How to determine the best exposure for raw files

Nope not right. Depends on the conditions.

The in-camera histogram loses spatial information (it doesn't show where the clipping is), zebras lose information on distribution/light intensity. Both are based on jpeg/rgb processed image, distorted against raw because of applied white balance. You seem to fail to understand that white balance skews original raw rgb channels.

RawDigger shows more accurate information on raw data, both histogram and spatial. Again it looks like you don't understand the difference between in-camera RGB histogram and raw histogram in RawDigger.

If RawDigger shows clipping, there's no way to recover the blown areas.
If RGB in-camera histogram or zebras shows clipping, there may still be room for recovery if you shoot raw.

That's why there's less information in in-camera histogram or zebras than in a raw histogram. Unfortunately there's no camera that shows a raw histogram as far as I know. At least it's neither of Sony cameras. The lack of raw histogram can be compensated by different techniques discussed in this thread but they all have caveats.

I made no such error. I used alternate words this time as you seemed to have some struggle with reading comprehension the first time. The original wording - “largely” invariant, meaning: to some extent, but not entirely invariant, was meant to succinctly get the point across that, if you are a RAW shooter, there may be options available that will allow you to, in some situations, make good use of the sometimes ISO invariant nature of some sensors to use a lower than normal ISO value in-camera, and apply any necessary “brightening” in post instead, dear.

If RD shows clipping in one channel, ACR and Lr can often do quite a credible job reconstructing the highlights. But the photographer's objective should be to avoid the need for that.

If we would only know what sacrifices are made for the lower gain ;)

I'd argue that reconstruction is not exactly the same as recovery (that is, just changed tone mapping) - reconstruction still implies the loss of information.

I don't disagree, but someone could read what you originally wrote and think that one blown channel will have a really bad effect on the resulting image. .

For some Canons, a modified firmware will do this:
magiclantern.fm
It's very strange that we still need to discuss crutches and compare workarounds to know clipping while shooting raw.

I know, ML seems to be abandoned since circa 2018 :( There's no support for Canon 5DIV and later.
In my R5, I don't even have live zebras (only in post-shooting previews)...

True, one channel can be fine to blow - I meant unrecoverable information, but it can be compensated to some extent, yes. From the experience it's usually ok to blow the red one in sunset sky. It's bad if it's green.

IMHO very low activity since at least 2020. I keep 5DMkIII, a rather competent camera.

Intriguingly I really like this picture (perhaps with slightly lighter processing), well done Erik.

was there ever a camera invariant all over the range? maybe an early sigma foveon?

Especially the ones that applied no variable gain at all to the raw signal and counted on the raw converter to do the push.

I don't think that that is a helpful way of looking at the problem.

Invariant or not invariant at a particular ISO setting is - to some extent - a matter of judgement. Like if I showed you a yellowy-green Pantone color card, and asked you whether you you would call it yellow, or green, if forced to make a choice.

The question is: At a particular ISO setting, does (amplified) read noise dominate conversion noise?
A similar question is: At a particular ISO setting, is output noise significantly different from (amplified) read noise?

Those are both slightly fuzzy questions.

If you're OK-ish with a noise model where amplified read noise adds in quadrature to conversion noise, then in black areas of the image we have:

(Recorded noise, in DN) = sqrt( (conversion noise, in DN)^2 + ( gain * (output referred read noise, in DN) )^2 )

Where "DN" means "Digital Numbers" - the numbers stored in a (lossless) RAW file.

Examples:
- If conversion noise is 1DN, and gain * output referred read noise is 1DN, recorded noise is sqrt(2)=1.414DN
- If conversion noise is 1DN, and gain * output referred read noise is 2DN, recorded noise is sqrt(5)=2.236DN
- If conversion noise is 1DN, and gain * output referred read noise is 3DN, recorded noise is sqrt(10)=3.162DN
...
As the gain increases, conversion noise has less and less effect, and the recorded noise becomes - arbitrarily - close to gain*(output referred read noise), regardless of what the conversion noise was at base ISO.

To me, a camera is "ISOless" at a particular ISO if the recorded noise is - for practical purposes - equal to gain*(output referred read noise).

I just pulled up a noise-model spreadsheet I have, that says "ISOless" ("invariant") or "not ISOless" for a camera noise model, at different ISOs.

It seems that it says "ISOless" if gain*(output referred read noise) is more than 90% of the recorded noise (in black areas of the image, remember).

I could have chosen some other number, like 80% or 75%. But 10% or 30% would be silly.

...

The Aptina dual-gain thing has the effect of reducing read noise at (or above) the gain setting at which it turns on. Perhaps enough to make conversion noise significant again, so that the camera goes from being "ISOless" to "not ISOless" until the gain has increased sufficiently further.

We could have a camera that behaves like this:
- ISO 100: not ISOless
- ISO 200: ISOless
- ISO 400: ISOless
- ISO 800: not ISOless (Aptina thing kicks in).
- ISO 1600: ISOless
- ISO 3200: ISOless
- ISO 6400: ISOless

That would mean that ISO 400, ISO 3200, and ISO6400 could be a bit pointless when capturing RAW images.

But the main thing is usually to try to capture as many photons as you can.

...

All uses of "ISO" above belong in scare quotes.

Both equations are fixed now.

Allan or Bob: now that this thread now seems to have the Sony "a74" (whatever that is) as the gold standard for determining the best exposure for raw files, should it be now moved to the "Sony Camera" category ?!

As an owner of a camera that shoots and outputs raw only, I was initially interested ... but all this hot air about zebras and specific Sony settings is way off-topic, grump.