See what supports a claim—and where the evidence stops.

1. The graph ontology

Every concept is a typed node (cell, molecule, receptor, pathway, tissue, pathogen, disease, therapy, biomarker, trial and more) with a stable ID. Every link is one directional claim — subject, predicate, object — with a stable relationship ID of the form REL-subject-predicate-object, a mechanism, a confidence and, where known, a tissue.

Release graph-0.3+draft-ext-4: core graph v0.3 has 260 concepts and 395 links. Extensions draft-ext-4 add 70 draft concepts and 154 draft links; drafts are always labelled Draft and never counted as approved evidence. Runtime total is 330 concepts / 549 links (same snapshot used on the homepage and footer).

v0.3 (unified): Immune City Layer 1 + Mechanism Layer 2 + base44 KG v0.2 merged; gap audit applied.

2. Relationship signs

• + carries an effect forward (e.g. produces, binds, activates, matures).
• − flips it (e.g. phagocytoses, kills, contains, impairs).
• r+ is descriptive; a positive effect flows from object back to subject (e.g. requires).
• Other predicates are descriptive only and are not used in simulations.

3. Evidence review

A source supports a link only when a curator maps it to that exact relationship ID and records that it supports the claim. Citations attached to a concept describe the concept and never count as support for a link. Coverage figures on the Evidence page use only claim-checked support.

Inclusion criteria: peer-reviewed primary research, systematic reviews or regulatory documents with a PubMed ID or DOI; identifiers are checked against PubMed and Crossref metadata before approval. Approval requires claim_support = supports (plus a verified/matched identifier) and records the reviewer and date. Live claim-support counts are on Evidence — Gate C progress is real but sparse overall versus runtime link count.

Sign-in and curation use Neon Auth with roles in user_roles. New accounts receive the reader role; admin is provisioned explicitly (not “first signup wins”). Approved citations are served by a server-side filter for public pages.

4. Model assumptions

• Perturbation (graph): signed spread along + / − / r+ links, decaying 28% per hop, weighted by confidence (high 1.0, medium 0.8, low 0.6) or a measured weight, up to 4 hops. It shows direction, not magnitude in real units.
• System simulator: a small set of coupled equations for pathogen, innate and adaptive response, cytokines, damage, memory and exhaustion, in normalised units with hand-set parameters. It is not fitted to patient data.
• Patient layer: transparent additive rules; each score lists the rules that produced it. Educational illustration — not patient-specific advice.
• Ask Immune Super Intelligence: queries are classified, graph context is retrieved, reviewed relationship-level citations are attached when they exist for those links, then an AI model may synthesise only from that extract (prompt ask-grounding-v3). Sparse approved cites mean Ask is graph-retrieved, not fully evidence-grounded. Unsupported, omics-not-wired and personal clinical questions abstain without model synthesis. Answers can still be wrong.
• Research workbench / literature / model registry: available on Research as candidate workflows and interface contracts (phases B–E partial) — never auto-promoted into the curated graph or live predictors.
• Advanced model adapters (omics, structure, pMHC/epitope, genomics, agents) are modular interface contracts only — none are implemented in production. Dry-run stubs return typed not_available and never invent scores or ranks. Full registry with request-shape previews on Research.

4b. Platform phase status

5. Limitations

• Most links have not yet been claim-checked against a source; claim-level approved support is still sparse overall (single-digit approved relationship claims vs hundreds of runtime links) and does not imply clinical validation.
• Exemplar progress is uneven: TNF→RA on-path edges have approved claim support; PD-1 / PD-L1 is partially reviewed; type-I IFN→SLE remains largely scaffold-ish outside a few reviewed edges.
• The Immunology Academy foundation track on Mechanism is a navigable teaching scaffold — not a finished curriculum, certification, or Gate C coverage claim.
• Links simplify dose, timing, species and individual variation.
• Tissue context is sparse; many effects are recorded without one.
• This product educates with references; it is not medical advice, clinician decision support, or patient-specific guidance.

6. Accessibility

The knowledge graph has a text table view with every link, keyboard focus is always visible, and detail drawers close with Escape. Please report barriers via the curation team.