Why we pivoted then
D2 is closed for one free alphabet: F1.4's one-alphabet residue (0253-0255) and F1.7 (0256) were decided cell by cell with the one-run decider and closed at their measured power, and D1b (0257) measured the decider for two free alphabets. D3's listing (2026-10-29) left one cell group to run, F13.4c at key lengths 6-19, and F13.4h (Alexis's call to chase 0237-0239's excess, #101) is the next open F13 row. Both are F13 work.
What happened before
- 0253 looked like a lead: K4 above all 12 of its copies in 299 of 2,093 key schedules where a fair test gives 161. 0254 and 0255 put shuffles of K4 through the identical test as stand-ins: they average 10.8% with a spread of 6.5 points, and K4 (10.0%) sits 21st of 41. Each text has its own rate, shared by every setting it is tested in, because all settings read the same 24 known places. A count of settings or cells where K4 beats its copies has to be judged against stand-ins, not a binomial.
- 0256 wrote plaintext autokey's chains in full and decided all 130 cells K4 fits; K4 sat inside 10 stand-ins, at low power (plants 12 of 48). F1.13 waits for a better autokey climb.
The change we believe is needed
Every experiment that counts K4's wins over many settings now carries stand-ins (shuffles of K4 put through the whole procedure). F13.4c at 6-19 ranks K4 per cell against 100 fresh shuffles climbed the same way, which is already a comparison over texts. F13.4h measures whether the B2 excess lives in width-19 reads and in the carved order itself.