AI workflow automation helped us reconstruct Decipher II, Message 1 by connecting hypothesis development, automated search, source research, and reproducible verification. The hardest parts involved recognizing a mechanism our early code could not represent and resolving a two-letter discrepancy in the source text.
Updated November 5, 2026: Our reconstruction of Decipher II, Message 1 has now been independently reproduced and computationally verified.
Independent researcher umbrellaid implemented our published decoding method in separate software, without importing our code. Using the published source paragraph and documented corrections, that implementation reproduced the same 314-character result and passed all 19 checks, including the published hashes. We reran the independent verifier and confirmed those results.
This verifies that the reconstruction works as specified. The complete decoding remains conditional on the documented working correction to 830 at observation 169; the verification does not settle that uncertain historical digit. Read the verification addendum.
Decipher II, Message 1 was solved by contestants in 1986. This project reconstructs and documents a working method.
A 492-letter passage from Gulliver’s Travels, followed by its reversal, creates the 984-position key.
AI tools helped develop hypotheses, write and run software, investigate sources, and challenge interpretations under human direction.
The complete 314-character decoding depends on two documented transcription changes, including one probable reading that remains uncertain.
I recently published An Independent Reconstruction of Decipher II, Message 1 (1985–86), a non-peer-reviewed research report describing the construction, evidence, and limitations. This article explains the puzzle and the investigation in more detail, especially how AI, orchestration, and conventional automation contributed to the result.
The project offered a concrete way to examine a question that matters well beyond cryptography: how do you turn AI-generated ideas into results that someone else can inspect and reproduce?
A Numerical Cipher Printed on a Jigsaw Puzzle
Decipher II was a commercial jigsaw puzzle carrying numerical messages and a prize competition. The photographed puzzle bears a 1985 copyright notice. Message 1 consists of 314 numbers whose interpretation depends on a literary source and a numbering rule, rather than a conventional alphabet-to-alphabet substitution.
The historical reporting gives the story a clear starting point. A March 25, 1986 Statesman Journal account named Eileen Novak and reported three other successful solvers, each receiving $6,250. Novak solved Message 1 on Valentine’s Day. Original solution credit belongs to those contestants.
The newspaper account also supplied two valuable facts: the source book was Jonathan Swift’s Gulliver’s Travels, and the decoded message concerned Bloom County, Berke Breathed’s comic strip. However, the reporting examined for this project did not explain the precise passage, the complete numbering method, or the full numerical correspondence.
That left a substantial reconstruction problem. Knowing a book’s title does not tell you which paragraph supplies the key, how to normalize the text, or how a number larger than the paragraph’s letter count should behave.
Karen Kavett’s community compilation preserved the transcription, photograph, clues, and historical material used in the investigation. Existing community research also mattered: umbrellaid had already publicly reported one important numerical correction. The reconstruction builds on that record and retains those credits.
Why Knowing the Book Was Only the Beginning
A book cipher uses a text as an addressable source of information. The address might identify a page, word, letter, or another unit. Decipher II-1 required reconstructing both the source passage and the rule connecting each recorded number to a letter in that passage.
Several plausible approaches fit the broad description. A program might number consecutive letters, extract word initials, reverse a source, or begin at a different offset. Every approach defines a different search space. A search can be exhaustive within one space while entirely missing the construction the puzzle uses.
The clue 492 pointed toward the eventual source length. Another clue read, “One way gets you half way there.” Those hints become easier to interpret once the construction is known, but they did not specify an executable algorithm by themselves.
The central tasks were therefore:
- Identify a compatible passage in the supplied book.
- Determine how that passage becomes a numbered key.
- Check the numerical transcription against the puzzle image.
- Resolve discrepancies between source representations.
- Test whether one fixed construction accounts for the message.
AI could contribute to all five tasks. Each task still needed an appropriate form of evidence.
Where AI Workflow Automation Helped the Investigation
The research used AI tools to develop hypotheses, implement software, execute experiments, analyze outputs, and investigate sources. I directed the process, set its scope, and reviewed subsequent research direction. The workflow alternated between exploratory reasoning and concrete tests as new evidence exposed weaknesses in the current approach.
The report’s AI disclosure names OpenAI Codex and Claude Code for computational investigations, Gemini for source research, and ChatGPT for checking and contextual interpretation. The tools contributed at different stages; this was an iterative investigation rather than one uninterrupted conversation.
