A knowledge graph of the Bible — where every connection carries its provenance

Under the hood the whole graph is deliberately boring: two database tables and an alias index. A node is one canonical entity — a verse, a person, a place, an episode, a theme, an object, a Hebrew or Greek lemma — with a stable identifier like verse:genesis:1:1 or person:god, one of 24 node types currently in the public graph. The alias index maps the names authors actually write to those canonical entities, so "the LORD," "YHWH," and "Yahweh" resolve to one node instead of three.

A link (an edge) is where the interesting part lives. Every edge carries a relationship type (522 in public use — appears_in, quotes, thematic_parallel, and so on), a provenance label, a source reference — a citation handle saying exactly which process created it and from what — a stored evidence excerpt, and a confidence score. Edge identity is unique on the combination of source, target, type, and source reference: the same two entities connected by two different pieces of evidence are two edges. That is deliberate — evidence records are never merged into an unaccountable blur.

Every edge is labeled with one of five provenance values. In the map itself you can filter by them — the "Evidence source" control. If you want the taxonomy in one sentence: two classes are deterministic, one is gated AI, one is hand-authored, and one is private curation that never appears in the public graph.

The inference class is the one that deserves suspicion, so it is the one wearing the most restraints. A model proposing a connection must satisfy every gate below; failing any one of them means the proposal is rejected before it ever touches the graph.

A provenance taxonomy earns trust by admitting its edge cases, so here are mine. First: a large share of the cross-reference parallels — quotation, verbal, and thematic links across the canon — were originally generated by the translation committee's model (grok-4.3) during the v3 run, then deterministically validated against the canon inventory and imported. They carry the "reference" label because the import path is deterministic, but their origin is AI. The raw generation records for that stage are public in the audit trail on Hugging Face, so that origin is downloadable, not just admitted.

Second: the inference class currently mixes true model output with locally authored pattern ontologies that made no model call. The discriminator is the source reference prefix — ai: for model-proposed edges, codex: for authored patterns — which is served on every edge but was, until this page, explained nowhere.

Third, on human approval: I will not pretend there is per-edge human sign-off at this scale — this is a solo project. Human review exists as targeted layers instead: reviewed interpretive ledgers for ambiguous cases, and every entity prose page passes a two-model gate — one model writes, a different model verifies against the evidence, and every reference the page cites must be on a server-side allowlist built for that entity. Pages that fail are held back.

None of the above asks to be believed. Open the Firmament, select any connection, and the evidence inspector shows its provenance label, its source reference, its confidence, and the full stored evidence excerpt — to anonymous visitors, with no account. The counts on this page are the public graph only: private research layers are excluded, and a link is counted only when both of its endpoints are public.