Oh, what is raw?

I used to work around telemetry from space and missile launches, any departure from the recorded stream off the sensor was considered "processed". Any of the above would be "processing". That distinction was particularly relevant in failure investigations, in order to establish the pedigree of fault assertions.

Here in photography-land, we can spend all day discussing lines in the sand, but the only useful one is what the photographer is presented by the camera. That's the starting point for any discretion one has in how the rendition will look. One has no control over what the camera manufacturer might want to cook prior to that. So, IMHO calling it "raw" is more a convention than a formal definition...

Unless the photographer is curious, or needs / wants more. Sony were pushed to offer lossless raw, Nikon - to stop subtracting black level.

Quite so. Only it isn't understood as a convention.

It is "RAW" ;)

Optionally? How do you get a JPEG without first having raw sensor data?

It's been many years since I've seen a new camera with raw files that didn't have spikes and gaps in the histogram from integer math. Even on those cameras, shooting at f/1.4 probably resulted in pushes that created histogram gaps and loss of DR due to the light loss of such optics with FSI sensors, but I wasn't aware of that behavior at the time that I was seeing perfect histograms (perfect except for mapped-out pixels, of course).

Speaking of mapped-out pixels, the Canon EOS 10D at ISO 3200 actually showed about half of the mapped-out pixels as a side effect of mapping them out after an arithmetic push, because the camera made the LSB of every horizontal line either 1-1-1-1..., 0-0-0-0..., or 0-1-0-1, so any pixel that was interpolated because it was bad had a 50% chance of falling outside of those consistent horizontal patterns.

Here's a 100% crop of the LSB (least significant bit in 12-bit raw):

What the *** is LSB?

IIRC, the whole pattern of analog gain and digitization on the 10D was that ISOs 100, 200, 400, and 800 had analog gain proportional to ISO setting, and ISO 800's analog gain was used for 1600 and 3200. At 1600, it was left in the original digitization, much like later Canons sported HTP at ISO 1600, but for 3200, they did a 1-stop arithmetic push and rather than having just only even numbers, they seemed to be trying to correct some horizontal banding read noise by adding/subtracting a mean of 0.5 or 1.0DN to some lines. Apparently, the mapped-out pixels were interpolated after the extra 2x "digital" gain, and that's why half of them leave those artifacts (the other half not falling outside the pattern).

Now I think you are thinking something like Plato's Idea of the picture you are taking.
I guess you have same kind of an idea about what you are looking and shooting. Like a Point and shoot Shooter?
The problem is Your camera do not materialize your idea automatically. It makes a "raw file".

In practice, the raw is something the sensor maker and the camera maker have "cooked in" or "baked in", something that breaks your idealism. You don't know what did happen, because sensors are not perfect, cameras are not perfect and you are not a perfect photographer*.

Is there any difference what is raw for a photographer and what it is for a program developer?

Is it possible to know what is raw?

Should I know what raw is for a photographer?

• According to Plato the Ideas are the only prfect things in the world. (Look at my typo!)

As I already wrote in the post you replied to, it stands for "least significant bit". We count to 7 digitally, like so:

0 = 000
1 = 001
2 = 010
3 = 011
4 = 100
5 = 101
6 = 110
7 = 111
      ^
      L
      S
      B

The LSB is the digit on the right, which determines if the number is odd or even in base 10.

So, I made an image that is one-bit from the least significant bit in the 10D ISO 3200 raw. Zero stayed zero or black, and "one" became white or 255.

I don't recall exactly why I decided to visualize a single bit from a 12-bit file; maybe I wanted to see how much pixel-level variance there was, but found this total artifice instead. I was shocked when it appeared on the screen, having expected to see an LSB shaped almost completely by noise.

I think that photography is a recording left by light, not a recording of light ;)

We could define what a raw file is by its benefits:
- the white balance and color profiles are not included in the data
- for color sensors, the channels are still separate (better highlight recovery)
- maximum possible data precision available (considering all possible workflows)

By that criteria, a linear DNG is a raw file, a file with applied shading compensation (Sony) is a raw file and a file with applied heavy NR (Pentax) is a raw file.

There are tonnes of versions of "raw" (according to manufacturer's naming) that have white balance applied to data, channels are YCC or some other form of non-Bayer scheme (while the sensor is Bayer), and linear DNG is a demosaicked format (channels are not separate). "RAW" is a meaningless marketing word.

Ok, perhaps the sentence is a little ambiguous or unclear.

Obviously the camera needs the raw data to output a jpeg.

What I meant was raw in layman's terms is the dataset the camera uses to output a jpeg image file when jpeg is set as the or one of the camera outputs - raw + jpeg, jpeg.

Linear DNG created by, e.g., Adobe's Denoise contains RGBG2 channels according to RawDigger (same probably for DxO's noise reduction software). What am I missing?

Well, is it linear DNG then?
"Linear (demosaiced): The image data is stored in an interpolated ("demosaiced") format. This option is useful if a camera's particular mosaic pattern is not supported by a DNG reader. The default mosaic format maximizes the extent of data preserved. Mosaic image data can be converted to linear data, but the reverse is not possible."

From all the information I can find, Adobe's Denoise and Super Resolution create a DNG referred to as a linear DNG. DxO writes that they create linear DNGs. Interestingly, DxO's output has only three channels (RGB), while Adobe's output has four channels (RGBG2).

Please check tags 0x0106 ("PhotometricInterpretation") in all IFDs:
$ exiftool -H -S -G -a -ee -PhotometricInterpretation IMGP1498.DNG

Here is from a regular DNG file:
[EXIF] 0x0106 PhotometricInterpretation: RGB
[EXIF] 0x0106 PhotometricInterpretation: Color Filter Array
[EXIF] 0x0106 PhotometricInterpretation: YCbCr

Here is from a linear DNG:
[EXIF] 0x0106 PhotometricInterpretation: RGB
[EXIF] 0x0106 PhotometricInterpretation: Linear Raw
[EXIF] 0x0106 PhotometricInterpretation: YCbCr
[EXIF] 0x0106 PhotometricInterpretation: Linear Raw
[EXIF] 0x0106 PhotometricInterpretation: Linear Raw
[EXIF] 0x0106 PhotometricInterpretation: Linear Raw

Whatever they call it, as long as it approximates the energy relationship of the scene's light I'm okay. Oh, and contains the metadata needed to interpret it for my own cooking... which it usually doesn't, color matrix for example.