“F15.3: held-out drafts planted under two free alphabets at key lengths 13 and 20, each climbed with English alone and with English mixed with the drafts' quadgrams, against 12 kept-place copies climbed the same way”
F15.3, power first (issue #118). 0257's harness (0144's compiled two-alphabet tempering, D1b's four runs, kept-place copies) with two changes: the plants are the 24 held-out drafts of F15.2 (written by another model and prompt, never in the prior), and every text is climbed under two score tables, English quadgrams (today's decider) and English mixed half and half with the quadgrams of the 329 corpus drafts (the draft prior). A third table (a fifth drafts) is reported, not used for the verdict.
We wrote about 350 guesses at what K4the carved letters The 97 letters cut into the copper at the CIA, shown at the top of every experiment page. might say, based on the first three messagesthe earlier messages The three solved parts of the sculpture. and on what Sanborn has said about it. Here we test whether those guesses help our search at all before using them on K4. We hide 24 other guesses, written separately, as secret messages, and see how often the search finds each one when it knows only English, and how often when it also knows the guesses' typical words. If the guesses do not help by at least two hidden messages in twelve, we drop the idea.
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
Guessed K4 texts (drafts) as extra knowledge: a score that rewards their four-letter pieces should pull a search towards Sanborn-like prose.
the check
24 drafts written by another model, never in the score, hidden as plants; found if above all 12 shuffled copies searched with the same score.
the bound
Found: English 18, half drafts 16, a fifth drafts 12 of 24; letters right 36.8, 37.9, 37.2 of 97 on average.
2
Experiment lineage
What this experiment was built on, from the lineage page: 18 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 a guessed K4 text no search has seen, for example one written by a different model.
Encipher it with two scrambled alphabets and a 13-letter key.
Make 12 copies of the ciphertext with the 73 letters at unknown places shuffled.
Search each for the most English-looking reading with the 24 known letters held in place.
Search them again with a score that also rewards the four-letter pieces common in the other guesses.
Under each score, the plant is found if its best beats all 12 copies' bests.
Count found plants under each score; the drafts help if the second count is at least four higher over 24 plants.
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.
F15.3 asks whether guessed texts help the search. For one hidden test message and its 12 shuffled copies:
score E English four-letter pieces only plant best vs copies' bests
score D50 half English, half the guesses' pieces plant best vs copies' bests
found when the plant beats all 12 copies under the same score.
24 test messages written by another model; the guesses help if D50 finds at least 4 more than E.
6
Hypothesis, for a cryptanalyst
Family, key space, the constraint under test, and the clues it reads.
F1.2's two-free-alphabet cells (Quagmire IV) at 13 and 20 are blocked for lack of checks: the 24 known letters fit copies as often as K4, and English alone finds about two thirds of plants (0257: 66 of 96 at four runs). If K4's plaintext is Sanborn's own prose on the subjects he named (reports/f15.1-content-model.md: a paced location, a meeting at the World Clock, the crowds of 1989, two breaches, light and shadow, K3 continued, a pointer onward), a score that knows that vocabulary should pull the climb towards the planted reading and away from English-like junk, and so find more plants against copies climbed with the same score. The prior is a quadgram table: log10 of 0.5 P_english(q) + 0.5 P_drafts(q), P_drafts from every quadgram of the 329 corpus drafts that touches a free place. Held-out drafts are never in it. Clue entries: C2 (instructions in the earlier text), C20(a) (K4 comes back to the first part), C27(c) (two alphabets).
7
Prediction
Written before the run; the result decides it.
English alone finds about 14-17 of 24 (0257's rate); the draft prior finds 2-5 more, because the held-out drafts share the corpus's subject words (CROWDS, PACES, SHADOW, MESSAGE). Copies climbed with the prior also score higher, so the gain is smaller than the gain in letters right. Pre-registered verdict: the bar (4 more of 24, the row's 2 of 12) is met or not; under it, F15.4-F15.6 close as weak.
8
Method
Plants: the first 24 held-out drafts of data/drafts/drafts.jsonl, 12 at key length 13 and 12 at 20 (alternating), each enciphered on two random alphabets and a random key (seeded), with an exact witness (0123's decider through 0144's witness4).
Copies: 12 per plant, the 73 free places shuffled, the 24 known places kept.
Tables: E (English quadgrams, 0144's), D50 (log10 of 0.5 English + 0.5 drafts), D20 (0.8 English + 0.2 drafts).
Every text (24 plants and 288 copies) climbed under each table: four runs of 0144's tempering from the witness.
Found: the plant's best above all 12 of its copies' bests under the same table. Also letters right.
Verdict: found(D50) - found(E) over 24, bar 4.
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
F15.3: plants found against 12 own copies, English 18/24, drafts 50% 16/24, drafts 20% 12/24; gain -2 against the bar 4: not met. Letters right 36.8, 37.9, 37.2.
What was searched, as the harness recorded it.
item
F15.3 (power first)
cells
F1.2 two free alphabets (Quagmire IV), key lengths 13 and 20
decider
0144's compiled tempering unchanged (16 x 625 x 1000), D1b's four runs, from the exact witness
plants
24 held-out drafts (claude-sonnet-5-5, prompt h1; never in the prior), alternating 13 and 20
own copies
12
null per plant
1/13
copies
kept-place: the 24 known places kept, the 73 others shuffled
tables
E
English quadgrams (0144's)
D50
log10(0.5 P_eng + 0.5 P_drafts)
D20
log10(0.8 P_eng + 0.2 P_drafts)
draft prior
25004 quadgrams (6529 distinct) from the 329 corpus drafts, windows touching a free place