Orchestration means coordinating that work: deciding which question a specialist should investigate, what inputs it receives, what result it should return, and how another part of the investigation should examine that result.
| Part of the work | Contribution to the investigation |
|---|---|
| Human research direction | Established scope, reconsidered assumptions, and redirected the investigation when the existing approach stalled. |
| Cryptanalysis and historical-design work | Proposed and implemented candidate mechanisms informed by the numbers, clues, and puzzle context. |
| Software execution | Enumerated candidate constructions, decoded outputs, and calculated comparison scores. |
| Source research | Examined literary witnesses, historical reporting, and contextual references. |
| Verification and challenge | Reproduced the stated construction and tested alternative passages, implementation details, and claims. |
During the original investigation, separate implementations provided useful checks, but the investigators shared clues and research context. Their agreement supports within-team verification. Subsequent external replication by umbrellaid is documented in the verification addendum.
The distinction matters because coordinated AI work can produce shared errors as well as complementary ideas. Assigning different roles does not remove the need to inspect what each result actually demonstrates.
The Early Searches Exposed a Missing Assumption
Initial one-way and word-initial searches failed to recover the real construction. Those failures described the limits of the represented search procedures. They did not eliminate book ciphers as a family, and they did not establish that the supplied literary clues were misleading.
Known-positive controls made the weakness more visible. One early procedure found an exact planted example but missed a plant containing two edits. A method that struggles with a known altered example needs careful interpretation when real historical transcriptions may contain similar discrepancies.
On October 11, 2026, I redirected the investigation toward missing mechanisms and mistaken assumptions. Cryptanalysis and historical-design agents then proposed and implemented bidirectional numbering. The passage search located the relevant tailor paragraph in Gulliver’s Travels.
The important change was the mechanism the software could test. More computation over the original menu would still have excluded the eventual construction.
The record also preserves a less convenient result: a promising candidate remained below a predefined acceptance threshold in one search. Later refinement did not turn that earlier failed gate into a pass. Keeping those outcomes distinct prevents the final answer from rewriting the history of the investigation.
The Key Uses 492 Letters in Both Directions
The reconstructed key begins with a passage containing 492 normalized letters. Write that sequence forward, append the same sequence in reverse, and number the resulting 984 positions. Each ciphertext number then selects the letter at the corresponding position in the combined sequence.
Normalization is part of the method. The verifier normalizes Unicode, converts the text to uppercase, and retains only ASCII A through Z. Spaces, punctuation, digits, and line breaks do not occupy key positions.
A small invented example shows the construction:
Source letters: A B C Source plus reversal: A B C C B A Numbered positions: 1 2 3 4 5 6
The repeated C at the turnaround is intentional. The reverse half includes the final source letter again. Omitting that duplicate would create a different key and change the addressing.
For the actual 492-letter source, positions 492 and 493 both address its last letter. Positions 1 and 984 both address its first letter. More generally, positions n and 985 − n address the same underlying source position.
This explains how a relatively short passage can support recorded numbers approaching 984. It also gives each plaintext letter multiple possible numerical addresses, since letters repeat within the source and the source appears in both directions.
The reconstruction specifies those available addresses. It does not identify the designer’s procedure for choosing one eligible number over another when composing the original message.
The Passage Describes a Tailor’s Numerical Mistake
The compatible passage appears in Part III, Chapter 2 of Gulliver’s Travels. A tailor measures Gulliver with mathematical instruments and produces badly fitting clothes after making a numerical error. The inspected text yields exactly 492 letters under the reconstruction’s normalization rule.
The paragraph is a memorable choice for a numerical puzzle. Its content makes the construction feel thematically appropriate, although that observation cannot establish the designer’s intent.
The compatible printed witness appears on page 97 of Raymond Macdonald Alden’s Readings in English Prose of the Eighteenth Century, with a 1911 copyright date. That witness provides the wording needed for the reconstruction.
Finding a compatible witness is more specific than finding the book, but it still leaves an important historical boundary. Different editions can contain the same normalized paragraph. The reconstruction therefore does not determine which physical edition the original designer used.
Two Missing Letters Explain the Alignment Problem
The investigation compared a 492-letter source representation with an online transcription whose paragraph contained only 490 letters. The difference was the word my. In a positional cipher, removing two letters changes the addresses of subsequent letters, even when the surrounding passage remains readily recognizable.
