Experiment 221
A6.1 (Stage 2, step 4, issue #84). Every sampled close in F1 compared K4's climbed reading with the copies' climbed readings by quadgram score, and the judge bench (reports/judge-bench.md) shows quadgrams find none of the damaged classes against English-like junk, which is what a climb returns on a copy. Here the six verdicts of F1.1 (0120 at 23 and 26, 0140 at 23, 0142 at 26), F1.2 (0144 at 13 and 16, 0146 at 8) and F1.3 (0125) are rejudged: each frozen attack.py is run unchanged with its own seeds (nothing was saved but K4's reading and the copies' scores), its process pool intercepted to keep every job's reading, K4's reading checked against the record, and K4's, the copies' and the in-run plants' best readings read by the judge, one blind call per verdict.
What the method does with the 24 known letters.
Several earlier experiments could not decide a question exactly, so they let a computer adjust the keythe secret word The word or string a cipher needs besides the method. for the best-looking English on the carved lettersthe carved letters The 97 letters cut into the copper at the CIA, shown at the top of every experiment page. and on scrambled copiesa shuffled control The same experiment run on the carved letters shuffled into a random order. of them, and compared the two with a simple letter-statistics score. That score turned out to be poor at spotting damaged English, with missing words, X's between words or a few wrong letters. Here we rerun those same searches exactly as they were and let the judgethe judge A language model that scores how much of a reading is English, from 0 to 100, allowing misspellings and X between sentences., a language model that reads one text at a time without knowing where it came from, score the carved letters' reading and the copies' readings side by side.
Words in the margin are defined on the glossary page.
The carved letters go in at the top; the reading comes out at the bottom.
Earlier verdicts compared K4's climbed reading with the copies' by a letter-statistics score, which is poor at damaged English. The judge reads for words and tolerates X stops and misspellings.
Six frozen searches rerun with their own seeds; every reading came back letter for letter. K4's and the copies' readings judged blind, one call per verdict.
K4 above every copy in 0 of 8 (chance 0.56). The judge found 20 of 20 planted running keys (quadgrams 11) but only 2 of 18 climbed free-alphabet plants: the climbed readings are about a third right, too little to read.
What this experiment was built on, from the lineage page: 13 experiments before it, back to the root of its line. Its number is in bold.
Laid out as the copper carries K4: OBKR ends the line that finishes K3, then three lines of 31. Left, as carved, the 24 known places in accent. Right, the candidate: a known place it gets right is an accent cell, one it misses is outlined. This is the candidate with most known letters right; none passed every gate.
in: the copper, always the same
OBKR ends the line that finishes K3, then three lines of 31
out: the candidate
checked: the known letters in place, and the reading turns back into the copper
The loss of the best candidate, part by part. Zero is solved.
| Known letters missed | 0 of 24, ten points each | 0.00 |
| English | the 73 free letters score -5.81; English text scores -3.6 | 2.21 |
| Key | 205 bits to write it down (key words are cheap, arbitrary letters 4.7 bits each); each 60 beyond 60 costs a point | 2.42 |
| Loss | of 240, the worst possible | 4.63 1.9% |
Redo the method on a few letters with pencil and paper.
The same steps run by the code on the real letters.
A6.1 rejudges a climbed reading. 0140's reading of K4 at key length 23 (one alphabet, Beaufort form):
reading BOARIMELRBCEHINVEHWATEASTNORTHEASTHATEGNANHIFPLLBADIRODMOROSORYBERLINCLOCKDAVARRUNDISHEDSAVEANLAP
judge sees BOARIMELRBCEHINVEHWAT#############HATEGNANHIFPLLBADIRODMOROSORY###########DAVARRUNDISHEDSAVEANLAP
quadgram mean -4.9891, rank 9 of 13 among its own copies
the judge reads the same 13 readings, one at a time, unlabelled; K4 must score above all 12 copies.
Family, key space, the constraint under test, and the clues it reads.
