a black point

An intraclass correlation of 0.23 is the strongest in the study and is still mostly shot variance. Half of all shots put nothing at zero at all, including shots from films that habitually do. The claim is about central tendency, not about every frame.

Abandoned: a crushed-versus-lifted axis

The report set out expecting films to sort along a toe-shape axis. They do not: the shape of the toe has an intraclass correlation of 0.003, and no clustering on shadow descriptors recovers anything resembling that axis once occupancy is accounted for.

One caution about the strongest number in the report. The census-tier intraclass correlation for absolute-zero clipping is 0.607 with an interval running from 0.02, which means it is being carried by one or two of seven films. The corpus-tier figure — 0.228 with an interval of 0.11 to 0.33, on 41 films — is the one the argument rests on, and the census tier is quoted only because it agrees in direction.

Four consequences at the desk, starting with the distinction the word crushed hides.

None of what follows reconstructs anyone's node tree. These are the practical consequences of the measurements, stated as things to look at and things to stop trusting.

Separate the two crushes

Material rendered too dark to see and material quantised to zero look identical on a picture monitor and are not the same event. The first is recoverable by anyone with a brighter display; the second is gone.

• A waveform with a trace resting on the floor is the second. A trace hugging the bottom but off it is the first.
• In this corpus half of all shots contain nothing at zero at all, so a floor reading of exactly zero is a choice, not a default.

Grade the budget, not the level

The measurement that separated good dark frames from indifferent ones here was not how low they went but where the remaining contrast sat. Structure inside the dark region runs at half a fair share; inside faces at 1.34.

• A local-contrast tool aimed at the whole frame spends the budget evenly, which is the opposite of what the corpus does.
• Windows and qualifiers exist to spend it unevenly; that is what they are for.

Expect layering to survive

The median frame holds seven distinguishable shadow populations at a measured floor of ΔE00 0.67. Eight bits is not the constraint people assume at the bottom of the range — the grade is.

• If two dark materials have merged, they merged in the grade or in the lighting, not in the encode.
• Three quarters of that separation is lightness. A curve that flattens the bottom twenty L* removes most of it.

Do not read a look off the toe

A film's toe shape carries essentially no film-level information. Matching a reference by matching its toe is matching one shot of it.

• What does travel is the floor decision and the shadow density — whether zero is allowed, and how many levels are spent above it.
• Those are worth deciding once, at the start, and holding.

One consequence for how these numbers should be quoted. Because shadow geometry is a shot property, a film-level average of it describes nothing. A summary saying a film's dark mass is 40% fragmented is averaging over frames that run from 0.00 to 0.93. Report the dispersion or do not report the measure.

Nine properties, each measurable on a rendered frame, and two words to retire.

Crushed and lifted are one axis for at least nine separable properties. Splitting them makes it possible to say which one a picture is short of, and each can be measured on a rendered frame without any claim about a transfer curve.

Two words to retire in technical contexts. Crushed, unless the speaker says which crush — floor contact or floor placement — because one is irreversible and the other is a display setting away from being visible. And rich, which in this corpus turns out to name at least three unrelated things: a low floor, many surviving levels, and a large subject-to-surround difference. Films can have any one without the others.

Black geometry earns its place as an umbrella only if it is understood to be an umbrella. The thing it covers is not one property of a film. It is a set of shot-by-shot spatial decisions taken inside two film-level constraints — whether zero is allowed, and how many levels are spent above it.

Every figure quoted in the prose above is declared in the report's source and re-verified against the measurement files at build time; a number that drifts fails the build rather than surviving in the text. The charts read the same measurement files directly, so no chart can disagree with the sentence beside it without the build noticing.

The largest limitation is the pixel tier. Complete frame-by-frame coverage exists for seven films because their Resolve exports happen to be on local disk; for the other thirty-six, pixel access is limited to curated frames whose displacement from their films was measured in section 02. Mounting the stills volume and re-running black_report/tools/measure_frames.py would extend every census-tier result to the whole corpus, and figure 12 is the chart most likely to move.

Three further limitations are structural. No perceptual experiment was run, so nothing here establishes what observers see; the face detector is a machine probe used for the difference between two runs and not as a model of a viewer. The compression-artefact question the brief raises — banding, macroblocking, chroma noise in graded darkness — is not answered here: separating codec failure from aesthetic shadow compression needs the delivery encodes rather than the frame exports, which this corpus does not hold. And every frame is an eight-bit display-referred master, so the study describes deliveries rather than negatives, sensors or grades.

• Corpus: output_*/shots.parquet, 43 films, 52,527 analysis-eligible shots. Eligibility is analysis_eligible where the extraction carries it and the published lab's picture range otherwise; the rule used is recorded per film in data/films.json.
• Census tier: the 1920×1080 Resolve exports under stills/, cropped to each film's measured active picture, for Amélie, A Very Long Engagement, Delicatessen, Kwaidan, Sisu, The City of Lost Children and Zatoichi.
• Curated tier: the lossless frame assets published by the slide labs, byte-identical to their audited sources.
• Colour: encoded RGB decoded with each film's recorded transfer (gamma 2.4, Rec. 709 or PQ anchored at 203 cd/m²) and primaries, then CIELAB under D65. The two PQ masters are excluded from every comparison made on a lightness scale.
• Readability probe: OpenCV's haarcascade_frontalface_default, run on the delivered grey and again after a fixed gamma-0.45 LUT, at a common 640-pixel width.
• Reproduce: tools/corpus_manifest.py --out black_report/data (shared with the highlight report), then black_report/tools/corpus_scalars.py, measure_frames.py, dark_census.py, analysis.py and render_frames.py, then black_report/build.py.
• Cross-study figures read white_report/data/analysis.json directly; if the highlight report has not been built, that section is omitted rather than estimated.