Words, repetition and feedback

Historical tokenization notice: this report splits apostrophes into separate tokens. Session 42 joins three apostrophe suffixes for lexical word analysis. Older word-based experiments retain their original segmentation and have not all been rerun. Read Session 42.

Result: no verified new plaintext. Completed the proposed scoring checks, word/phrase map, and a bounded test of word- and rune-dependent rules. The most useful finding in this pass is a large deficit of identical neighboring runes, present throughout the unsolved text under the specified shuffle comparison. This is a property to explain, not an identified cipher. Its novelty in wider Cicada research has not been established.

1. Check the language scoring

Added a word-boundary-aware, smoothed trigram model in word_core.py. It models the start and end of each rune word, with sparse trigram contexts backing off to bigrams and unigram frequencies. The smoothing concentration is fixed at 5 rather than chosen against unsolved text. Training uses 1,963 runes from the same established LP1 plaintext material as the earlier experiments; no speculative decryption is added to training.

Compared this with the earlier bigram-plus-trigram scoring formula, retrained on matching folds. Five LP1 groups were each held out in their entirety: warning, wisdom, the four-page koan, the four-page divinity section, and the final instruction. Each original plaintext was ranked against its 811 nonidentity invertible affine transforms plus 128 rune shuffles. Two further held-out LP2 plaintexts, pages 56 and 57, were compared against all 811 nonidentity affine transforms.

Both scorers ranked the correct plaintext first in all seven cases. Both also ranked the original first in all 29 shorter, word-aligned LP1 windows against the affine alternatives. Short windows accumulated about 64 runes at a time; any final window needed at least 32 runes.

This does not demonstrate that the new scorer is universally stronger: the benchmark did not distinguish the models' success rates. It adds explicit boundary information and verifies useful behavior on these cases. It also gives no evidence that the earlier negative findings arose from a completely broken scorer. Different models' raw score scales are not directly comparable. Neither model is a solution verifier.

2. Map words and phrases

Parsed all 12,956 runes in unsolved LP2 pages 0-55 into 2,899 rune words. Line and page wraps do not themselves split words; word, clause and larger delimiters do. The mapping records page number, section, paragraph, word index, and rune offset. Rune offsets and LP2 page numbers are zero-based. The separate nine-rune segment is included in this descriptive map.

Findings:

• No complete word of four or more runes appears twice.
• No pair of words of five or more runes shares the same pattern of within-word rune differences modulo 29. Such a pattern would survive adding a different constant to every word.
• The known six-rune repetition is exactly the two-word phrase ᛞᛄᚢ ᛒᛖᛁ, at LP2 page 27, rune offset 28, and page 55, rune offset 70. The individual words are three runes each. Their transliteration labels are DJU BEI, not decoded English.
• No other repeated two-to-four-word phrase with at least five total runes was found under the recorded tokenization and within-paragraph/section restriction.

The community quickstart already highlights the six-rune repetition. This session does not claim to have discovered it. Word boundaries make it a useful constrained pattern to retain.

Shuffle comparison

Used 999 shuffles that permute runes only within each original page, preserving each page's exact rune inventory and all original word lengths and delimiter positions. Five exploratory statistics were evaluated, with one-sided upper-tail comparisons and a five-statistic Bonferroni adjustment.

For the repeated phrase statistic, the simulation comparison is 0.001 before, and 0.005 after, the five-statistic correction. The phrase was known beforehand and motivated this analysis, so this is not independent confirmation of a discovery. It indicates rarity under this particular shuffle model, not a probability that a proposed translation is correct. No translation was inferred.

3. Follow the unexpected deficit of repeats

The low within-word equal-rune count prompted a post-hoc examination of adjacent positions. The main deficit is at immediate neighbors; for example, the complete corpus has 63 equal pairs among 10,057 within-word neighboring positions, but 252 equal pairs among 7,244 positions separated by one intervening rune.

For consistency with decryption targets, repeated the adjacency comparison separately on the eight substantial sections, excluding the nine-rune segment. Used 999 fresh shuffles per section, preserving rune inventories within each section's page fragments and keeping layout fixed. These are exploratory follow-ups, not a new untouched confirmation set.

