RUNA-1: a typed biosemiotic knowledge-graph embedding for European ecology (v1.4.0)

RUNA-1 is a typed knowledge-graph embedding (PyKEEN, model: BoxE, embedding_dim 128) over open European ecological data, in which species, discretized environmental-state nodes, detected-community nodes and signal/element nodes share one geometric space where proximity encodes ecological and biosemiotic relatedness.
On top of conventional ecological relations (predation, pollination, mycorrhizae, parasitism, etc., mapped where possible to the OBO Relations Ontology), it adds biosemiotic sign-relations trained as ordinary typed edges: indicatorOf (a species as a sign of an environmental state, from EIVE/Ellenberg) and keystoneSignProducerIn (keystone sign-producer within a detected community).
New in v1.4.0 — the navigation layer + blind-spot fixes. v1.4 adds navigatesBy, linking a species to how it orients: magnetic_compass, sun_compass, star_compass or olfaction_homing. This makes explicit that a magnetic, celestial or chemical sense is used for finding the way (salmon and eel by smell + magnetism; blackcap by stars + magnetism; honeybee and painted lady by the sun). It ingests literature-confirmed magnetoreception facts surfaced by using the model itself as a screen and then verified against the primary literature (wood mouse, common toad, common frog, chaffinch, Atlantic herring and European sprat at species level; blackbird, song thrush, chiffchaff and goldfinch at group level). It removes a 15-edge data-contamination cluster that had buried the ash-dieback host signal, so Hymenoscyphus fraxineus now correctly resolves Fraxinus excelsior as its host; and it anchors sparse endemics to real interactions (Saimaa ringed seal diet, aspen-bracket host, stag-beetle deadwood dependency).
v1.3 carried forward — the signal-web layer. Following the relational semiosis of Kohn's How Forests Think: castsSignal (which species produce which kinds of signal — sound types song/call/alarm/flight-call/echolocation, from xeno-canto) and readsSignal (which species read the elements as signs — the magnetic and electric fields, from the confirmed-taxa literature). Every signal edge is provenance-typed (MEASURED at species level vs confirmed-at-group level).
v1.2 carried forward: 96 place-based dependency edges curated by the nine agnt eco agents and fact-checked by a multi-model consensus (claude-opus-4-8 and gpt-5.4 reading the cited sources; included only where Claude confirmed and GPT did not refute).
Priority claim: the first trained relational embedding that operationalizes biosemiotic relations for ecology.
Validation (honest scope, v1.4.0): held-out link prediction filtered MRR 0.301 (Hits@10 0.50), unchanged from v1.3. On the independently-validated indicator axes (model placement vs real GBIF occurrence × environment, non-circular): temperature Spearman rho = 0.576 vs CHELSA bio1 and soil pH rho = 0.316 vs SoilGrids (above the expert-input ceiling of 0.213). These are within run-to-run variance of v1.3's 0.602 / 0.340 — a same-data control run under a different random seed placed them at 0.569 / 0.298 — so v1.4 holds the validated axes. Moisture and nutrients remain unvalidated; keystoneSignProducerIn, the signal layer and the new navigation layer are learnable but not yet independently validated.
Contents: the frozen v1.4 triple set, the relation schema, the full derivation/verification code, the trained BoxE v1.4 model, the multi-axis validation data, a benchmark plus the v1.4 result/variance notes, documentation, and a SHA-256 manifest. Derived from GLOBI, Mangal, EIVE 1.0, GBIF, CHELSA, SoilGrids, xeno-canto, plus place-based agent research and literature-verified sensory facts.