The 17 Demonstrations¶

One per engine, per agent, and for the TSC disease program. Each is written to the same shape: a person at stake, the factory compressing their odyssey, and an honest statement of what the room is actually seeing.

Every demo carries exactly one label. This is not decoration — it is the project's honesty gate:

Label	Meaning
LIVE	Running now, in front of the audience, on real input, on this box
REPRESENTATIVE	A pre-computed or curated result standing in for a long or gated step — said out loud
BURST	Running live, but on remote GPUs over the private mesh — say "elastic burst", never "all on one box"

Current distribution: 13 LIVE · 4 REPRESENTATIVE — and as of 2026-09-15 all seventeen RUN, on real input, writing a transcript to demo/transcripts/.

The REPRESENTATIVE ones are not stubs. Each runs the part of its pipeline that is not gated and states the boundary out loud: E1 loads the real GIAB HG002 genome into the variant store and runs the ACMG SF panel but does not call variants (Parabricks); E3 generates and runs real RDKit chemistry on candidates but does not dock them (MolMIM/DiffDock are gated NIMs, now correctly registered planned); E7 folds and optimises on CPU. What each does NOT do is printed in its own output, not buried here.

A LIVE label is a claim about the demo, not a promise the box is up. A LIVE demo needs the platform running and its corpus seeded; ANTHROPIC_API_KEY is needed for answer synthesis, not for retrieval. Ask the runner rather than trusting this page:

.venv/bin/python scripts/run_demo.py --check-all

It refuses to run a demo labelled LIVE whose service is unreachable rather than returning a canned result — so the honest count on a cold box is lower than the count above, and it says so.

Every demo states decision support, not diagnosis the first time clinical output appears.

Engines¶

E1 · Genomic Foundation — "The variant that was always there"¶

Label: REPRESENTATIVE (LIVE after Parabricks · G2) Audience: clinicians, general · Runtime: 6 min

Weight. A child has had three years of tests and no answer. Their genome was sequenced eighteen months ago; the answer was in the file the whole time.

Compression. Load HG002 — a publicly consented reference sample, never a patient — into the variant store. Query the region. Apply ACMG secondary-findings logic. Show the variant surface: SNVs, indels, CNVs, structural variants, and the transition/transversion ratio as a quality signal.

What actually runs: DuckDB variant store, ACMG SF module, GWAS association (156 tests green). What does not: alignment and variant calling. Parabricks is not installed, so FASTQ→BAM→VCF is pre-computed. Say so. The "under five hours" claim belongs to Parabricks, not to this demo.

Hope. The same store answers the next child in seconds.

E2 · Precision Intelligence — "Ask the evidence layer a question"¶

Label: LIVE (needs ANTHROPIC_API_KEY + seeded Milvus) Audience: clinicians · Runtime: 5 min

Ask in plain English: "What variants are associated with frontotemporal dementia?" The engine retrieves across ClinVar and AlphaMissense vectors, reasons over them, and returns cited evidence — including VCP on chromosome 9, the thread E3 and E7 pick up.

Impressive because the citation is checkable. Click through to the source record. Honesty: retrieval quality depends on what is indexed; an empty collection returns nothing, and the demo should show that failure mode rather than hide it.

E3 · Therapeutic Discovery — "From one protein to a hundred candidates"¶

Label: REPRESENTATIVE (LIVE/BURST after MolMIM + DiffDock · G3/G4) Audience: general — this is the flagship · Runtime: 8 min

Seed compound CB-5083 against p97/VCP. Generate novel analogues, dock them, score, rank.

What actually runs today: RDKit property calculation, the ranking and filtering logic (124 tests). What does not: MolMIM generation and DiffDock posing — 252 combined code references, neither installed. The molecules and poses shown are pre-computed. If the NIMs are x86-only, this becomes BURST and the narration must say "elastic burst", not "one box".

Do not run this as LIVE until §4 of the gated build guide is complete. It is the most-watched demo and the easiest to overclaim.

