Image values that describe the light that was in front of the camera, as opposed to display-referred values, which describe the light a particular kind of screen should emit.
Camera log files and ACES frames are scene-referred; a Rec. 709 master or a PQ deliverable is display-referred. An output transform converts the first into the second, compressing the scene's range into the display's. ITU-R BT.2100 shows both kinds in one standard: PQ is defined by a function from signal to display light, HLG by a function from relative scene light to signal.
Scene-referred does not mean 'accurate' or 'ungraded'. It describes what the numbers refer to, and a creative grade can be carried out entirely on scene-referred data before the output transform.
One scene-referred grade can feed a cinema output transform, an SDR television one and an HDR one, each squeezing the same scene values into a different display range.
Work done on the scene side of the output transform survives a change of display; work done after it is tied to one kind of screen and has to be redone for the next.
Whether image data is scene-referred is a property of its encoding, read from the specification and not from the pixels. ITU-R BT.2100 gives a clear pair. It defines PQ by a reference function that maps the signal to the light a display emits, which makes PQ display-referred. It defines HLG by a reference function that maps relative scene light to the signal, which makes HLG scene-referred. The ACES documentation describes its core as a standardised scene-referred encoding, with input transforms bringing camera data into it and output transforms preparing it for a display.
A film is graded after the output transform, on values already made for one display.
The grade is tied to that display's range. One camera maker's pipeline note advises that all grading take place before the output transform, which it describes as the step that gives monitor independence.
Display-referred graphics or archive footage are cut into a scene-referred timeline unconverted.
The output transform is applied to material that has already been through one, so its contrast and saturation are altered a second time. Such material has to be converted into the working encoding first.