Why we pivoted then
After 118 experiments nothing had beaten the scrambled copies of the carved text, so there was no result to follow up. The work had become a sweep, and the switching between families looked unsystematic: each next experiment was picked from whatever the last one suggested. Alexis asked for a plan (issue #29).
What happened before
- 0089-0096 worked through types 3 to 8 of the directions: keys read off the sculpture, the Berlin clock with every move, messages written along grid paths, a carving error, riddle words. All at chance or ruled out.
- 0096 corrected 0087: a column word plus a row key does not line doublets up by column. Any key of period 3, 7 or 21 after a move makes the width-21 repeats.
- 0100 found one setting above the scrambled copies (an unknown alphabet pair, key length 3); 0102 showed it was chance (p 0.043 against 400 copies, 0.10 against copies that keep the width-21 columns).
- 0105-0118 closed Hill with a periodic key, the dictionary-keyed hand moves (columnar, Myszkowski, AMSCO, rail fence) and the remaining running-key ideas.
- A grading of all 118 found 49 ruled out, 62 at chance, 6 with no comparison and one statistic beyond chance (0056, a pattern in the carved text, not a reading).
The change we believe is needed
Work the families in a fixed rank order and leave a family only when every item in it is closed or blocked. Every experiment names the item it works on and the strength of its test. A result outside the scrambled-copy range stops the queue until it is replicated. The first family is F1, a single keyed stage, starting with unknown alphabets (F1.1), the largest axis the earlier sweeps left untested.