Welcome. Inside is a living map of oncology that accretes as institutions privately share attributed, privacy-floored measurements — extreme collaboration for patient health, with attribution and privacy held as one.
Oncology.Study.SpecimenSeal@v1. This is your
“transect”: the already-open specimen onto which co-measurements can be honored.
No contributions yet. Honor a request to record attributed, privacy-floored credit.
Every analysis on the shared map runs on real infrastructure — GPUs, CPUs, storage, network. This view traces how the field's models actually run (training, inference, simulation), ties each quarter's PathwayRuns to the resources they consumed and the domains they served, and shows how projects — within and across organizations — can pool compute by shared objective rather than by bureaucracy. Because raw data never moves — only private cryptographic substrates and derived signal — patient privacy is the default, and compute can be allocated to the most valuable work wherever it lives.
Three workload classes with very different resource shapes and duty cycles — each mapped to the domains it serves on the map:
Multi-node pretraining over integrated multi-omics from hundreds of thousands of patients; B300 blocks cut months to days (seed §1).
Vision-language inference over gigapixel WSIs, mapping immune cells relative to tumor to stratify trial patients (seed §2).
Millions of parallel cellular-level simulations over RWE-grounded digital twins and synthetic control arms (seed §3).
A PathwayRun is a sealed, provenance-bearing execution of a Pathway (a composable workflow
recipe). Every run in 2026-Q2 carries a provenance root that binds what ran to the
exact network resources it ran on, and the domains / subdomains it served:
| Run | Pathway | Domains · subdomains | GPU (type × hrs) | Storage · network | Provenance root |
|---|---|---|---|---|---|
pr_7f21 | Compute.TrainFoundation@v1 | multi-omics · genomics / epigenomics / transcriptomics / proteomics | 512× B300 × 9,800 | 1.2 PB · 400 Gb/s IB | 0x9c4e… |
pr_3a8d | Study.SpecimenSeal@v1 | pathology · genomics | 4× B300 × 12 | 24 TB · 25 Gb/s | 0x1d07… |
pr_b2c5 | Assay.HonorInSitu@v1 | spatial biology · immuno-oncology | 8× B300 × 30 | 8 TB · 25 Gb/s | 0x77af… |
pr_c810 | InSilico.SimBatch@v1 | digital twins · synthetic control | 256× B300 × 4,200 | 300 TB · 100 Gb/s | 0xa1b9… |
pr_5e33 | Map.EmergeComposite@v1 | stratification (immune_strat) | 2× B300 × 3 | 0.5 TB · 10 Gb/s | 0x44d2… |
pr_e0f1 | Consortium.QuarterlyClose@v1 | commons · attribution | CPU-only | 2 TB · 10 Gb/s | 0xf3aa… |
How the quarter's 16,000 GPU-hours were actually spent — one bar, split by share of the total, so the shape of the work is legible at a glance:
The three workload classes have complementary duty cycles. Scheduled against one another, a shared allocation runs far hotter than any single project's private cluster — and the savings fund more science:
Across organizations, allocation follows the shared goal and objective space — compute goes to the work most likely to move patient outcomes, not to whichever org has seniority. Each party keeps single-tenant sovereignty; only derived signal and cryptographic substrates cross the boundary, so pooling never implies data pooling.
| Mode | Siloed | Pooled by objective |
|---|---|---|
| Utilization | ~45% | ~85%+ |
| Allocation basis | org budget / seniority | outcome value of the work |
| What crosses the boundary | (nothing — capacity stranded) | derived signal + crypto substrate |
| Privacy posture | isolated but idle | isolated and busy |
Where each collaboration's workloads actually run. Pick a collaboration to see the institutions, their places, and the datacenter regions involved. Links carry derived signal over private cryptographic substrates — raw data stays local to each site.
Because each analysis's provenance chains down to the CPU / GPU / storage / network it ran on, the whole pipeline is auditable and, crucially, plannable: you can price any future analysis from the measured cost of the ones that produced today's map.
| Provenance layer | Bound resource | Planning signal |
|---|---|---|
| PathwayRun | GPU type × hours, node count | accelerator demand curve |
| Coverage root / seal | storage class × volume | hot / warm / cold tiering |
| Cross-party contribution | network egress + crypto compute | interconnect & secure-agg budget |
| Attribution ledger entry | who ran it, on whose capacity | chargeback & fair-share pooling |
Rolling the measured footprint forward, against the roadmap of models to train and studies to run — the basis for a structured, multi-year offtake (predictable CapEx vs. volatile OpEx, seed §4):
| Period | Headline workload | GPU-hours | Peak nodes | Note |
|---|---|---|---|---|
| 2026-Q2 (actual) | foundation v1 + sims | 16,000 | 512 | baseline measured here |
| 2026-Q3 (plan) | + immune-strat inference scale-out | 21,500 | 640 | pooled off-peak with partner org |
| 2027 FY (plan) | pan-cancer v2 pretrain | 88,000 | 1,024 | 3-yr dedicated block, single-tenant |
Pooling and planning use resource telemetry and derived signal only. Raw patient data never leaves its sovereign boundary; parties collaborate through private cryptographic substrates — federated learning, secure aggregation, and proofs of computation — so the default is privacy, and shared compute never means shared records.
