A knowledge graph for culture and developmental history where nothing is a bare fact: every statement is a claim, with sources, a confidence and a trust tier.
Early build Docs partly planned
Illustrative diagram. The field names come from the engine’s schema; the example values are invented.
Kosmographica aims at a record of human thought, culture and development through an integral, developmental lens. It does not start from scratch: it pulls together four existing sibling datasets, namely comparative mythology, guru-to-disciple lineages, developmental models such as AQAL and Spiral Dynamics, and a historical timeline, into one graph.
The design is a thin universal core plus pluggable domain modules. The core defines only what every domain shares: entities, relationships, claims and sources. Modules such as Religion and Mythology add their own types and vocabulary on top.
The engine is a Python service on Postgres with an author-and-verifier pipeline, and a Next.js web app. A second Next.js app is a read-only public encyclopedia that only shows records past a trust threshold.
Illustrative diagram Drawn from the engine’s database models and the decision log. The boxes and field names are real; it is a simplification, and no data is shown.
The unit of trust is the claim: a sentence about an entity or a relationship, a confidence between 0 and 1, the exact supporting text from the cited sources, and a disputed flag. A contradicting claim opens a dispute and both are kept with their provenance, rather than one overwriting the other.
Records are never changed in place. A new write supersedes the older one, so every change, including each AI write, stays auditable and reversible. Each record also notes which generator and batch produced it.
The tiers are defined in code. The decision log (ADR-013) sets a publish-then-verify rule: an AI author may write straight to the canonical store, and an automated verifier plus later audit decide what readers can see. The public API hides the lowest tier.
Specified, not confirmed in code
The core design doc also describes derived confidence bands (high from 0.8, medium from 0.55, low from 0.3, speculative below), claim types such as historical, mythological and doctrinal, and a developmental layer that records stage, state and quadrant as claims of their own. We read these as the design; we did not check that every one is implemented.
Adapters for two sibling datasets (Mythographica and Sacred-Lineage). Records sharing an external id are linked automatically; fuzzier name matches go to a review queue.
A provider-agnostic LLM client drafts claims; a deterministic structural gate and a verifier check them before they count. An eval command compares the verifier with a plain lexical baseline.
A read-only Next.js app for entities, claims, disputes, a review queue, reconciliation and search.
A second Next.js app with browse, search and entity pages, each showing claims with trust badges. It sends no tier override, so the engine’s public clamp applies.
Records carry a sensitivity level (public, sensitive, sacred, restricted), and sacred or restricted material is gated.
The plan lists a D3 neighbourhood view around an entity as shipped. It is not yet a general graph editor.
A cosmograph is a culture’s map of reality. The repository ships a catalog of 162 of them as seed data (an older 74-row CSV is the starting subset), with tradition, region, dates, topology, the human’s position and the path of liberation. Five real rows:
| Cosmograph | Dates | Topology | Human position | Path | Primary sources |
|---|---|---|---|---|---|
| Vedic Three Worlds | 1500 BCE | Triple Realm | Sacrificer | Ritual ascent | Rig Veda |
| Sumerian Cosmos | 3500 BCE | Layered | Servant of gods | Divine order | ETCSL, Enuma Elish |
| Egyptian Cosmos | 3000 BCE | Layered Geography | Participant in Ma’at | Afterlife navigation | Pyramid Texts |
| Zoroastrian Cosmos | 1000 BCE | Dualistic | Cosmic combatant | Frashokereti | Avesta |
| Jain Lokapurusha | 500 BCE | Cosmic Person | Jiva | Liberation | Tiloya Pannatti |
Source: data/cosmographs/catalog.json. Copied as text; no images are shown.
About the images
The catalog does not contain image files. Every one of the 162 rows has image metadata that points to a Wikimedia Commons file, with its title and licence (73 are marked public domain, the rest mostly Creative Commons). A script is meant to fetch them. We have not downloaded or reproduced any of them here, so this page shows no cosmograph images.
The core meta-model, the Religion and Mythology module (over a thousand lines), the architecture, governance and AI workflow papers, and a large ontology folder with a cosmograph catalog spec and a gap analysis.
Architecture decisions, including the move to publish-then-verify.
The glossary and the Philosophy and Science, Art and Culture, Polity and Society and Technology modules are listed as stubs in the plan. About two dozen doc files carry a stub or outline note near the top, so some may be older than the plan’s status table.
Waves 1 and 2 are marked shipped. Wave 3, the public encyclopedia, is in progress: the site, stable URLs and graph view are done; the vector search backend and a human review layer are pending.
Last commit is dated 14 July 2026. The README still says design stage; the plan and the code have moved well beyond that.
Engine code with models, an API, reconciliation, a verifier, a command-line tool and migrations; an audit console; a public site; a seed catalog; two database snapshots.
A snapshot from 4 June 2026 is recorded as roughly 4,376 entities, 3,989 relationships and 6,985 claims. That count is from the snapshot’s own readme, not something we counted.
The engine has a tests folder with a few dozen test functions across fourteen files. We counted them but did not run them, and the web apps have none.
We did not start the engine, restore the snapshot, or run either app. Statements about behaviour come from reading code and docs.
Vector search, a human review and promotion layer, the four stub modules, and a licence: the README says licence information will be added soon.
AI-authored claims are only as good as the verifier and the audit. Treat anything below human_reviewed as a lead, not a finding, and respect the sensitivity flags.
A local project folder named kosmographica. No public address is stated in its README.
docs/PLAN.md is the master index and build order; docs/core-meta-model.md is the model.
The engine’s own README describes a Docker database and a command-line tool; the public site’s README describes how to point it at a running engine.
