An integrated research environment for asking what can actually be learned from Linear A without pretending the script has been deciphered. The project combines structural testing, corpus reconciliation, modern epigraphy, publication monitoring, linked archaeological data, and a strict prospective-evidence firewall.
The strongest result is structural: recurring construction families are highly predictive, but the evidence does not support turning those regularities into claimed meanings or grammar.
These are project results, not accepted translations. Each statement is deliberately narrower than the data it summarizes.
Within H11 model selection, the lexical/construction-network model scored 16, compared with 6.5 for a common-chunk/null model and -4 for a limited productive-morphology model. The leader remained stable under leave-one-independent-evidence-out testing.
Interpretation is structural only. It does not identify meanings, morphemes, grammatical categories, sounds, or a historical language family.
Across 204 masked tests, the shared-chunk model achieved MRR ≈ 0.944, 100% top-five recovery, and about 89.7% rank-one recovery. Adding family-specific edge information did not improve the model.
This is evidence for stable constructional organization, not proof of productive affixes.
40 of 43 H11-C families occur across more than one site. That makes a purely local scribal or site-specific explanation less attractive.
At the same time, support specialization was absent and no FDR-significant scribe concentration survived.
Exact one-edge density was not exceptional, secure positional slots did not survive FDR control, and a frozen productive-operator transfer test performed 0/18 while the lexical/formula baseline recovered 18/18.
That negative evidence is part of the result. We do not hide failed hypotheses.
The next phase moves outward: better objects, better provenance, independent controls, modern images, and automated discovery of genuinely new material.
Recent candidate publications discovered through scholarly metadata APIs. Results are leads for human review, not automatically admitted evidence.
These are the services we should actively integrate. The priority labels reflect usefulness to this project, not prestige.
Some are upstream data sources, some are toolkits, and some are competing hypotheses. Those roles must remain distinct.
Network access makes contamination easier, not harder. The web platform therefore separates discovery from admission.
RILA-S1: obtain and ingest the full object concordance lawfully, metadata first.
Post-2023 objects: continue object-level watch without opening sealed predictions.
Landau 1958 pp. 268–273: obtain lawful access for controlled onomastic/reference work.
Build canonical object records with authority IDs, publication history, image/3D links, place identifiers, corpus aliases, conflict flags, and a machine-readable contamination state. Add automated literature and dataset monitoring around that object graph.
BOASTRADAMUS