10(log10 143 - log10 106) gives the decibanage in favour of a theory
for each occurences of a J. Each pair of letters can be decibanned in
exactly the same way. Now we can take (say) the count for X3 = 01
as standard, and the other two counts can be decibanned up or down
according to the excess or defect of the good letters over those in
the 01 count.
A warning should be given about the decibanning of pairs
of letters that go the wrong way; unless this feature can be satisfactorily
explained, the scores should not be accepted at their face value. No
satisfactory way of allowing for this is to hand : it probably be
safest to leave out such pairs.
When the total decibannages have been added up, the relative
probabilities of the 3 settings can be irnmediately stated.
Summary of Techniques.
(i) For long runs. - 3.5 sigma is borderline, 4 sigma is liable to be
about 3:1 on, 4.5 sigma is generably reliable.
(ii) For short runs, -2.5 sigma is borderline, 3 sigma is liable to be
about 3:1 on, 3.5 sigma is generally reliable.
(iii) If a long run gives settings with probability P, and these are
used to get further settings, the odds in favour of the further
settings (worked out from the chart) ?comment? P gives the odds in favour
of the wholwestory.
If a run based on other settings fails, those settings lose a
factor of 2.
(iv) Final counts must always be looked over
(a) to check individual impulses.
(b) to see that there are no abnormal and inexplicable bulges.
For X3, 11 of the 16 pairs should go the right way
For X5, 8 of the selected pairs should go the right way.
Slide rivals can be decibanned from the full counts.