Could it find a known key?
All 45 constructed positive cases recovered the observed key and every held-out plaintext character. These cases established capability for the restricted search.
Explore / Where the Keys Stop
A key has to
earn its reading.
First recover a known message. Then apply the same search to the manuscript, compare it with shuffled text, and check whether the key stays stable.
45 / 45Known-answer checks recovered
0 / 3Language objectives meeting all manuscript criteria
Restricted bijective search · recorded experiment, 17 August 202601 / One mapping, tested beyond its training text
Follow a proposed key through the checks.
The search mapped 25 single EVA characters to distinct ASCII letters. Each run used four restarts of 10,000 proposals and a connected character four-gram language objective. Five splits kept physical leaves together; three optimizer seeds were used per split and language.
Recorded run · selectors retrieve stored results
Could it beat both controls?
Did the key remain stable?
The output of the selected run
Read what the key actually produced.
Manuscript units
Proposed output · unvalidated reading
Inspect this run’s key and source record
The displayed input reverses the stored bijection on the first 60 held-out output tokens. Selection always starts at token 1; words have not been chosen for apparent readability.
Five splits, two controls
Each dot shows the best real score across three optimizer seeds for that split. The two other marks show the best scores across the corresponding shuffled runs. Higher is better within the same language objective.
The full decision
How the comparison was made
Each null family contains 119 joint realizations. The empirical p value is (exceedances + 1) / 120. The line control shuffles units within lines. The folio control moves intact tokens among equal-length positions on the same folio. Each null realization repeats the fitting procedure.
The decision also required strict superiority in all five splits and median key agreement of at least 0.80 within every split and 0.70 across splits. Holm adjustment included six hypotheses, retaining the three homophonic objectives stopped during calibration. These p values concern the specified controls, not the probability of a language or a cipher.
The homophonic objective that stopped earlier
The earlier many-to-one objective allowed every cipher unit to map to the same plaintext letter. A legal constant mapping outranked the generating key in all 15 matched positive checks. That objective stopped at calibration. The bijective search above was a separate, restricted experiment.
02 / Before testing the manuscript
Could the detector trust its own alarms?
The second experiment measured adjacent-token dependence. Its 2,700 constructed positive trials had recoverable messages; 900 destructive controls tested false alarms. Explore all 54 measured conditions.
Measured grid · 60 trials per positive condition
The selected recoverable messages
False alarms in matched controls
Read the interval and stopping rules
Each bar shows the observed rejection fraction and its recorded two-sided Clopper–Pearson interval, adjusted simultaneously across all 54 cells at family alpha 0.05. The per-interval alpha is 0.05 / 54. The intervals describe the finite control experiment.
The easiest positive condition required its lower bound to reach 0.80. Each negative class required a point rate no greater than 0.05 and an upper bound no greater than 0.10. All nine negative cells failed the upper-bound rule, so manuscript testing stopped. Increasing a positive detection rate cannot repair that failed check.
The homophonic transformation here operates on whole plaintext tokens. The selected depth is a construction parameter for controls, not an estimate for MS 408.
Open the complete 54-condition table
The current research record
Two experiments, two different stopping points.
The restricted key search reached the manuscript and returned no candidate meeting its rules. The dependence detector stopped at its controls. The scope is the stated mappings, samples, objectives and budgets; other cipher systems remain separate questions.
The inherited CR01 and CR02 articles describe earlier exploratory work. The 17 August adjudications and the current book chapter govern this page. Download every displayed result, source path and SHA-256 hash.
Try the mechanisms yourself
What changes when you hide a message?
Measurements, sources and reproducibility
H1 measures the observed distribution of individual units. H(next|current) measures uncertainty in the next unit, given the preceding one. Both use plug-in frequency estimates, in bits. IoC is the fraction of ordered pairs of distinct positions carrying the same unit. Small windows give unstable estimates. Here they describe the displayed sample.
Source excerpts come from the frozen computational inputs to the book’s restricted key-search experiment. Each retains its source classification and normalization record. Download the recorded results and source hashes. The browser experiments use this readable implementation.
For substitution, polyalphabetic, homophonic and transposition definitions, see Menezes, van Oorschot and Vanstone, Handbook of Applied Cryptography, §7.3. The verbose two-letter code is a constructed teaching example.