Why we pivoted then
Every item in F1's table is now closed or blocked. P6 reopened F1 to work on the shared blocker of the free-alphabet cells, no decider where the known letters leave too few checks (F1.5), and the pivot audit and the field-manual sources added F1.6-F1.11. All of them have now run. F6 is second in the ranking, which is unchanged.
What happened before
- 0136 bounded every blocked cell: the harness's quadgram scorer extracts 1.78 bits per letter of English, so 73 unknown letters carry about 130 bits. One free alphabet at every length, two free alphabets to length 16, and Hill with any key are decidable in principle. Two alphabets from length 19 and plaintext autokey at lags 53-84 need a better English model; lags 85-96 no English test can decide.
- 0137-0139: design 1, a climb that keeps every known-letter check, seemed to have power (planted texts beat K4's scrambled copies 10 of 12), but 0138 compared each planted text with its own scrambled copies and the power fell to 6 of 12: most of it had come from letter counts. Design 2, parallel tempering in the same space, passed that like-for-like bar (6 of 8).
- 0140-0146: design 2 decided F1.1 (Beaufort 23 and 26) and F1.2 (lengths 8, 13, 16); K4 sat inside its own scrambled copies every time. 0141 was an error (control generation too slow).
- 0147-0152: one carved letter wrong, missing or extra, through every exact free-alphabet verdict: at the copies' rate, or dead; the one cell revived by meaning (K1 and K2's system at length 19 with the S at 33 a slip) was too weak to judge (0148), though one planted text there was read back whole.
- 0149, 0154-0157: autokey on a free alphabet, Carter as a running key on a free alphabet, progressive keys on a million tables, Friedman's key-advance schedules, split key phrases: dead or at the copies' rate. 0153 was an error (a loop without a clock check).
The change we believe is needed
- F6 next, in table order: F6.3b (mixed triples at wide widths, compiled), F6.5 (compass directions as routes), then F6.6-F6.19 from the audit and the manuals, grouped where one route library serves several items.
- F6.1 and F6.2 (free alphabets after a move) are decidable in principle and a decider now exists, but tens of thousands of fitting (move, length) settings at 20-30 minutes each put them beyond any run: they stay blocked, now by cost rather than by principle, with the numbers in their rows.
- Lesson carried forward: a decider's power is measured against each planted text's own scrambled copies, never against the target's.