Predict what will happen, try it, then explain the result. The messages, keys and case notes in the six numbered activities are invented teaching examples. The linked papers supply the research ideas; these activities do not reproduce their datasets or trained models. A separate historical handbell case follows the exercises. Everything runs in your browser, with no upload or account.
For your own documents, use the Evidence Workbench. For sourced historical cases, visit Decipherment Investigations.
1. Which key belongs to this message?
A loose sheet and three key cards have been filed together. The folder label suggests one connection—but a folder can contain material from several dates. Which card deserves a closer look?
Exact recorded symbols: 1 2 2 3 / 4 3 / 5 4 3 2. In this exercise, slashes are known word boundaries and each number represents one letter.
- Key A: found in a folder labeled “1625.” The loose sheet has no date.
- Key B: an incomplete card, with no surviving date.
- Key C: a later copy of a key, filed elsewhere. Its relationship to the loose sheet is initially unknown.
Entries:
Literal reading:
Additional evidence
The case notes supply a separate plaintext copy: MEET AT GATE. Treat its identification as a matching copy as a given for this invented case. In archival work, the copy’s identity and independence would themselves need verification. Spacing is retained; no spelling has been silently repaired.
Explain: Why does mapping every symbol fail to distinguish A from C? What does the separate copy add that the folder label cannot?
Research credit: Eva Pettersson and Beáta Megyesi, Matching Keys and Encrypted Manuscripts (2019), §§1 and 3–5. Their work connects transcribed keys with encrypted manuscripts and evaluates readings using historical language models. This three-card exercise uses literal decoding and supplied corroboration, not their scoring system.
2. Not everything in a cipher letter is encrypted
A letter may mix readable language, codes, dates and damage. Classify these six passages before revealing the case notes. Cleartext was left unencrypted; decrypted plaintext is recovered by applying a key. A readable-looking word alone does not settle which you have.
The chunks are supplied for you. “Uncertain” is a valid initial judgment when evidence is missing. Brackets mark a modern transcriber’s note.
Explain: Which digits belong to ordinary text? Why should a French greeting not decide the language of the whole letter? What should stay unresolved?
Research credit: Maria-Elena Gambardella, Beata Megyesi and Eva Pettersson, Identifying Cleartext in Historical Ciphers (2022), introduction and conclusion. Their study distinguishes cleartext detection from language identification and discusses short passages, mixed languages and numeral-filter errors. These invented chunks are a manual lesson, not their classifier or evaluation set.
3. How much can one misread letter change?
Start with the clean message, then misread or lose one ciphertext letter. Does the search still find the key? Does finding the key restore the missing information?
This bounded experiment searches all 26 Caesar shifts. It ranks each reading by exact matches to a five-word vocabulary shown below. Spaces are supplied, not recovered. Ties remain ties. This is a deliberately favorable classroom example, not a general historical-cipher solver.
Original:
Clean ciphertext:
Selected transcription:
Complete scoring vocabulary: . These word lists were chosen for this exercise, not learned from a historical corpus. The older-spelling sentence is invented and is not assigned to a particular date or dialect.
How the experiment is measured
The generator encrypts with shift 7. The search receives only the ciphertext and selected vocabulary; it tries all shifts without being given the answer key. The original is used afterward to count edits. A correct key can coexist with an incorrect recovered message. Changing vocabulary can change a score without changing any letters in the source. No automatic spelling correction is applied.
Compare the older sentence under both vocabularies while leaving noise unchanged. Then change noise while keeping the sentence and vocabulary fixed. These comparisons separate model mismatch from a transcription error.
Research credit: Micaella Bruton, Meriem Beloucif and Beáta Megyesi, Bridging the Low Resource Gap in Historical Cryptology: A Multilingual Diachronic Synthetic Dataset for Reproducible Cryptanalysis (2026), introduction and dataset design; publisher PDF. HistCiph motivates testing transcription noise and language change. This Caesar experiment uses neither HistCiph records nor its homophonic generator and does not reproduce the paper’s results.
4. Different symbols, same hidden pattern
Rename every occurrence of a symbol consistently. What changes—and what stays? Give the first distinct symbol the number 1, the next new symbol 2, and so on. Whenever a symbol returns, reuse its number.
Original
Pattern:
Compared sequence
Pattern:
What does this reveal?
It records equal-symbol positions, not plaintext. Renaming is one-to-one: different symbols keep different names. In a homophonic cipher, different ciphertext symbols can still stand for the same plaintext letter, so this pattern is not necessarily the plaintext’s pattern. This does not apply as a decipherment method to modern encryption or a properly used one-time pad.
The single-occurrence action replaces the first repeated occurrence with a new symbol. If all symbols are distinct, it replaces the last with the first. With only one symbol, no equality relationship can change. Separating spaces here are token boundaries, not symbols being encoded.
Research credit: Nishant Kambhatla, Logan Born and Anoop Sarkar, Decipherment as Regression: Solving Historical Substitution Ciphers by Learning Symbol Recurrence Relations (2023), §3.1. This activity implements their first-occurrence numbering definition on an original example. It does not run their Transformer model or claim their decipherment performance.
5. What survives when words are replaced?
A repeated label can reveal a useful relationship without revealing an entire message. In this original task, every source has three words drawn from PINE, OAK, ELM and ASH. The question is always: were the first and last original words the same?
