Same numbers. Different result.

Start with independent questions. Which primaries and white? Which transfer function? RGB or Y′CbCr? Which matrix? How are normalised values mapped to codes? How many bits, and what is the component grid? Only then ask which codec and container store it. This is an inspection map, not a mandatory sequence of operations inside every encoder.

“10-bit ProRes” does not specify gamut, curve or whether a LUT was baked in. A MOV extension does not answer those questions either. Compare shooting notes, stream properties and decoded output. If several layers are called a “profile”, establish whether the word refers to a codec, colour definition or camera mode.

RGB and Y′CbCr represent image components differently. Conventional NCL Y′CbCr is calculated from nonlinear R′G′B′. With full sampling, a matching matrix and no rounding, the conversion is reversible; it need not degrade the image on its own.

Losses arise from clipping, quantisation, subsampling and compression. These often accompany a representation change in practice, so their effects get incorrectly attributed to the name Y′CbCr. An RGB file can likewise be heavily compressed or low precision. Compare actual properties and the next processing task, not merely the representation name.

Matrix coefficients define how to derive Y′ and colour differences from R′G′B′ and reconstruct them. BT.709 uses weights 0.2126 / 0.7152 / 0.0722; BT.601 uses 0.299 / 0.587 / 0.114; BT.2020 NCL uses 0.2627 / 0.6780 / 0.0593. A wrong inverse matrix alters coloured regions differently.

Changing coefficients from 709 to 2020 does not convert RGB primaries into another colour space. That requires a separate colour transform. Primaries and matrix are different metadata fields because they describe different operations. Resolution alone cannot establish the matrix: resized older material may retain its original encoding.

In nominal 8-bit video-range Y′, black is code 16 and white is 235; in 10-bit they are 64 and 940. Chroma uses a different span: 16–240 centred on 128, or 64–960 centred on 512. Applying the Y′ span to all three components introduces additional colour errors.

Nominal endpoints do not make every outside value meaningless or require immediate clipping. Excursion handling and reserved codes depend on the transport and format. In the lab, Full for Y′/RGB maps 0…1 to 0…2ⁿ−1. Full-range chroma must follow its particular format definition rather than inherit the luma formula.

When Video codes are read as Full, nominal black becomes positive and white falls short of one, often appearing washed out. Reading Full as Video instead reconstructs values below zero and above one. If a later stage clips them, endpoint detail disappears.

This is an affine change of code scale and offset, not a new gamma, white balance or primary set. The lab retains out-of-range values in its readouts and clips only the screen swatch. Appearance alone is not a diagnosis: a soft look can resemble the error. Confirm it using metadata, a known ramp and re-import.

n bits provide 2ⁿ possible integer words: 256, 1024 and 4096 for 8, 10 and 12 bits. The range determines where nominal signal values sit within those words. Codes 64…940, for example, contain 877 values but 876 intervals between endpoints; do not confuse these counts when computing a step.

Additional bits reduce quantisation error for a fixed range. They do not expand sensor headroom or the primary triangle. Converting 8 to 10 bits permits finer subsequent arithmetic but cannot recreate distinctions already merged into one source code. Storage depth is also not necessarily the effective precision of the measurement.

In 4:4:4, Y′, Cb and Cr have equal spatial sample counts. In 4:2:2, each chroma plane has half as many horizontal samples. Common progressive 4:2:0 halves both chroma dimensions. The luma grid need not shrink.

This does not mean 4:2:0 “has no blue channel”: both Cb and Cr remain, just at lower density. Fewer component samples save data before entropy coding. For single-pixel coloured lines this is a different limitation from 8-bit steps on a gradient. Interlaced material additionally requires correct field handling.

Grid density does not specify chroma position relative to Y′. A sample may be co-sited with luma or lie between luma samples. Full-grid reconstruction needs the correct phase. A wrong assumption shifts a colour boundary even when sample counts and depth agree.

Resizing and transcoding therefore need compatible filters, chroma positions and, where relevant, field handling. Metadata may be absent, leaving format conventions and documented decoder assumptions important. The lab explicitly uses centred blocks and a simple reconstruction rather than claiming to reproduce every codec’s 4:2:0 behaviour.

A codec defines coding and compression methods; a profile constrains available tools and representations. A container combines streams, timing information and metadata. Neither MOV nor MP4 is itself a colour space. Files sharing an extension can contain very different video streams.

Even within one codec family, sampling, depth, alpha support and loss behaviour differ. Apple’s documentation distinguishes ProRes 422 and 4444 variants; these are not one identical quality at different sizes. Visually lossless does not mean mathematically identical either. For important intermediate processing, inspect the exact profile and decoded result.

A stream can signal primaries, transfer, matrix, range and chroma location, while its container may carry additional information. These guide interpretation but do not prove that the encoded samples actually follow them. Fields can be missing, wrong or contradictory.

Changing a tag does not perform a colour conversion. The range lab makes this explicit: identical codes yield different normalised values under different interpretations. Correct assignment fixes an incorrect description; assigning a desired label does not convert the material into that standard. Keep the source and document which values were actually transformed.

Compare source properties with decoding first. Range, matrix and transfer mismatches produce different classes of error. Data Levels controls numeric-range interpretation; Input Color Space describes the colour state. They are not interchangeable “fix image” buttons.

For a suspect clip create a test version without creative nodes and with known colour management. Compare source and intermediate on identical frames, checking neutral ramps and coloured regions. A change that helps only one object may hide the cause rather than remove it. Automatic detection is useful but does not replace verification of ambiguous media.

Inspect the rendered file itself: dimensions, pixel format, profile, colour fields and frame rate. Then re-import it into a controlled project and compare it with the timeline under matched viewing. The first check reads file descriptions; the second exposes decoding consequences. Neither alone certifies the entire delivery.

For external inspection, ffprobe uses -select_streams v:0 to select the first video stream, -show_entries to choose fields and -of json for structured output. A missing or unknown field indicates missing information, not automatically Rec.709. Do not deliver a “fixed” file solely because one player makes it look more pleasing.