Those eight sections contain 10,050 within-word neighboring positions. The observed equality rate is approximately 0.63%, versus 3.40% in the shuffle comparison. Every section's count was below all of its 999 shuffle counts. The simulation resolution is 0.001; this is not an exact probability estimate for the tail.

Across word boundaries, ignoring delimiters but staying within each section, the corresponding count is 23 of 2,889, versus a summed shuffle mean of 98.47. Thus the deficit is not confined to the inside of words. It does not by itself identify the mechanism, prove a feedback cipher, or justify treating all rune transitions as independent.

Source checks

Extracted rune text from the separate scream314/cicada3301 archive and compared it with the master transcription for LP2 pages 0-2 and 27-32. Both sequences match exactly: 729 + 1,433 = 2,162 runes across nine pages. The respective counts of identical neighboring runes are 4 and 13 in both archives when read as continuous rune streams. These sources can share ancestry; this is a consistency check, not independent authentication. Not every original image was newly transcribed.

The two extracted snapshots needed to reproduce this check are saved in archive_rune_excerpts.json, with their source URL and headings.

4. Test bounded word/rune feedback rules

The adjacency result motivates testing dependence on the previous rune. To also complete the proposed word-level investigation, tested eight feature definitions in

decoded[i] = (a * ciphertext[i] + b * feature[i] + k) mod 29.

All a=1..28, b=0..28, and k=0..28 were evaluated. Features:

1. Previous ciphertext rune, continuous within a section.
2. Previous ciphertext rune, resetting to zero at every word.
3. Previous ciphertext rune, resetting at pages.
4. Previous ciphertext rune, resetting at paragraphs.
5. Current word's zero-based ordinal within the section.
6. Current word's rune length.
7. Sum of the previous ciphertext word's zero-based rune values.
8. Sum of the previous ciphertext word's Gematria Primus prime values.

The previous-word sums start at zero. Counters ignore non-runic material. No selectively unencrypted F rules, plaintext-autokey, nonlinear state changes, or unknown initial-state searches are included.

1,507,072 settings across eight sections. Counts include duplicate transformations, including the repeated b=0 affine baseline. This is not a count of distinct ciphers.

Parameters were selected using the word-aligned prefix ending at or before min(256, floor(60% of section runes)). The same parameters then decoded the remainder, which was never used to choose them. The twenty best prefix candidates are retained per section. Best prefix scores ranged from about -4.81 to -4.64, and their continuations from -5.44 to -5.23 on the boundary-aware model. Inspection of leading beginnings and continuations showed gibberish, not a coherent passage.

Two constructed positive controls used the held-out LP2 page-57 plaintext: one encrypted using previous-rune feedback, the other using the previous ciphertext word's rune sum. Both used a=5, b=7, k=11 in the decoding equation and were encrypted by its inverse. The full search recovered both exact plaintexts at rank 1.

What this changes

The results provide no readable key and no support for the simple linear feedback schedules tested. They do show a specific structural property that arbitrary independent-rune shuffles fail to reproduce: immediate repeated runes are strongly suppressed throughout the substantial unsolved sections.

The next research target is a mechanism that explains that property while preserving occasional repeated neighbors and the known repeated two-word phrase. Candidate mechanisms might involve a changing alphabet or nonlinear state, but none is established by these statistics. A useful next comparison would preserve bigram structure when evaluating repeated longer phrases, and test any proposed mechanism on independent continuations. Wider community research should be checked before claiming novelty.

Reproduction

Python 3 and NumPy are sufficient. Run in this order from the extracted package:

python word_diagnostics.py
python adjacency_probe.py
python feedback_pass.py

word_core.py and lp_core.py supply shared parsing, training and numeric utilities. The scripts use included snapshots and need no network or Tor access. Corresponding JSON files contain mappings, benchmark ranks, null comparisons and failed decryption candidates. No ongoing background task remains after the scripts exit.

Sources:

• Master transcription
• Second archive's rune transcription
• Community quickstart, including the established six-rune repetition

Earlier reports, source snapshots and results remain in the package unchanged, except the checkpoint index and provenance record are updated.