E4 · Clinical Imaging — "The scan that had already answered"¶

Label: LIVE · Audience: cardiologists, radiologists · Runtime: 7 min

Coronary CT → measured stenosis → CAD-RADS 2.0 category with modifiers (P, HRP) → cross-modal hand-off to genomics.

This is the strongest demo in the catalogue and the only one already proven end-to-end on this box: API on :8524, 1,365 tests green, accuracy verified 2026-08-13. It reads a scan, produces a graded report, then names the genomic follow-up — 9p21, LPA — and hands to Engine 2 · Precision Intelligence.

Honesty inline: decision support for a qualified clinician; CAD-RADS is a reporting standard, not a diagnosis.

E5 · Precision Oncology — "The molecular tumour board, in one afternoon"¶

Label: LIVE (needs seeded Milvus) · Audience: oncologists · Runtime: 7 min

A pediatric solid-tumour case (demo/pediatric_oncology_case.json). Variant → actionable target → therapy options → evidence tier. 556 tests green.

Impressive because it shows the tier, not just the answer: what is FDA-approved, what is off-label, what is trial-only. Clinicians trust the gradation more than the recommendation.

E6 · Cardiology — "Risk that changes management"¶

Label: LIVE (needs seeded Milvus) · Audience: cardiologists · Runtime: 5 min

demo/cardiology_risk.json → structured risk with the reasoning shown. 1,966 tests green — the deepest suite in the platform.

Pair with E4. Imaging finds the lesion; cardiology contextualises the risk; pharmacogenomics (A3) decides whether the statin will work. Three engines, one patient — that sequence is the point.

E7 · Structural Biology — "The shape you have to fit"¶

Label: REPRESENTATIVE (LIVE after ESMFold · G5) Audience: protein scientists · Runtime: 6 min

PDB 5FTK — cryo-EM p97/VCP with ADP bound (a nucleotide, not a drug). Show the druggable pockets a molecule must fit.

What actually runs: protein search and re-ranking, developability scoring (34 tests). What does not: ESMFold prediction and ProteinMPNN design — both need CUDA, which the environment serving them does not have. Structures shown are deposited PDB entries, not predictions made in the room.

The engine capability structural-biology-engine:8581 is planned — nothing binds that port. Demo the model-level services (8570/8571/8578), not the engine endpoint.

E8 · Single-Cell — "Nine populations from one sample"¶

Label: LIVE · Audience: researchers · Runtime: 5 min

Real scanpy on the bundled PBMC 3k: QC → normalise → HVG → PCA → neighbours → Leiden → marker DE, then annotate clusters by marker overlap. Verified output: 2,700 cells → 9 clusters → CD4 T, CD8 T, B, NK, CD14+ and FCGR3A+ monocytes, dendritic, megakaryocytes.

Impressive because it is small and completely real — no gated model, no GPU, runs on the Grace cores in minutes. After E3's honesty caveats, this one lands harder precisely because nothing is withheld.

Prerequisite: pip install scanpy anndata (ungated, step 0 of the gated build guide).

Intelligence Agents¶

Each agent reasons over a curated corpus and returns cited output. All eight are LIVE once Milvus is seeded and ANTHROPIC_API_KEY is set; none needs gated software. Payloads already exist in demo/requests/.

A1 · CAR-T — "Why this construct, for this patient"¶

LIVE · 6 min · cart_query.json Eleven collections — constructs, assays, manufacturing, safety, biomarkers, regulatory. Show construct selection with its safety counterweight: on-target off-tumour risk and CRS/ICANS monitoring. 415 tests. The reference implementation — the only service using the governance gate.

A2 · Precision Biomarker — "The marker that changes the decision"¶

LIVE · 5 min · biomarker_query.json · 709 tests Predictive vs prognostic, stated explicitly. Clinicians conflate them; the agent does not.