A6 and A13 (knowledge/12-assumptions.md): the sampled closes of F1.1 (one free alphabet, Beaufort form, key lengths 23 and 26: 0120, 0140, 0142), F1.2 (two free alphabets, lengths 8, 13, 16: 0144, 0146) and F1.3 (a running key from an unknown English text on the 208,062 dictionary tables: 0125) rest on the quadgram mean of a climbed reading. C5 and C27(c) (a keyed alphabet 'slightly different' from KRYPTOS, the K1/K2 system) are the clues those closes read; C19 (X stops, Scheidt 2015) and the bench's damaged classes are why the score is in doubt. Statistic per verdict: the judge's score (j3, Sonnet 5.5, kryptos/judge.py) of K4's best reading against the best of each copy's, copies made by the frozen code (0119's exact decider for 0120/0140/0142, kept-place shuffles for 0144/0146, plain shuffles for 0125), K4 counting only strictly above every copy, None as not found. 0125's reading is the forced running-key stretch (24 letters at 22-34 and 64-74), shown to the judge at places 35-47 and 49-59 with '?' elsewhere so its masking of the known places does not hide it. Eight cells; chance puts K4 above all 12 (50) copies with p = 1/13 (1/51) each, about 0.6 cells in all; a cell flags at p < 0.05 / 8, which needs 160 copies: a K4 above every copy here is a candidate for that replication, not a result.
Written before the run; the result decides it.
K4's reading scores above every copy's in none of the 8 cells (chance expects 0.6). K4's six saved readings are reproduced exactly. The judge finds at most as many of 0120's and 0140's planted texts as quadgrams did (0120: 3 of 6 at 23, 1 of 6 at 26; 0140: 5 of 6), because their climbed readings are only a third right (29-43 of 97 letters), and at least the 11 of 0125's 20 planted running keys that quadgrams found. Zero candidates pass gate 4.
The harness's record: candidates, gates, loss, and the negative result this experiment leaves behind.
All 4 candidates, lowest loss first.
saved reading of 0140: Beaufort form, one alphabet RTKDPNVSXJQMUAZGOBILEYCHWF on both sides, key length 23
saved reading of 0142: Beaufort form, one alphabet UTVXRSWEPNHFLZDAOQBJICKGYM on both sides, key length 26
saved reading of 0120: Beaufort form, one alphabet NDPEAFCIXWRMUHGYZKOBTJVSQL on both sides, key length 26
saved reading of 0120: Beaufort form, one alphabet JCRBOQYZPXVWHSMGAUKNFTIELD on both sides, key length 23
0142 L26: K4 6, copies best 15 (rank 13 of 13); 0146 L8: K4 8, copies best 12 (rank 11 of 13); 0125 running key: K4 8, copies best 22 (rank 42 of 51); 0140 L23: K4 12, copies best 18 (rank 7 of 13); 0144 L13: K4 8, copies best 12 (rank 13 of 13); 0144 L16: K4 8, copies best 18 (rank 13 of 13); 0120 L23: K4 12, copies best 12 (rank 6 of 13); 0120 L26: K4 8, copies best 22 (rank 11 of 13). Plants: 0125 plants 20/20; 0140 plants L26 0/6; 0120 plants L23 2/6; 0120 plants L26 0/6
What was searched, as the harness recorded it.
QINARCEOUSEAZEYNONENGEASTNORTHEASTEDTUDECRSABSOSCIVEFUSSHKRLDRABERLINCLOCKCLARINRUNGONTAGTOLEMASCZEMYBMWDBURGHAPTGCWEDEASTNORTHEASTYIDWCKUNDTOFLYPSHAYKINGSADVERBERLINCLOCKAVEFLIVATQUODEMOGOOFGLABOARIMELRBCEHINVEHWATEASTNORTHEASTHATEGNANHIFPLLBADIRODMOROSORYBERLINCLOCKDAVARRUNDISHEDSAVEANLAPDBUGHHPROLEITEDREOFVIEASTNORTHEASTLYOWNENWASINENOHWCITAFFMYSBURBERLINCLOCKOBAYGUIORYFAKELYUSFUMWOKDOWBOOFGOLPREATINCEREASTNORTHEASTYIRTIDONTTOSTENEETSBEIPHYAVAMBERLINCLOCKYWASTILINGLDPREVELOSINPWRYZNNIPNMUPTOMEORJZEEASTNORTHEASTRBYEASILETFIPLSNGESMUCHASTDCEBERLINCLOCKHASANDWLANDWOHLUCEKSENPQBHEFEULCLATLYEPHUVALEASTNORTHEASTPOILTYNESTUCOOTEVELHOURKMRTISBERLINCLOCKTBYINAROMTHEENEIGNORGHFHow this experiment sits among the others is on the lineage page; its family's open hypotheses are on the families page.