Find your way through the investigation
Seventy-three sessions ask a single question from different directions: what can the unsolved Liber Primus actually support? This guide connects the strongest results, the corrections that matter, and the questions still worth testing.
Research snapshot: Session 73. No new verified plaintext, real decoding table, or author identity has been recovered. This guide organizes the existing research; it is not a new cryptanalysis session.
Start with the kind of question you have
| Your question | Read first | Then inspect |
|---|---|---|
| What does a successful method look like? | Known-solution controls · 1 | Boundary and counter controls · 62 |
| Can I trust the word boundaries? | Apostrophe correction · 42 | Full singleton image audit · 43 |
| Why are repeated neighboring runes unusual? | Inventory-conditioned comparison · 29 | Common-letter counterexample · 70 |
| What does the latest test say? | Opening-section source comparison · 73 | Singleton structural roles · 72, original singleton audit · 43 |
| What do the tree and colored marks tell us? | Tree passage structure · 55 | Derived boundary map · 62 |
| Did the Berkeley remailer lead go anywhere? | Expanded source comparison · 65 | Verification records and manifest |
| Are the literary parallels a key? | Literary source map · 15 | Supplied book-page comparison · 24 |
The current position, in plain language
1. One fixed table is severely constrained
A homophonic cipher can let several ciphertext symbols mean the same plaintext letter. But a single shared decoding table runs into the corrected one-rune-word inventory. If those groups are ordinary A/I words, the fixed-table model forces 96.58% of the text into those two letters, leaving at most three plaintext symbols. That is a statement about this model and its word premise, not every possible cipher. Session 68
2. Changing tables creates capacity, not a recovered key
Compact selectors were tested in Session 69. Under its reference screens, a shared bank needs at least 20 tables for A/I, or ten when O is allowed. Independent section banks remain more flexible. Passing a capacity screen says a model has room for certain outputs; it does not demonstrate that the unknown tables contain English. Session 69
3. Repeat avoidance adds a measurable prediction
Session 70 adds a specified group-based repeat channel to section-specific banks. The smallest banks passing its full-section screens need five or six tables. Four later singleton words contradict selectors frozen on earlier text. Three frozen selectors retain full-section capacity, and only two also pass the continuation-only frequency screen. No real codebook is recovered. The frequency caps come from a limited solved reference sample; they are not universal limits for English. Session 70
4. A common letter can still avoid exact ciphertext doubles
The constructed controls reversibly encode solved prose with 248 scored E positions and zero neighboring ciphertext repeats. Choosing among homophones makes this possible. So a shortage of repeated ciphertext symbols alone does not identify a rare plaintext letter. These controls verify a possible mechanism with known tables, not unknown-key recovery from the real book. Control records
5. Alphabet coverage adds another conditional limit
Session 71 asks whether each state can encode the same declared plaintext alphabet. Under that additional premise, the three frozen full-section survivors from Session 70 fail the saved common-letter frequency profiles. No tested bank with up to eight tables and 25 or more declared types passes. Smaller and restricted state alphabets remain open; fifteen known-text controls recover exactly. These are conditional reference screens, not universal English limits or a recovered key. Session 71
6. Location does not establish a marker role
Session 72 compares the audited singleton positions with seven mapped structural features. The primary placement test gives p = 0.3027 after taking the feature maximum; short-word and red-text sensitivity checks are also unremarkable. Of 81 singletons outside red text, 61 have none of the listed features and still span 26 symbols. A strong planted location signal is detected. This gives no positive support for a general marker explanation and does not prove that the words are English. Session 72
7. A guided source comparison has a limited evidential role
Session 73 retains a frozen first pass and a reviewed reading of LP2 pages 0–2. The rough first pass differs by 136 minimum edits, largely label confusions. The master-guided review matches the existing 729-rune stream and retains four neighboring repeats against 24.812 inventory-matched expected, five singleton types and two words across page seams. No source correction is established. This is a reproducible comparison fixture, not independent confirmation or new plaintext. Session 73
