Read a manuscript, preserve competing interpretations, and test a cipher hypothesis. Everything runs in this browser. Export your record before closing; source images stay on your device. Start with the historical investigations or return to Cipher Detective.
Start with a transcription or import a Cipher Detective JSON export.
1. Preserve the source
Keep case, diacritics, spaces and line breaks. A missing symbol is different from a null that deliberately contributes no plaintext. Revisions preserve the original transcription and its annotations.
Retained source revisions
2. Record readings and uncertainty
Select text above, then use “Use selected text,” or enter positions. Positions count Unicode characters from 1 and include both endpoints. New interpretations are appended; earlier readings remain visible.
For example, 0.1 is 10% of the image width or height. Coordinates attach to the recorded image, not a changing screen size.
Derived readings
Keep the symbol sequence, diplomatic reading (original spelling), normalized reading and translation separate. Record which earlier layer you transformed.
3. Apply a known key
Historical nomenclators mix letters, syllables, whole words, names and controls. This tool applies an entered key; it does not recover one. Multiple readings stay unresolved. The explicit teaching convention for repeat and erase acts on one preceding resolved Unicode character; check the manuscript’s actual convention.
Types: letter, syllable, word, name, null, space, repeat, erase, unknown. This example is invented, not Mary Stuart’s historical key.
Link sources, keys and readings
Use IDs such as revision-1, layer-1, key-1, or an external archive identifier. A relationship records a claim and its evidence; matching metadata alone does not establish a historical connection.
4. Compare numerical segmentations
A known key can leave the boundaries ambiguous: 2228 can be 2|2|2|8, 2|22|8 or 22|2|8. Compare readings without changing the original digits. This is bounded known-key path enumeration, not unknown-key cryptanalysis.
At most 32 complete paths are retained. Scores compare candidates; they are not probabilities. The English model is unsuitable for historical French, German or Latin, and assigns zero to readings shorter than three letters.
5. Search a homophonic letter key (experimental)
Several symbol IDs may map to the same letter. A seeded simulated-annealing search tests consistent mappings using a four-letter model built from museum exhibit prose, with a declared letter-frequency penalty. Short texts can still produce persuasive errors; inspect alternatives and use independent evidence. It does not handle nulls, word codes, control signs, uncertain symbols or transposition. Resolve those hypotheses first.
A crib becomes a declared lock. If every symbol is locked, the result is known-key application, not recovery.
Maximum 20 million token evaluations. Five distinct candidates are retained. Use Cancel to stop the background worker. Larger budgets do not guarantee recovery.
No search performed.
What would count as evidence?
The receipt records the exact input, locks, seed, budget, model, candidate keys and score trace. A high score alone does not establish a historical reading. Check the mapping against every occurrence, account for exceptions, and seek evidence that did not select the candidate. Short or damaged texts can yield convincing errors.
The public generated exercise reads THE SECRET LETTER IS IN THE CHEST… It is a practice fixture, not a blind benchmark. Read the evaluation protocol and limitations.
Research behind the workbench
- Historical-cipher transcription conventions (2020): preserve the source and uncertainty.
- Numerical cipher segmentation (2022): known keys can still leave multiple paths.
- Lasry, Biermann and Tomokiyo (2023), Appendix A: annealing helps with letter homophones; nomenclature and archival corroboration still require human work.