This report is a way to discover the idea behind the visualization — from the problem in oncology, through the category-theoretic machinery, to the collaboration mechanism that makes the map draw itself. Read top to bottom, or jump to a section. Two kinds of links appear throughout.
Modern oncology's frontier — pan-cancer multi-omics foundation models, spatial-biology patient stratification, in silico trials with synthetic control arms — is gated less by compute than by cooperation. The signal that would move patient outcomes sits in silos held by institutions that compete, and that are bound by strict privacy law. Compute is necessary but not sufficient.
Represent the oncology Shared Domain Graph (SDG) not as a fixed taxonomy but as a living latent space: a hypergraph whose most valuable regions are hidden until collaboration brings them into being. The organizing claim is a single equation — a pullback that lets two things usually traded against each other be held at once:
Full credit, zero identity leakage. When credit is guaranteed and re-identification is impossible, the rational move for a competitive institution flips from hoard to contribute.
The map holds 34 conceptual domains in 9 categories, joined by morphisms and grouped into hyperedge regions. Tap a category to ▸ highlight it live (tap again to clear):
Every domain on the map is oriented toward Patient Outcomes — survival, response, quality of life. Hover any node on the map for its description and its grounding in the seed context.
Formally, Onc is a category. Its objects are the domains; its morphisms are
structure-preserving maps (a practice area projects into Law-like bases, an entity grounds
a domain, a temporal constraint binds a process); its hyperedges capture irreducibly multi-way
relations rendered as regions.
The interesting domains are pullbacks (fiber products) D₁ ×B D₂:
a composite that exists only once both of its legs do. These are hidden by default and
crystallize in situ when a co-measurement supplies the missing leg. Tap go
to see where each lives:
The pullback of §2 works because attribution and privacy operate on different layers and never contend for the same object:
| Layer | Object | Guarantee |
|---|---|---|
| Data | Raw records / sequence | Never leaves the sovereign boundary |
| Derived | Calls, panels, aggregates, model cards | Consent-bounded, privacy-classed |
| Ledger | Contribution + credit | Fully attributed; references only the derived layer |
The shape is borrowed from field “water reporting” and rebuilt for oncology as shared assay reporting. None of the prior work's code or assets are used.
| Water reporting | Shared assay reporting |
|---|---|
| Transect walk | A lab's own study on a specimen / cohort |
| Bounded sample point | The already-open specimen |
| Private request for extra readings | Private request for an extra derived assay |
| Honor on the same walk | Honor in situ on the same specimen |
| Shared water map | Shared oncology SDG map |
| Cross-project opportunity | An emergent composite domain |
Each turn of the loop adds an attributed contribution to the commons, under the privacy floor, and may emerge a new composite — five pathway templates encode it:
The concept is stated as pre-registered, falsifiable hypotheses — the ones this demo exercises:
H-ONC1Honoring a request materializes a new pullback with no top-down edit.
H-ONC2Attribution and privacy hold simultaneously (sovereign attribution).
H-ONC5An already-open specimen: near-zero marginal cost, high network value.
Everything on the map is a view onto a signed-off collaboration bundle. Start anywhere:
Oncology's hardest problems — pan-cancer multi-omics models, spatial-biology patient stratification, in silico trials — require compute at serious scale.
They also need it run with efficiency, transparency, and governance — reliability built through redundancy across collaborative dimensions, not siloed isolation.
The signal that would move patient outcomes is scattered across institutions that cannot safely combine it.
This is an interactive model of a different arrangement: a living map of the oncology SDG that accretes in situ as competitors privately share attributed, privacy-floored measurements — and new regions of knowledge emerge that no one directly authored.
And yet “no one directly authored” is not the same as unauthored. Every collaborator authors the context from which a new region crystallizes — the questions asked, the panels run, the annotations left, the objectives held in common.
Large language models compress that distributed human context into dense representations, and honoring a co-measurement hardens that compression into a durable, verifiable substrate — the way sustained pressure turns loose carbon into diamond.
So each emergent region carries the fingerprints of everyone whose context made it possible, even when no single hand drew it.
Every observation, panel, and model card carries provenance and a credit split. Contributing is on the record.
Single-tenant sovereignty, consent scope, federated computation. Nothing re-identifiable is ever exchanged.
Sovereign Attribution = Privacy × Attribution. Full credit with zero identity leakage, held at once.
A hypergraph of 34 conceptual domains of oncology — molecular, imaging, clinical, therapeutic, patient, compute, governance, and the collaboration fabric — all organized around patient outcomes (survival, response, quality of life). The most valuable domains are latent pullbacks, hidden until collaboration brings them into being.
Grounded in domain-expert context from the field — the challenges and opportunities around pan-cancer models, spatial pathology, and data sovereignty.
Want the full picture? The Latent-space report is a guided, ten-section tour that makes the whole concept discoverable — from the problem in oncology, through the category-theoretic machinery (objects, morphisms, pullbacks), to the collaboration mechanism that makes the map draw itself. It carries two kinds of links: ▸ interactive ones that act on the live map (filter a category, reveal latent nodes, emerge a specific pullback), and ↗ document ones that open the underlying bundle artifacts — registry YAML, technique specs, pathway templates, sidecars, hypotheses — in a new tab.
Tap a category to highlight it live on the map (tap again to clear):