The error bars in those charts are showing about ±1 million km² uncertainty. I think that's extremely accurate considering the available data. It's about twice the error bar for the satellite age.

We have little ice data, but I think we should take the many reports from amazed fishermen, trappers and scientists of the 20's seriously.

*picture*

Approximately twice that of the satellite age. I don't know, that does not sound too unreasonable to me. There is some error associated even with satellite observations (precisely determining areal extent is more difficult than you might imagine). And the available observational/proxy data is perhaps better than you might think. But to be fair, I am not an expert in sea ice extent reconstructions or the uncertainty associated with them. Are you? Do you have a quantitative assessment for why the uncertainty bars should be of different size, or know of studies that yield strongly contrasting results?

I think I can see what you are trying to show, though it's not very clear and I don't have a clue what the source figures were. I'm not even sure which report they came from. Can you please help me out by posting the originals? I've been trying to avoid any confusion by giving clear citations, let's please try to do that both ways.

basically a sanity check that the models are not just going crazy when run forward

Now with some relief we've finally resolved your questions about the error bars and lack of response to the internal variability during the 20s. In looking at the correct figures, you see exactly what you intuited.

So when the models fail to reproduce in the past a known physical property, we shouldn't have very high expectations for them in the future.

we're not looking for a model which will show us exactly what the ice sheet extent will be in the year 2047

That wasn't really what was in question here, but rather within what range we should expect the extent to be in a year like 2047. I'm addressing the uncertainty.

So when the models fail to reproduce in the past a known physical property, we shouldn't have very high expectations for them in the future. The extra JAS sensitivity surely doesn't only apply to natural variability, but surely also oce-atmosphere interactions.

Do not confuse the effects from natural variability for the effects from the underlying global warming trend, which is what the models are looking at.

(a) A 2007 prediction of ice-free summers by 2013 has clearly failed

When we do so, we see that the reconstructions do show the natural variability that you were pointing out, and the relative sizes of the error bars make sense, too.

There is nothing insightful to be gained here by using the wrong figures for the wrong discussion points.

Does it really matter? As long as you can accurately predict long-term trends then what does it matter if you know the uncertainty at any given point or not? In this particular case I don't see what practical difference it makes, aside from being able to check when we arrive at that point in time whether the observed data is within the range of the prediction, and by that time the current models will be obsolete anyway.

Yes, since it's a corner stone of science that any claim must have clear ways of disproving it. That you already know. There can be no exceptions even if we expect the models to become obsoleted. The more permissive a model is in being proven correct, the less its scientific value is.

The discussion that I invite is whether that certainty about the 20's is credible in face of the accounts available and the lack of them.

A projection only based on simple ice-atmosphere interaction may already be on this ice in this respect (sorry for the pun) in assuming that all else remains the same. But in atmospheric science there is very little "all else remaining the same". It's a difficult task to extract the right signal and perform verification.

I would very strongly recommend a reading of AR5 Section 12.4.6.1 here... may I ask that you please do so before we continue?

No, because validation does not require prediction of exact values. Probability distributions and correlation functions are standard scientific principles that can be and are applied here.

The long term trend has been that of loss so you expect a priori the satellite era to be lower than the early 20th century on average.

The sea ice extent is not going to completely recover to levels from 50 years prior by just one decade of cooler temperatures.

it reproduces the historical sea ice extent extremely well, minus the differences that it does not account for -- natural variability.

The CMIP5 model simulations exhibit an ensemble-mean polar amplified warming, especially in winter, similar to CMIP3 model simulations (Bracegirdle and Stephenson, 2012; see also Box 5.1). For RCP4.5, ensemble-mean winter warming rises to 4.6°C over pan-Arctic land areas by the end of the 21st century (2081–2100), and about 7.4°C over the Arctic (Table14.1). Throughout the century, the warming exceeds simulated estimates of internal variability (Figure AI.8).
...
In summary: It is likely Arctic surface temperature changes will be strongly influenced by anthropogenic forcing over the coming decades dominating natural variability such as induced by NAO.