Read the four labeled examples, then try a new one under the same replacement rule. Here a token is a whole word; the spaces and three positions stay visible.
| Example | What you receive | Original ends |
|---|
New transformed message:
Its visible recurrence pattern: . These numbers describe the transformed words, not an inferred plaintext.
Inspect the source, rule and all compatible readings
Supplied original: . It is shown for checking the exercise, not supplied with the question.
Complete toy rule:
The check enumerates all 4 × 4 × 4 = 64 possible source triples. For a fixed substitution, the count below assumes the whole mapping is known; the labeled examples alone do not establish that a learner has discovered it. For fresh replacement, every output is possible from every source under the stated independent, uniform model.
Try the counterexample: choose the many-to-one rule and compare A with B. Then choose C: can you answer the equality question even though several original messages remain possible?
Scope: this is a deterministic teaching display, with frozen possible draws for the fresh-replacement condition. It runs no language model and reports no AI accuracy. It distinguishes a task clue from full message or key recovery. Whole-word replacement, a historical letter cipher, and modern encryption are different systems; random replacement here is not an encryption scheme with a decoding key.
Research credit: Zhouxiang Fang, Aayush Mishra, Muhan Gao, Anqi Liu and Daniel Khashabi, ICL CIPHERS: Quantifying “Learning” in In-Context Learning via Substitution Ciphers (2025), §§2.2–2.5, pp. 25913–25914. Their comparison uses stable token permutations and fresh per-occurrence replacements; task success need not recover the whole mapping. This four-way comparison, word task, fixtures and exact enumeration are our teaching model, not a replication of their experiment or results.
6. A hidden message still has to reach its reader
Would the recipient recover the intended message? Would the carrier attract attention? Would a discovered message remain secret? These are separate questions. The three original examples below expose the extraction rules and keys so that you can test them.
In her study of seventeenth-century German handbooks, Eveline Szarka discusses Harsdörffer’s ambiguous invoice numbers, musical ciphers whose rhythm or melody might look unusual, and Friderici’s need to keep reserved words out of surrounding text. Those are accounts of proposed methods, not evidence that every proposal was used successfully.
6a. An invoice with two legal readings
Our invented word code is 1 → MEET; 2 → DAWN; 3 → DUSK; 12 → WAIT; 13 → CANCEL. Concatenate codes to form an amount. The recipient knows this code but must also know where its boundaries are. This is not Harsdörffer’s historical letter key.
| Quantity | Unit price | Amount field received |
|---|---|---|
|
Every legal decoding under the selected boundary rule:
Check the sender’s record
The invented sender intended . This record lets you check the exercise; a real recipient cannot resolve ambiguity by consulting an unavailable original.
6b. An ordinary word can become an accidental marker
This separate toy rule extracts every whole word amber → MEET, silver → AT, copper → DAWN, ignoring case and punctuation. The surrounding words are ignored. Predict what happens when an ordinary sentence uses a reserved word.
Words extracted:
Recipient’s reading:
6c. Decodable rhythm, conspicuous carrier
Our tiny duration code is WAIT → 1, 1, 1, 1 beats and MEET → 1, 1, 1, 2 beats. Each sequence is presented as one complete, single-voice bar labeled 4/4: it must contain four quarter-note beats. Only duration is modeled; this is not a historical score or a melody.
Decoded with the shown duration key:
Explain: Which change helps the recipient but changes the cover? Which ordinary word damages the message? Which code still decodes despite a visible inconsistency? None of these checks measures an interceptor’s detection rate or establishes a secure covert channel.
Continue with the Cardano grille’s selected positions, null-cipher extraction rules, or Bacon’s two-symbol representation.
Research credit: Eveline Szarka, On the Combination of Cryptography and Steganography in 17th Century Germany (2023), §§4–6, pp. 163–166, and conclusion p. 167; publisher PDF. All invoices, word codes, sentences and rhythms in this activity are new teaching fixtures. No source illustration, historical key or musical composition is reproduced.
The handbell: when a date rules out a cipher
Historical case — unresolved. Rudolf II’s alchemical handbell, made around 1600, carries a Greek-letter inscription. Bean, Gannon and Lang’s 2023 study corrects an earlier transcription and compares possible interpretations.
The same inscription, represented as digits or Latin letters, prompted different suggestions from Mason’s cipher classifier:
| Input representation | Leading suggestions |
|---|---|
| Digits | Monome-Dinome; Tridigital |
| Latin letters | Bazeries; Checkerboard |
These tools used ACA-oriented training and English-language assumptions. Their rankings are clues, not calibrated probabilities of the bell’s cipher type.
Could Monome-Dinome explain an inscription from around 1600?
The paper dates Monome-Dinome in its present form to about 1936. That chronology rules out this particular suggestion for the bell. Eliminating a candidate does not identify the correct one.
Keep the question open: the paper does not settle whether the inscription encodes recoverable plaintext or represents another kind of meaningful symbolism. No decipherment is supplied here.
Source: Richard Bean, Corinna Gannon and Sarah Lang, The cipher of Emperor Rudolf II’s “Alchemical Hand Bell” (HistoCrypt 2023), pp. 13–17, §§2–3 and 5. The classifier reports and chronology come from that study; the museum has not inspected the artifact.
Sarah Lang’s Situating ciphers among alchemical techniques of secrecy (2023) adds a useful distinction: secrecy can involve cover names, symbolic conventions and social practices as well as reversible encryption. Transcription alone need not reveal meaning.
Continue with recording uncertain readings, applying your own key, or the experimental homophonic search.