Structure Before Language: a pre-registered computational study of Minoan Linear A (Phases 0-22.D) — corpora, runners, pre-registrations and reports.

Complete data and code archive for a pre-registered computational-linguistic and palaeographic study of Linear A, the undeciphered Minoan writing system (c. 1800-1450 BCE). The study characterises the script quantitatively; it does not attempt decipherment. Twenty-two phases were executed under a locked pre-registration protocol in which every threshold and decision rule was fixed before execution and every outcome reported regardless of direction. The archive contains the analytic corpora, every deterministic runner with its seed, every pre-registration document, every metrics file, and the reports that record the verdicts. Principal results. Mycenaean Greek is rejected relative to a same-lexicon control: a Knossos-derived lexicon covers 18.2% of the Pylos corpus but only 9.2% of readable Linear A words. Six further hypotheses (Luwian, Hittite, Sumerian, Akkadian, Hurrian, Ugaritic) could not be tested by the Phase 6 method, whose positive controls were blind; they are neither rejected nor supported. Sign shapes are classified from images at 85 times chance across 129 classes. The SigLA sign-role classification is shown not to be recoverable from either the distributional or the graphic channel (unsupervised ARI = 0.007), and is therefore a functional interpretation anchored in Linear B rather than an independently verifiable property of the corpus. Correction (September 2026). The Phase 6 reports in this archive state that seven language hypotheses were mechanically rejected and that the Mycenaean Greek hypothesis is structurally infeasible. A later pre-registered power audit (Phase 6.P) found that the Phase 6 coverage criterion could not recognise a known language. The structural-infeasibility argument rests on the word-final -o gap, which is suspended as evidence. The statement above replaces both claims. The archived files are unchanged. Details: Papanikolaou (2026), technical report, https://doi.org/10.5281/zenodo.22811049. Withdrawn findings. Four positive findings obtained against correctly specified permutation nulls were subsequently withdrawn when controls external to the null were added: consonant dissimilation and a distinctive lexical share of sequential structure (same-script control in a known language), vowel dissimilation (marginal equalisation by iterative proportional fitting), and the distinctiveness of the sign-sequence constraint (normalised same-script control). Nineteen null-model design lessons are documented, extended to twenty-two in the technical report cited below, of which seven arose from controls applied after a finding had already been reported and five from the Phase 22 cycle described below. Scope and continuation. This archive covers Phases 0 to 22.D. Phases 22 to 22.D, which attempted to break the information ceiling identified here by bringing in external Mycenaean Greek evidence, are also deposited separately, each as its own pre-registration and report with its own DOI and OpenTimestamps anchor; those records are listed under related identifiers and are the timestamped evidence that each analysis plan preceded its data. Phase 22.E is pre-registered but not yet executed, so neither its pre-registration nor its code is included here. All of them returned negative or void outcomes, and one consequence binds on the papers in this archive: the Linear A / Linear B word-final -o gap is suspended as evidence of a non-Mycenaean language pending a range test on previously unmeasured data. Sources. Linear A data derive from SigLA (Salgarella & Castellan), used under CC BY-NC-SA 4.0. The Linear B control corpus derives from DAMOS (Database of Mycenaean at Oslo). See LICENSE_NOTICE.md in the archive.