“F1.12: F1.1 Beaufort 23 and 26 and F1.2 at 13, K4 against 100 of its own copies with the one-run decider, then MIX8m and the blind judge as second deciders”
F1.12 (Alexis, 2026-10-28, #100: close at stated power and still rerun, so a missed lead gets a second chance). The three cells closed at measured power on 2026-10-28 had K4 ranked among 12 copies only (9th, 12th and 13th of 13). Here each cell's K4 is ranked among 100 of its own copies, each climbed with the one-run decider (D1: one tempering run, 0140's climb for one alphabet; 0144's for two, with as many runs as D1b qualifies); the best readings then rescored by MIX8m (0233) and read by the blind judge, K4 and copies together. 8 plants per cell in the same run.
Three versions of the cipher table with a scrambled alphabet were closed after K4the carved letters The 97 letters cut into the copper at the CIA, shown at the top of every experiment page. read no better than 12 scrambled copiesa shuffled control The same experiment run on the carved letters shuffled into a random order. of itself. Twelve is a small crowd. Here K4 faces 100 copies in each, and two more readers judge the results: a stronger statistical model of English, and a language model that reads for words.
Words in the margin are defined on the glossary page.
1.1
The method at work
The carved letters go in at the top; the reading comes out at the bottom.
the idea
F1.12: three cells closed at measured power on 2026-10-28 with K4 ranked among only 12 copies; rerun against 100, with two more readers.
the check
One-search decider (four searches for two alphabets, D1b); MIX8m and the blind judge read the best readings, K4 and copies together (303 judge calls).
the bound
K4 is never above more than 90 of its 100 copies on any reader; planted messages found 7, 5 and 5 of 8. Closed.
2
Experiment lineage
What this experiment was built on, from the lineage page: 17 experiments before it, back to the root of its line. Its number is in bold.
keptreverted, in its family's colour×error or timeoutreverted under today's rulesbuilt on a parent in the same familybuilt on a parent in another family
3
Best candidate against the known letters
This experiment produced no candidate: the search ruled everything out before a full text was built. Its product is the bounded negative under Result and bounds below.
4
Walkthrough, by hand
Redo the method on a few letters with pencil and paper.
Take one cell, say a scrambled alphabet on both sides of the table with a key of 23 letters, Beaufort form.
Search for the alphabet and key that make K4 read most like English while keeping the 24 known letters.
Make 100 scrambled copies of K4 that fit the same 24 letters equally well and search each the same way.
Rank K4's best reading among the 100 copies' by the search's score, by a stronger English model, and by a language model reading for words.
A real answer should come out on top on all three.
the worksheet: the first steps by hand, on squared paper
5
Worked example, generated
The same steps run by the code on the real letters.
F1.12: three cells, K4 against 100 of its own copies, three deciders.
cell alphabets key length checks
F1.1 B23 one 23 7
F1.1 B26 one 26 1
F1.2 13 two 13 11
each text: one tempering run keeping the 24 known letters -> best reading
rank K4 among its copies by quadgrams, by MIX8m, by the blind judge.
6
Hypothesis, for a cryptanalyst
Family, key space, the constraint under test, and the clues it reads.
F1.1 (Quagmire III with an unknown order, Beaufort form) at key lengths 23 (7 checks) and 26 (1 check), and F1.2 (Quagmire IV, two unknown alphabets) at 13 (11 checks): the known letters leave thousands of alphabets, so English decides. Design 2 (0139-0146) put K4 9th, 12th and 13th of 13; D1 (0228, 0229) and A6.2 (0226) measured power on clean and Sanborn-style plants: L23 33/40 and 32/48, L26 13/24 and 17/24, L13 7/10 and 17/24. With 100 copies K4's rank is measured to 1%; MIX8m (1.95 bits per letter against the quadgrams' 1.75) and the judge (which reads telegraphese and abbreviated English, 0251) are second deciders that do not share the climb's score. Clue entries C27(c), C5(a); A6, A13.
7
Prediction
Written before the run; the result decides it.
K4 inside its copies in all three cells on all three deciders: rank by the climb's score within the middle 90%, MIX8m and judge ranks likewise; no cell with K4 above all 100. Plants found about as D1 measured (5-7 of 8).
Copies: one alphabet, shuffles of K4 that 0120's exact decider fits at that length (as 0140); two alphabets, kept-place copies (the carved letters at the 24 known places kept, the other 73 shuffled, as 0144), started from K4's witness.
Every text climbed with the one-run decider (RUNS_ONE runs of 0140's tempering; RUNS_TWO of 0144's), its best reading kept.
Deciders: the climb's quadgram score; MIX8m on the reading; the blind judge on K4's and the 100 copies' readings in one call per cell. K4's rank among its copies on each.
Plants: 4 clean and 4 Sanborn-style per cell, 6 own copies, through the same climb; found when above all own copies.
9
Result and bounds
The harness's record: candidates, gates, loss, and the negative result this experiment leaves behind.
0candidates returned
0passed every gate
0/24known letters, best
–lowest loss
nonechampion before
F1.12: 1.1-B23: K4 quadgram 68/100 at or above, MIX8m 55, judge 46 (K4 12), plants 7/8; 1.1-B26: K4 quadgram 70/100 at or above, MIX8m 10, judge 50 (K4 12), plants 5/8; 1.2-13: K4 quadgram 100/100 at or above, MIX8m 97, judge 86 (K4 8), plants 5/8
2 x (101 + 56) one-run climbs (about 17 s) and 157 four-run climbs (about 70 s): about 4.5 core-hours, under $1 of Modal credit (about $3.7 left after the margin); the judge 303 calls
How this experiment sits among the others is on the lineage page; its family's open hypotheses are on the families page.
Rendered from the record on 2026-11-05 17:14 UTC; updated after every experiment.