Dorabella Experiments

is it three rows, or one string? The transcription records 29 / 31 / 27. But the two symbol pairs that straddle the breaks — 2nw→1n and 3nw→2se — both recur elsewhere in the text, so the message runs across the breaks using ordinary pairs, as continuous prose does. And 29/31/27 are unequal, which is what hand-wrapped lines look like; a deliberate division would more likely be regular.

So one continuous string is the default here. Row-based operations are still offered, but they are marked, because they assume a structure the evidence does not support.

stack them Each step feeds the next. Everything except the two LOSSY drops is a pure rearrangement, so symbol counts — and therefore index of coincidence, entropy and the frequency profile — cannot change. Only n-gram order moves, which is the thing the solver actually judges.

why this belongs on this page A string operation and a rotating key are not independent. With a static key, reversing the string is absorbed by the solver — it differs only by a global rotation. With a key that moves, reversal changes the unwound stream for every schedule, because the key state at each position is no longer the same one.

So the chain above is applied first, and the key motion is unwound after it — in the new order. Testing a reordering without this only ever tested the one case where reordering barely matters.

the only honest scoreboard Anneal a substitution over the rearranged stream and compare with real text solved at the same effort, against the matching baseline — English controls for English, German controls for German, compressed-English controls for shorthand. Comparing across targets without that would manufacture a discovery: raw scores from different language models are not on one scale.

Watch the percentile, not the prose. Human eyes find words in anything at this length.

German is scored by a 3+4-gram model built from 2.5 M letters of 19th-century prose, with its controls drawn from a held-out 15% the model never saw. That last part is not fussiness: in-corpus German controls measured 0.43 SD too high, which would have raised the reference band and made any candidate look worse than it is.

every chain you have tried Kept so you can see the whole picture rather than the single best row. If you try twenty chains, the best of twenty will beat the rest by chance alone — the count is shown for exactly that reason. A result is only interesting if the rule behind it had a reason before you ran it.