Auditing the Unread — A Corpus Validation Software Framework for Undeciphered Scripts, with a Demonstration on Linear A (including open-source code)

Abstract

🔗 This paper is accompanied by open-source software and a fully reproducible pipeline. Code and data-generation scripts are available at:https://github.com/ChristosTsirkas/corpus-validation-for-undeciphered-scripts-linear-aLicense: https://github.com/ChristosTsirkas/corpus-validation-for-undeciphered-scripts-linear-a/blob/main/LICENSE.mdFor decades, research on Linear A has relied on the assumption that digital corpora provide a reliable foundation for applying increasingly complex methods—including large language models. Attention has focused on decipherment, yet a fundamental question remained unanswered: is the corpus itself sufficient? This paper approaches the problem from the opposite direction. Before examining hypotheses about the language, it tests the reliability and informational adequacy of the corpus itself. Three main independent sources were compared: GORILA, Younger's transcription, and the "SigLA" database, plus secondary ones. The comparative analysis revealed systematic deviations in the rendering of two signs, and in every case of disagreement, Younger and "SigLA" agreed completely. This demonstrates that even basic digital sources require systematic auditing before computational analysis. The most significant finding concerns the amount of information the corpus contains. "Minimal pairs" occur less frequently than chance. "Conditional entropy" cannot be calculated reliably, as the corpus falls short by two orders of magnitude. "Vowel harmony" tests detected no statistically significant signal. The conclusion is not that methods are flawed, but that their application presupposes far more information than Linear A currently possesses. Techniques such as "morphological segmentation", HMMs, "Bayesian cognate detection", "sound-correspondence search", LLM-based approaches, and proto-language reconstruction cannot be supported by the present corpus. At the same time, the paper re-examines Packard's (1974) statistical hypothesis that Linear B phonetic values can be applied to Linear A. Using thousands of permutations instead of his original nine, the analysis confirms the stronger strand of his argument with far stricter controls. The framework is script-agnostic and can be applied to any undeciphered script: Cypro-Minoan, Indus, Proto-Elamite, or Rongorongo. No method, however sophisticated, can recover information that is not in the data. In the case of Linear A, the paper concluded that the available corpus does not contain enough information to substantiate any computational decipherment. The paper develops a reproducible audit and adequacy-testing framework for digital corpora of undeciphered scripts, demonstrated on Linear A. The framework comprises four core components: sign-identity resolution (treating Unicode sign identifiers as authoritative data rather than lossy ASCII labels), structured preservation of editorial uncertainty (damage, reading certainty, and measure confidence as structured fields), multi-witness comparison against independent editions with documented conventions folded before counting divergences, and an adequacy protocol that determines whether a statistic is computable by counting cells against observations before computing it. The framework is script-agnostic and transferable to any undeciphered script corpus with multiple editions. The demonstration on Linear A produces several findings. A widely used digital corpus systematically inverts AB21 (sheep) and AB22 (goats) across fourteen documents; a third witness agrees with the alternative reading at fifteen of fifteen attestable sites, in both directions, establishing a mis-mapping rather than scholarly disagreement. Three adequacy tests show the standard routes to a structural fingerprint are closed: minimal pairs fall below chance (z = −1.63), conditional entropy is inestimable by two orders of magnitude, and vowel-sequencing dependency is undetectable under both five- and three-vowel specifications, with harmony of strength ≥ 0.10 excluded by power calibration. A replication of Packard's (1974) statistical defense of transferring Linear B phonetic values to Linear A — the first in fifty-two years — reproduces his weak evidence (1.55 : 1, p = 0.17) and independently confirms his strong evidence (4.74 : 1, z = +4.34, p = 0.0020) on a different corpus with four thousand times his permutation count. Affixation exceeds chance in both registers (z = +5.91 administrative, +6.74 religious), while the prefix/suffix ratio does not (z = +0.23), meaning the direction of affixation carries no information at this corpus size. The paper contributes a methodological framework, a corrected corpus, and a set of stated bounds on what a corpus of this size can support. Four times during the work a defect in the comparison tooling produced what looked like a substantive finding, once at conventional significance — demonstrating that in this domain a divergence count measures the comparator at least as much as the corpus. The framework is fully reproducible: all code, data-generation scripts, tests, and documentation are available in the accompanying repository, with checksums, a verification script, and byte-identical rebuild from source.This paper aims to contribute to the literature on undeciphered scripts by developing a reproducible audit and adequacy-testing framework that determines, in advance, whether a corpus can support a given statistical claim. The framework comprises sign-identity resolution, structured uncertainty preservation, multi-witness comparison, adequacy testing, null-model permutation, power analysis, and a machine-readable divergence register. The demonstration on Linear A establishes three adequacy bounds showing that minimal pairs, conditional entropy, and vowel harmony are not supported by the current corpus, and replicates Packard's (1974) statistical defense of Linear B sound values in Linear A with four thousand permutations—the first such replication in fifty-two years. The framework is script-agnostic and transferable to any undeciphered script with multiple editions. It aims to contribute to the literature on undeciphered scripts by developing a reproducible audit and adequacy-testing framework that determines, in advance, whether a corpus can support a given statistical claim. The framework comprises sign-identity resolution, structured uncertainty preservation, multi-witness comparison, adequacy testing, null-model permutation, power analysis, and a machine-readable divergence register. The demonstration on Linear A establishes three adequacy bounds showing that minimal pairs, conditional entropy, and vowel harmony are not supported by the current corpus, and replicates Packard's (1974) statistical defense of Linear B sound values in Linear A with four thousand permutations—the first such replication in fifty-two years. The framework is script-agnostic and transferable to any undeciphered script with multiple editions. The repository contains a complete audit framework for digital corpora of undeciphered scripts, demonstrated on Linear A. The framework includes: Sign identity as authoritative data (not lossy ASCII labels). Structured preservation of editorial uncertainty (damage, reading certainty, measure confidence). Multi-witness comparison with a machine-readable divergence register. Adequacy testing: counts cells vs. observations; determines if a statistic is computable. Null models: document-level permutation preserving length and positional frequencies. Power analysis: calibrates sensitivity of negative results. Entropy estimation: H1 unigram; demonstrates H2 is unestimable. Minimal-pair grid analysis: tests whether a syllabic grid is recoverable. Positional preference testing: edge preference with Benjamini-Hochberg correction. Affixation analysis: tests both existence and direction of affixation. Libation formula segmentation: dual-reading, pairwise precedence. Arithmetic control (KU-RO): null model for "total" reading (z = +13.4). Phonological inheritance test: vowel harmony under five- and three-vowel specifications. Sparse-data estimator testing: whether smoothing can circumvent H2 failure. Packard replication: first in 52 years; internal alternations and name-parallel. SigLA decoder: OCaml Marshal parser; makes 802 documents accessible. Divergence register: machine-readable log of every disagreement. Reproducible pipeline: end-to-end, checksummed, byte-identical rebuild. Regression test suite: licensing hygiene, invariants, published figures. The demonstration on Linear A produces: A systematic correction of AB21/AB22 (sheep/goat) inversion across 14 documents. The first replication of Packard (1974) in 52 years, with 4000 permutations. A decoder for the SigLA database (OCaml Marshal format). Adequacy bounds: minimal pairs below chance; conditional entropy inestimable; vowel harmony undetectable. A positive structural finding: affixation exceeds chance in both registers (z ≈ 6). KU-RO arithmetic control: z = +13.4.