The inspected Alden witness reads “take my measure.” The online transcription attributed to an 1801 edition reads “take measure.” A reader understands both phrases, but the key-building program receives different sequences.
The effects were substantial. A 492-letter window beginning at the online paragraph’s start agreed with the compatible decoding at only 55 of 314 observations. Moving the anchor two letters earlier increased agreement to 260 of 314, because that shift aligned a different part of the text.
| Source representation | Observed effect |
|---|---|
| Compatible Alden paragraph | Contains 492 normalized letters and supplies the reconstructed source. |
| Online paragraph without MY | Contains 490 letters; a paragraph-start window agrees at 55 of 314 observations. |
| Window anchored two letters earlier | Agrees at 260 of 314 observations, without repairing the complete construction. |
Computational alignment identified the two-letter shift. Context suggested the missing word, and the inspected printed witness supplied the supporting wording. Detecting a gap and establishing its contents are separate steps.
The original 1801 printed page was not inspected. The report therefore describes a difference between the examined representations; it does not claim to prove that the historical printings differed.
For automation work, this is a concrete data-quality lesson. Two sources can describe the same thing and remain incompatible as exact computational inputs. When every character carries an address, a small transcription difference becomes a structural error.
Two Ciphertext Readings Also Needed Attention
The complete reconstruction uses a corrected working transcription with two numerical changes. One correction has stronger support than the other. The distinction remains part of the result because a coherent output cannot, by itself, settle every ambiguous digit in the source image.
| Observation | Original reading | Working reading | Evidence status |
|---|---|---|---|
| 169 | 330 | 830 | Probable. The competing 330 reading remains documented. |
| 306 | 338 | 388 | Supported by the photograph; umbrellaid’s earlier public correction receives credit. |
The four AI readings supporting 830 were candidate-aware: the readers knew the digit under consideration. They were not blind human examinations. Agreement under those conditions should not be presented as independent confirmation.
With the corrected working input, the key produces a complete 314-character output. The completeness claim remains conditional on the probable 830 reading at observation 169.
Preserving the original transcription alongside the corrections makes that condition inspectable. A reader can distinguish the original observations, the working changes, and the result calculated from them.
Verification Tested the Rule, the Inputs, and the Alternatives
The reconstruction rests on an explicit source representation and a fixed decoding rule, supported by correspondence checks and scoped controls. Language-model scores helped compare candidates, but the final verifier performs deterministic lookups. The proposed plaintext is not an input to that decoding program.
For each observation, the program can record the ciphertext integer, the selected key position, the underlying source position, and the resulting letter. This connects the output to its inputs at the character level.
Coverage also needs a precise description:
| Quantity | Coverage |
|---|---|
| Decoded observations | 314 of 314 under the corrected working input |
| Underlying source positions exercised | 224 of 492 |
| Numbered key slots exercised | 261 of 984 |
The complete rule defines all 984 slots, but the message uses only a subset. Unused mappings follow from the construction; the message does not independently authenticate them against the designer’s historical key.
The research supplement documents several further checks:
- Alternative windows. Under the two-way, 492-letter rule, the true source window ranked first among 454,033 windows in the supplied-book comparison. That control used a representation with MY restored, so it belongs to the post-answer evidence.
- Uncorrected input. The true window also ranked first under the original numerical transcription in that comparison. The two working corrections did not create the passage’s entire ranking advantage.
- Wrong-passage repair. Forty selected alternatives received a generous automated repair budget and remained below the true passage under the reported settings. This tested those alternatives and that repair strategy, rather than proving universal uniqueness.
- Separate implementations. Within-team implementations reproduced the stated construction. Later known-answer CPU and GPU checks agreed on the relevant hashes.
The candidate scorer used a character-based statistical language model. That scorer estimated how language-like a candidate output appeared. Its score was a ranking aid, not a verdict on historical correctness.
GPU work expanded the computational checks after the mechanism had been identified. One known-source search evaluated 16,345,616 hypotheses and ranked the true candidate first. Because the book and relevant answer information were already supplied, that result must be described as post-answer validation rather than blind discovery.
Other tests exposed weaknesses. Five one-edit perturbations and five two-edit perturbations produced zero exact recoveries in each group. Some outputs retained high partial agreement, but partial agreement and exact recovery are different outcomes. The investigation therefore does not establish a general, reliably calibrated cipher detector.
