> Both dosings use two shots, and the "90%" number from the smaller dosing is from an unreliably small sample set of 33 infections."

Can you count infections as a metric of efficacy for a vaccine that's designed to stop infection?

Consider [1]:

"However, protection was 90% in an analysis of around 3,000 people on the trial who were given a half-sized first dose and a full-sized second dose."

AFAIK it's not known how many of those 3,000 in that trial were exposed to potential infection.

[1] https://www.bbc.com/news/health-55040635

Couldn’t the same be said in reverse? I.e., due to sampling error the reports could be amended to a 95% efficacy when more data comes in?

I don't think we know based on public information. Some media report "up to 90%" which suggests the upper bound of the confidence interval, rather than the centre. Could also be a very wide CI, that overlaps with the standard dosage.

That's a fairer statement than the original which seemed biased towards only reporting the lower end of the interval without acknowledging the upper bound.

At the moment possibly wider error bars don’t come from just “33 infections” but more from possibly smaller number of those who received the observed treatment.

The bars are to be calculated using conditional probabilities.