2653
concepts mapped across the library
VictorOS treats a disease as a landscape to be mapped. It builds and maintains a governed map of what is known, uncertain, contested, and unexplored, with every claim carrying its provenance and evidence tier. One map per disease, each one a guide you can read today.
VictorOS is named for Victor A. McKusick (1921–2008), whose pioneering work helped establish the foundations of modern rare disease research. McKusick believed that meaningful progress begins with the careful organization of knowledge. His Mendelian Inheritance in Man, the first comprehensive catalog of inherited disorders, transformed scattered observations into a usable system for discovery.
VictorOS carries that principle forward by creating continuously evolving AI specialists that develop deep expertise in individual diseases, turning fragmented information into actionable understanding.
Every statement in a guide is tied to a published study, shown with its evidence strength and a link to its source. Nothing is filled in from guesswork. That has a visible cost: if a source has not been mapped yet, it does not appear, so a gap here almost always means not yet mapped, not unknown to medicine. Newer guides are marked as early maps and grow as sources are added.
The full account of how guides are built, why they abstain, how you can verify any claim yourself, and who stands behind the work is on the how it works page.
The asset is the disease model that organizes the literature into a coherent, evolving territory. A paper that cannot be placed on the map (anchored to a node, carrying provenance, assigned a tier) has only been read. Placement is what brings it into the model.
What is known, suspected, debated, emerging, and unexplored. Every claim carries source, date, evidence tier, replication status, and known contradicting evidence. Observation stays strictly separate from interpretation; collapsing the two is the origin of most medical misinformation.
The breakthrough is storing relationships, not papers. Questions that would take a full literature review become single graph traversals across the causal spine:
A runnable, standard-library-only engine fills the front of the pipeline without ever letting unverified claims into the model. The canon gate does not move.
A live PubMed pull adds provenance-bearing source nodes, writes full observation records, and scans abstracts for proposed typed relationships into a review queue, where they wait for a human to promote them before any reach canon.
Turning free text into typed claims is the error-prone step where a model invents plausible but unsupported relationships. So that step stays governed: a human promotes queue entries before they are committed.
Run traversals across the graph, regenerate the offline navigator, run drift checks that flag single-source strong claims and genotype-only phenotypes, and trace a new high-tier source's second-order consequences before acting on it.
The governing principles are enforced at the write boundary. They are refusals, not guidelines.
Every node and edge must carry source, date, source type, evidence state, and layer, or the write is refused. An assertion without a traceable source fails to count as a claim at all.
A single source cannot push a claim above the ceiling for its type. A lone case report can never assert "well established." Strength comes from independent corroboration or replication.
"Five cohorts report elevated cytokine X" is an observation. "Cytokine X drives progression" is a hypothesis with its own evidence for and against. The two are kept strictly apart.
The model always answers not only "what is known" but "how well, by whom, replicated how many times, and disputed by whom." Absence of evidence is recorded as a gap, not omitted.
When two representations are equally valid, prefer the one that claims less. A marked unknown is more valuable than a confident invention. Significance emerges through recurrence.
Twelve named failure modes are machine-checked: single-source strong claims, recency-only promotions, drugs with no mapped target, genotype-only phenotypes, over-certain plain language, and more.
One of a growing library of guides, and the deepest, the one we use to stress-test everything else. Seeded from eight real PubMed sources with DOIs and extended by a live pull of 2025 literature, it deliberately exposes the hard cases rather than hiding them. Browse all the guides →
Schema v1.1 added a first-class non-genetic layer. All 12 molecular subtypes now carry sourced penetrance and expressivity, and vascular EDS carries explicit Modifier nodes (sex, mutation class, pregnancy), discharging drift check DR-008 with evidence rather than by papering over the gap. hEDS stays defined by comorbidity and a proposed polygenic basis rather than by a single gene.
symptoms-without-mechanism automatically surfaces chronic widespread pain and orthostatic intolerance, because the map has no path from them back to a mechanism.
The hEDS / POTS / MCAS relationship is held as two disputed edges carrying both supporting and contradicting sources. The 2025 matched-cohort study (118,256 individuals) was reviewed and redirected onto the hEDS subtype, strengthening the disputed POTS edge as a supporting source while preserving its disputed flag; controversy reinforced with evidence and kept visible.
The reusable substrate is the product: a governed, continuously-extendable knowledge environment that can be stood up for any condition. The library grows steadily, and every guide is built only from sources you can check.
Browse the rare disease guides →Because every map is built on the same substrate with the same controlled identifiers, a characteristic that appears in two conditions resolves to the same node. The commonality view lists those overlaps and nothing more. It is a read-only join over nodes that were each grounded independently, in their own condition, against their own source. Each overlap is presented as a bare commonality. No relationship between the conditions is asserted, suggested, or implied, and no link connecting them exists in the graph. Every appearance keeps its own source beside it, so the chain of custody stays separate per condition.
View cross-condition commonalities →VictorOS organizes evidence; it does not practice medicine. It does not give medical advice, diagnose, or recommend treatment. It builds the evidence landscape so patients, caregivers, and advocates can engage their care teams as informed participants. Disease facts in each map are based on articles retrieved from PubMed and cited with DOIs in the canon and observation files.