Linear A: a reproducible statistical framework

Kober-style combinatorics for an undeciphered Aegean script

This project analyses the Linear A corpus (Minoan Crete, ~1800–1450 BCE) using only combinatorics, arithmetic and permutation statistics — no decipherment claims, no translations. Every claim is of the form "candidate", "class", or "consistent with", and every number in the report is reproduced by a seeded script in this repository.

Main document: linear_a_full_report_v2.md (sections §0–§DS; in Russian). The journal-style English preprint (v1.6) condenses it; a popular-science summary is available in English and Russian. See also the FAQ.

Headline results (all permutation-controlled, seed=42)

• LB sound values transfer to LA: signs that are vowels in Linear B behave positionally like vowels in Linear A (p≈0.0005); final-sign alternations share the consonant more often than chance (p≈0.002).
• A toponym + personal-name layer survives into Greek Linear B archives: exact homographs of length ≥3 are 11 observed vs ~2.4 expected (p<0.0001 for ≥4 signs); six candidates are confirmed in name slots of Knossos records (da-i-pi-ta, i-ta-ja, ki-da-ro, pa-ra-ne, ta-na-ti, i-ja-te; p<0.0001). Short-word "identical word" lists are shown to be statistical noise.
• Calibrated morphology: suffix -JA and prefix A- survive family-wise permutation control and replicate across independent archives (plus PI-); -JA marks the non-administrative register (p≈0.01–0.016).
• A "geography slot" in the libation formula: position 2 of the pan-Cretan formula is site-specific (p=0.012); its Palaikastro cluster (DI-KI-TE stem) sits exactly where Greek cult of Diktaian Zeus is later attested.
• Commodity-dependent metrology: grain is counted on a pure binary fraction grid (1/16 quantum, 100% coverage), oil/wine on a mixed duodecimal grid (Fisher p=0.021); people are counted in integers.
• Typology: word-shape profile of Linear A is incompatible with Greek (final -o: 4% vs 45%), Akkadian, Hurrian and Georgian profiles; the remaining candidate cluster (Etruscan/Anatolian/isolates) is not statistically separable at current sample sizes.
• 34 "bimorphs" (two-sign elements living both as free words and inside longer words) vs 19.8±3.6 expected (p<0.001) — a first data-driven inventory of Linear A lexical elements, split cleanly by register (administrative A-DU vs cultic I-DA).

Data

• corpus.pkl — frozen research database: 1721 inscriptions parsed into typed tokens, validated against a hand-checked mini-corpus (44 exact / 3 fraction upgrades / 5 documented revisions / 0 unresolved). Derived from the tabulations of G. Douros and the editions GORILA (Godart & Olivier) and J. Younger, via the open dataset of lineara.xyz. See DATA_SOURCES.md for provenance, credits and takedown policy.
• underdots.tsv / underdots_layer.pkl — a layer of epigraphically uncertain sign readings (Younger's underdots), machine-aligned to the corpus (717 marks, 345 documents).
• lb_lexicon.tsv, lb_name_slots.tsv — Linear B comparison word lists (3694 types) and record-slot personal-name candidates (1641), derived from public scholarly transliterations (Killen & Olivier; Bennett) via archived pages of minoan.deaditerranean.com.
• lb2_lexicon.tsv, lb2_name_slots.tsv — an independent second digitization (4791 types; 1751 slot names) from the linearb.xyz dataset, used for cross-replication.
• sigla_docs.tsv — document-level dating/metadata layer (687 corpus documents get an MM/LM period) extracted from the public data of SigLA (Salgarella & Castellan) by our OCaml Marshal reader (parse_sigla.py).
• damos_lexicon.tsv, damos_name_slots.tsv, damos_context.tsv — word-level indexes derived from the curated DĀMOS corpus (F. Aurora, University of Oslo; 5,832 documents fetched politely via its public per-document JSON), used as the held-out source for the pre-registered confirmatory onomasticon test.
• Raw third-party source files are not redistributed here; run tools/fetch_sources.sh, tools/fetch_lbxyz.sh, tools/fetch_sigla.sh and tools/fetch_damos.py to obtain them from the pinned upstream commits / public endpoints.

Reproduce

Python 3.12 (uv), dependencies in requirements.txt:

uv venv .venv --python 3.12
uv pip install --python .venv/Scripts/python.exe -r requirements.txt
PYTHONIOENCODING=utf-8 PYTHONHASHSEED=0 .venv/Scripts/python.exe validate_pkl.py   # expect 44/3/5/0
bash tools/stress_run.sh    # full battery; expect empty `git status --porcelain`

All scripts are seeded (random.seed(42)); additionally pin PYTHONHASHSEED=0, otherwise rows with tied scores may print in a different order (statistics are unaffected — verified by a clean-clone stress test, report §BE).

Method in one paragraph

Freeze the corpus; never edit it (additions go through corpus_supplement.py). For every observed regularity, build an explicit null model (label permutation, positional sign shuffle, subset resampling), report R and an exploratory p-value, and state honestly whether the effect survives. Negative results are reported with the same care as positive ones; two of our own early claims were retracted by later tests (§AF, §BG).

Citing / contact

Cite as: Zolden & Claude (2026). Linear A: a reproducible statistical framework. v1.4. Zenodo. https://doi.org/10.5281/zenodo.21298882 (concept DOI for all versions: 10.5281/zenodo.21262274) — and please cite the underlying editions (GORILA; Younger; Douros; lineara.xyz; Killen & Olivier). Issues and questions: open a GitHub issue.