Evidence ledger
| Claim | Standing | What would change it? |
|---|---|---|
| The known prime-based LP2 method can be reproduced | Reproduced control; Session 1 | A conflicting source or failed reproduction on the same input |
| Corrected one-rune groups span 28 symbols | Image-audited inventory: 83 tokens; 42, 43 | A documented glyph or boundary correction |
| One-rune groups generally act as structural markers | No clear concentration at seven tested features; 72 | An independently motivated role with source-confirmed predictions |
| Opening-section source comparison changes the master | No correction established; same-reader guided review, 73 | A source-confirmed disagreement from an independently validated reader |
| A repeated six-rune phrase occurs at two locations | Image-audited observation; 32 | A source-image correction |
| Simple recency models explain the unsolved text | Tested versions fail reference screens; 66, 67 | A defined untested mechanism that also passes held-out checks |
| Berkeley wording links the earlier remailer to Cicada 3301 | Not established in the sampled material; 65 | A verifiable cross-reference, distinctive dated passage, or specific reused implementation |
| Literary and religious parallels supply an unsolved key | Not established; 15, 24, 28 | An independently motivated rule producing reproducible continuation |
| Section-specific homophonic tables solve the book | Open model family, no recovered codebook; 70, with full state alphabets further constrained in 71 | A frozen table and rule that recover coherent held-out text and explain cross-section recurrence |
Corrections worth reading before reusing old results
- Words: Session 42 corrected three apostrophe suffixes previously split as words. Session 43 completed the remaining source-image checks. Use this corrected lexical inventory for singleton claims.
- Punctuation: Sessions 58–61 audit image/transcription discrepancies. Session 62 creates a derived boundary map while preserving the master transcription. Session 63 traces which earlier clause-reset results change.
- Steganography: Session 37's OutGuess crashes are not evidence of hidden payloads. Session 38 diagnoses the iterator issue, preserves known outputs, and recovers no new payload.
- Historical sources: Session 65 expands the Berkeley comparison to 28 distinct verified message bodies from 2012–2014 and separately checks 94 solved-text clauses. No shared four-word phrase was found with the sampled Berkeley material. The sample cannot rule out all historical connections.
Word correction · Boundary impact audit · OutGuess retest · Berkeley expansion
Research roadmap
These are proposed next tests, not completed findings or promised breakthroughs.
- Improve recognition before more source readings. Session 73 completes the opening-section comparison but exposes substantial first-pass mistakes and a guided-review limitation. A glyph-matching reader calibrated on solved images should pass other solved-page controls before the unknown section is compared. Do not expand the same manual procedure and claim independent confirmation. Smaller or restricted cipher alphabets still need independent evidence before more tables are fitted.
- Preserve the cross-section phrase. A section-specific mechanism must account for the repeated phrase rather than treating every section as unrelated.
- Freeze the mechanism before continuation. Specify state, boundaries, exceptions and score in advance; use later text for validation and retain contradictions.
- Keep positive controls. Recover constructed messages and known solved material with the exact implementation and input conventions being tested.
- Reopen parked leads only with new evidence. More arbitrary selector settings, extra passage splits, or a shared insect name do not supply independent support.
These priorities follow the source-comparison limits, the structural-role test, the limitations in Session 71 and Session 70, the recency-family audit, and the Berkeley conclusion.
Reproduce a claim
Start from the report's Reproduce and inspect section. Keep its source version, dependencies, hash checks, and boundary policy together. Compare your output with the saved results before changing the model.
For the current source-comparison fixture: report, fixed protocol, frozen first pass, Python implementation, complete results, master transcription. The report requires Python, NumPy and Pillow; no fresh experiment was run for this guide.
Use the searchable journal for reports and the searchable evidence files for code, inputs and outputs. A negative result is useful only with its assumptions attached.