The Decoded Letters Still Need Careful Interpretation
The fixed construction yields 314 literal characters under the working transcription, consistent with the historical description of a tribute to Bloom County. The verifier does not insert word spacing, punctuation, line divisions, or spelling repairs. Those are later interpretive choices rather than additional cryptographic observations.
The report reproduces a short opening that reads, with editorial spacing, “In homage to Bloom County.” The complete poem is excluded from the public package and from this article.
Two details illustrate why the literal output matters. The output contains PISTACHIO, but not YAZ. ChatGPT assisted with the later contextual association to Yaz Pistachio, and a 1983 Bloom County strip supports the character reference. That is contextual interpretation, not an extra name recovered from the numbers.
Similarly, the literal output contains BRINKLEY. The reconstruction retains those letters rather than silently changing them to the likely intended Binkley.
Keeping calculation and interpretation distinct allows readers to inspect the result without inheriting every editorial decision made afterward. It also prevents a model’s plausible explanation from becoming an undocumented alteration to the evidence.
The Finished Decoder Runs Without AI
The public decoding application needs Python 3.10 or later and only its standard library. Once the source, construction, and corrections are specified, reproduction requires no model, API key, network service, GPU, or additional package. The exploratory research and the finished verifier have different computational requirements.
The public reconstruction repository provides the literary-source transcription, correction specification, verifier, synthetic tests, aggregate results, hashes, report, and supplement. Readers obtain the original numerical input separately through the source referenced in the project instructions.
The reproduction sequence is explicit:
- Authenticate the original numerical input against its recorded fingerprint.
- Apply the documented working corrections and their preconditions.
- Normalize the supplied literary source and check its 492-letter length.
- Construct the forward-and-reverse key and decode the message.
- Write fresh outputs and compare their hashes and correspondence checks.
A matching hash shows that two representations contain identical bytes. It does not resolve the uncertain photographed digit, identify the designer’s physical edition, or establish historical publication priority.
The fixed decoding is more reproducible than the complete exploratory history. Some exact historical software versions and execution records are unavailable, and the public release does not support replaying every whole-book or GPU search. Readers can verify the specified construction while retaining those limits on the research record.
What This Case Teaches About AI, Orchestration, and Automation
The case shows a workable division of research responsibilities: AI contributed hypotheses, code, analysis, and source investigation; automated programs performed defined computations; human direction changed scope and reviewed the path forward. The value came from connecting those activities to evidence that could be examined outside the original conversation.
Five lessons carry into practical operational work:
- Represent the right problem before expanding computation. The initial searches could not express the eventual construction. In a business workflow, the equivalent failure is automating the wrong decision or excluding the exception that explains why the process keeps breaking.
- Give different tasks different standards of evidence. A language score ranks candidates. A printed witness supports wording. A deterministic verifier checks a construction. An uncertain image reading remains uncertain until stronger evidence resolves it.
- Make handoffs concrete. A useful research handoff identifies the proposed rule, supplied inputs, observed output, and remaining question. The same discipline helps teams move from AI suggestions to reviewable software actions.
- Retain failed gates and corrections. A final success does not make every earlier procedure reliable. Failed controls reveal where a workflow needs better inputs, narrower claims, or a different method.
- Turn a successful method into ordinary software where possible. Exploration may require flexible reasoning. A stable calculation benefits from a small, inspectable implementation that gives the same result each time.
This case does not measure whether multiple agents outperformed one agent, reduced total cost, or were necessary for success. There was no matched comparison supporting those claims. What the record does show is how coordinated AI-assisted work contributed to a specific reconstruction and how subsequent checks narrowed the account.
The same questions guide human review in AI workflows: what has the system established, what still needs judgment, and what evidence should accompany the next action?
For Decipher II-1, the tangible result is a compatible literary source, an explicit numbering rule, and a reproducible decoding under documented working corrections. Original solution credit remains with the 1986 contestants. The designer’s exact edition and number-selection procedure remain unknown, and the examined sources do not settle whether a complete method appeared elsewhere previously.
You can read the research report on Zenodo or inspect the software to follow the construction yourself. If your team is exploring a similarly complicated workflow, Eckman Design can help map the research, automation, and review steps into a practical system.
AI disclosure: The research report describes AI contributions to hypothesis development, software engineering, experimental execution and analysis, source research, and manuscript drafting and editing. Christopher A. Eckman directed the research. This article was also prepared with AI assistance.
Discussion
0 Comments