A convergence model of bioelectric, gap-junctional, and Hippo-YAP signalling in oral cancer stem cell maintenance knowledge graph: evidence-tiered snapshot and pipeline (2026-09-28)

Evidence-tiered knowledge graph of the proposed convergence between bioelectric signalling, gap-junctional communication and Hippo-YAP signalling in oral cancer stem cells, together with the pipeline that builds it and the complete validation record. WHAT THIS IS A narrative review proposes a mechanistic model. A model stated in prose cannot be queried, cannot be contradicted by a new paper, and does not record which of its parts rest on direct evidence and which on extrapolation. This deposit is that model converted into a graph in which every assertion is a node carrying an evidence tier, the quote it was extracted from, its source, and its current validity — so a claim can be superseded when stronger evidence contradicts it. CONTENTS (118 assertions, snapshot of 28 September 2026) 29 experimental in oral squamous cell carcinoma, 20 experimental in other oral tissue, 20 human clinical associations, 32 experimental in non-oral systems, 13 mechanistic inferences, 2 computational, 2 unclassified. 55 were curated from the source review; 63 were added by automated literature surveillance. Five untested predictions are held as separate nodes, and 89 claims whose entities could not be grounded are quarantined rather than discarded. The graph's substantive result is an evidence asymmetry, and the three pillars fail differently. Hippo-YAP carries 27 oral-experimental assertions of 63. Gap-junctional carries none of 18, but six are human oral clinical associations, so its deficit is intervention rather than orality. Bioelectric carries 1 of 14, with 12 resting on non-oral systems, and is the extrapolated arm. HOW TO USE IT Open index.html in a browser for the whole graph with filters by pillar, tier and origin — no build step, no server, no external requests. For a database instead, run the ten Cypher scripts listed under "Rebuild the graph" in DEPOSIT_MANIFEST.md, in the order given there, against a Neo4j instance; verify_graph.cypher reproduces the audit values reported in the paper. assertions.csv and assertions.json carry the same content for programmatic reuse. WHAT IS ALSO HERE, AND WHY The deposit includes the full validation record, not only the result: three rounds of blinded expert adjudication with thresholds fixed in writing beforehand, four pre-registrations saved before the analyses they score, identifier and cell-line audits, and a ledger of 32 defects found during development with the direction of each error and the test that now pins it. The pipeline was written with large language model assistance, and several of those defects would have produced a graph that was confidently and silently wrong. The record is deposited because the failures are the part a reader needs in order to judge the result. DEPOSIT_MANIFEST.md describes every file, states what was deliberately excluded and why, and derives its own completeness from the artifact list. LIMITATIONS The graph is small and covers one model. Extraction is reproducible under deterministic decoding but the claim set is only partly so. Steady-state accrual is unmeasured. Adjudication inter-rater agreement was not computed. These are stated at length in the accompanying paper. Zenodo presents the files as a flat list; the group column of deposit_manifest.csv maps each file to its section. LICENCE: CC BY 4.0. The two third-party articles discussed are cited, not redistributed.