Experiment 228
D1 (issue #77), first measurement. Design 2 (F1.5) costs 20-30 min per setting: 10 independent tempering runs (16 replicas x 625 rounds x 1000 moves) on K4, 12 copies and a power test. Two cuts are measured here on the same plants: fewer runs (each run's best is recorded, so power at 1-10 runs comes from one pass) and runs a tenth as long (62 rounds, a second pass). Plants: clean Carter (0139's and 0143's rule) and A6.2's Sanborn-style classes (0226), at F1.1's Beaufort L23 (0140's compiled climb) and F1.2's L13 (0144's).
What the method does with the 24 known letters.
The best tool we have for judging the hardest cases is slow: for each setting it runs ten long searches on K4the carved letters The 97 letters cut into the copper at the CIA, shown at the top of every experiment page. and on a dozen scrambled copiesa shuffled control The same experiment run on the carved letters shuffled into a random order.. Here we hide test messages, run the tool, and note after every one of the ten searches, and with searches a tenth as long, whether the hidden message was already picked out. If it usually is, the tool can be made much cheaper without missing more, and settings we had to skip for cost can be judged.
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.
The decider for the hardest cells runs ten long searches per text. If a hidden message is usually found in the first runs, or in short ones, it can be cut.
259 texts (37 planted messages, clean and Sanborn-style, and 6 scrambled copies of each) climbed twice, every run's best written down.
At length 23 one full run finds as many as ten; at 13 nothing under nine does. With copies climbed one at a time, 3.5 times cheaper, short of ten. The ten-run power itself came out lower than recorded: messy 6 of 12 at 23, clean 2 of 5 at 13.
What this experiment was built on, from the lineage page: 15 experiments before it, back to the root of its line. Its number is in bold.
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.
Redo the method on a few letters with pencil and paper.
The same steps run by the code on the real letters.
D1 measures how fast design 2 finds a planted message. For one plant and its six copies:
run 1 2 3 ... 10
plant best of runs 1..r
copies best of runs 1..r, each
found at r when the plant's best is above every copy's best.
Repeated with runs a tenth as long; power at each budget is the share of plants found.
Family, key space, the constraint under test, and the clues it reads.
F1.5 design 2 (parallel tempering in the set of keys that keep the 24 known letters; 0139-0146) decided F1.1 L23, L26 and F1.2 L8, L13, L16 at measured power; F1.4's free-alphabet residue (about 1,900 settings), F1.7's fitting lags (about 100), F6.1 (about 443,000) and F6.2 3-14 are blocked by its cost. A tempering run either reaches the basin of the planted reading or not; if most plants are found in the first runs, or in short runs, the decider can be cut without losing power. Separately and exactly, copies can be climbed one at a time, stopping at the first that beats the target: the decision 'target above every copy' is unchanged, and under the null the target is exchangeable with its copies, so the expected number climbed is H_12 = 3.1 instead of 12 (no run needed; stated here). Test: per plant, found at budget (rounds, r) when the best of its first r runs is strictly above the best of the first r runs of each of its 6 own copies (null 1/7). Pre-registered: the reference is the full budget (625 rounds, 10 runs) measured in this run; a cheaper budget qualifies when its power is at least the reference's on clean and on Sanborn-style plants at both lengths; D1 is reached when the cheapest qualifying budget times the sequential-copies factor is at least ten times cheaper per setting. Also reported: design 2's power on 12 fresh Sanborn-style plants at L13, beside 0226's 9 of 12 (the reopened F1.2 L13 cell).
Written before the run; the result decides it.
Power saturates early: at full length 3 runs reach the reference power at L23 and 5 at L13; runs a tenth as long lose power (about two thirds of the reference at L23, half at L13). With sequential copies, 3 full runs give about 3.3 x 3.1 = 10 times cheaper per setting at L23: D1 reached at L23, not at L13. Design 2 on fresh Sanborn-style plants at L13: 9-11 of 12. Zero candidates.
The harness's record: candidates, gates, loss, and the negative result this experiment leaves behind.
reference power L23 clean (6, 8), messy (6, 12); L13 clean (2, 5), messy (8, 12); cheapest qualifying budget (625, 9); per-setting factor with sequential copies 3.5
What was searched, as the harness recorded it.
How this experiment sits among the others is on the lineage page; its family's open hypotheses are on the families page.