Verification Methodology

For specialists: how every number was checked and what remains open

Code-first arithmetic

The core implementation (soyga.py) is an independent reimplementation of Reeds’s generation algorithm from his prose description, not from any prior code. The f values are derived symbolically as f = {c: (V[c] - 1) % 23 for c in ABC} — a single line that makes the identity explicit — rather than hard-coded from Reeds’s Table II.

All 46,656 cells of all 36 tables are generated and stored. Every claim about cell counts, match rates, or error coordinates is verified computationally.

Line-number sourcing

Every quotation from the Kupin edition is cited with its line number in kupin_soyga.txt. Key loci:

Seven adversarial passes

The findings were reviewed through seven independent passes before publication:

1. Initial derivation. Arithmetic verified in code against the Kupin primary text.
2. Cross-check against Reeds’s full typeset transcriptions. NISRAM 216/216, Leo 466/468 (2 diffs = Reeds’s own catalogued original-errors), Moon 24/24.
3. Gemini independent review. A second AI system given only Reeds’s paper and the Kupin text was asked to search for the origin of f. It independently located the Section 18 verse and confirmed the identity.
4. Self-correction pass. A deliberate search for overclaims in the initial findings. Five secondary figures were corrected on the record (letter-frequency wording, consonant offsets, hidden-text ratio, o/t/s attestation status, selection-rule count). The corrections are listed in SOYGA_SECOND_REVIEW_2026-06-09.md, not hidden.
5. Second independent review (2026-06-09). A fully fresh reimplementation from scratch — new code, data re-transcribed from raw Internet Archive OCR of Reeds, verse re-parsed directly from the Latin. Every Tier-1 claim confirmed. Five secondary figures corrected. This review strengthened the o/t/s attestation from “inferred” to “attested in the printed edition.”
6. Blinded replication. A reviewer given only Reeds’s empirical f values and the raw Kupin text — with no access to these findings — was asked to search for the source of f. Without prompting, they rediscovered the verse and the identity. This is the strongest single confirmation: the result is findable independently from the primary sources alone.
7. Hostile-room review. The confidence ledger and all open questions were reviewed from the perspective of a hostile specialist audience (Enochian scholars, historians of cryptography). Every vulnerable point was identified and the explicit not-claimed list was written. See Confidence.

Five published self-corrections

These corrections are on the record as a feature of the methodology, not as a sign of weakness:

1. Letter-frequency wording. “Sources = under-represented” was stated absolutely; corrected to “strong tendency — r is a source but slightly over-represented (1.03–1.07×).”
2. Consonant offsets. “+15/+16 two-value rule” corrected to “+15 for 11 consonants, +14 for r and s, k = +3, q = +32. No +16 anywhere.”
3. Hidden-text ratio. “0.80× chance” corrected to “0.99× chance” (dictionary-dependent artifact; conclusion unchanged).
4. o/t/s attestation. Upgraded from “inferred” to “attested in the printed Kupin edition” after the second independent review located the line.
5. Selection-rule count. “22/23 derived” (which required per-table fitting) corrected to “13–19/23 clean depending on variant lane.”

The null model for hidden text

The claim that the Soyga tables contain no hidden plaintext was tested with a scoped, enumerated null model — not a universal negative.

Standard traversals enumerated: rows, columns, main diagonals, broken diagonals, boustrophedon (snake-path), inward spirals — all directions, all 36 tables.

Test: substring matches against a 72,789-word English dictionary (second review figure), versus letter-preserving shuffled controls (same letter frequencies, randomised arrangement).

Result: real tables score 0.99× the shuffled controls — at chance. Entropy of rows, columns, and diagonals sits at the random baseline.

The shuffled-assignment null (code-word construction). For the claim that the code words are drawn from the adjacent Liber Radiorum divine-name triads: a shuffled-assignment null was run in which the same 23 real code words were randomly reassigned to different triads, and the spellability of each code word from its (randomly assigned) triad was measured. Over 20,000 draws: mean 7.3/23 spellable, maximum 15/23. Observed: 23/23 with the real assignment. This is far beyond any reasonable null (p ≪ 0.001).

What this does not claim. The scoped null covers standard traversals. It does not cover every conceivable path through a 36×36 grid, and this work does not claim to have done so. The book’s own text states the tables are operated (as a mirror and binding grid), not decoded; the operative reading and the cryptographic null are complementary.

The 21-test public suite

The public verification bundle (soyga_verification_bundle.zip) includes a 21-test suite covering:

• The 23-letter alphabet definition
• The identity f(W) = V(W) − 1 (mod 23) for all 23 letters
• Machine-verified NISRAM table row prefixes (rows 0, 1, 2)
• Full recurrence integrity across the NISRAM 36×36 table (1,296 cells)
• Book-arithmetic sums for Latin words from Soyga Section 18
• The normalize() function (j→i, v/w→u, length enforcement)
• Independent second-reviewer scripts: verify_soyga.py, verify_section28.py, verify_null.py

Run with: pytest (requires Python 3.10+, pip install pytest).

What remains open — stated explicitly

1. Paleographic check of Section 18. The headline confidence of 0.97 rests on the Latin as edited by Kupin. A paleographer checking the verse against the manuscript (Bodley 908 or Sloane 8) is the single verification that would move this from 0.97 toward 0.99.

2. The Moon row. The printed Section 28 recipe for Tabula 31 yields UISEUR; the unique one-ordinal repair (final xvi → vi) gives UISEUA, which Reeds independently confirms as the table keyword from the full manuscript. The repair has not been verified against a manuscript image. Un-digitised folios: BL Sloane 8, fols. 140r–141r; Bodleian Bodl. 908, fols. 168v–169r. Anyone who can image these is invited to write.

3. The selection rule. The code words are drawn from the letters of the adjacent divine-name triads (23/23 spellable, p ≪ 0.001), but the exact ordinal selection rule is 13–19/23 clean depending on variant lane. Cancer, Pisces, Mercury, and Moon are underivable from any printed reading without emendation.

4. “Soyga alca miketh." The diary’s own gloss (judgment / wisdom) is the safest local reading; translation-versus-comment is not resolved. Laycock and Rowe independently index the words in those terms.

Download soyga_verification_bundle.zip — full verification suite including independent second-reviewer scripts.