A3 · Pharmacogenomics — "Two patients, same dose, different outcome"¶

LIVE · 6 min · pharmacogenomics_query.json · 1,001 tests The most rigorous content in the platform — 44 CPIC pairs across 13 genes, all verified genuine. Show CYP2C19 → clopidogrel, or DPYD → fluorouracil where the stakes are toxicity, not efficacy. Honesty: CYP3A4 substrate relationships are pharmacology, not CPIC guidance — label them so.

A4 · Precision Autoimmune — "Before the third flare"¶

LIVE · 5 min · autoimmune_query.json · 455 tests

A5 · Neurology — "NIHSS, and what comes next"¶

LIVE · 5 min · neurology_nihss.json · 208 tests Now on UI 8535 / API 8536 after the port re-seat.

A6 · Clinical Trial — "The trial that was open all along"¶

LIVE · 6 min · trial_match.json · 769 tests Close the loop: hand E3's discovered molecule to this agent and ask what it would need to enter trials. That hand-off is the single most effective closing move in the catalogue.

A7 · Rare Disease — "Ending the odyssey"¶

LIVE · 6 min · rare_disease_diagnose.json · 206 tests HPO phenotype terms → ranked differential with evidence. The demo closest to the mission.

A8 · Single-Cell Intelligence — "The engine computes, the agent interprets"¶

LIVE · 5 min · single_cell_query.json · 185 tests Run immediately after E8. The engine produced nine clusters; this agent explains what they mean clinically. The clearest illustration of the engine/agent split in the whole platform.

Disease Program¶

P1 · Tuberous Sclerosis Complex — "The whole factory, one child"¶

Label: MIXED — LIVE agents, REPRESENTATIVE discovery · Audience: everyone · Runtime: 12 min

The flagship. An infant with seizures and cardiac rhabdomyomas. Five disease-specific agents composing the horizontal engines: variant curator → trajectory modeller → therapeutics strategist → phenome mapper → TAND surveillance.

Weight. The need for this program came from paediatric clinicians. TSC1/TSC2 act as a brake on mTOR; lose the brake and growth runs unchecked.

Compression. Genome → variant curated → mTOR pathway → everolimus considered with its real tolerability profile → trial matched → TAND surveillance scheduled.

Honesty, inline and non-negotiable: - decision support for a qualified clinician, never diagnosis or prescribing; - gene-therapy correction of TSC1/TSC2 is preclinical — the open design bench, not a cure today; - pediatric caution at full force.

Hope. Open-source, one box, and the next child's afternoon is faster because of this one.

Public wording: "No one should wait years for a disease we could have understood in a day." Not the internal phrasing.

Honest summary¶

Label	Count	Which
LIVE	6	E2, E4, E5, E6, E8, + the 8 agents once Milvus is seeded
REPRESENTATIVE	9	E1, E3, E7 and the agents until Milvus is seeded
MIXED	2	P1, and E3 if the NIMs burst

The honest count improves in one step: execute GATED_SOFTWARE_BUILD_GUIDE.md, and E1, E3 and E7 become LIVE or BURST.

The strongest three to show today: E4 (imaging — proven, end-to-end, on this box), E8 (single-cell — small and entirely real), A3 (pharmacogenomics — 44 verified CPIC pairs).

The one to be most careful with: E3. It is the flagship, the most watched, and currently the least live.

Relationship to the D1–D7 portfolio¶

These are a different axis, not a replacement. docs/demos/index.md defines the established D1–D7 portfolio: seven patient-story demonstrations (D3 is the TSC flagship, D1 is secondary genomics → novel molecule, D2 is imaging → cardiac → pharmacogenomics, and so on). That portfolio is the coverage contract the honesty framework and the public site refer to.

This catalogue is keyed per subject — one demo for each of the 8 engines (E1–E8), 8 agents (A1–A8) and the TSC program (P1) — because the brief asked for one per capability.

Axis	Keys	Organising principle	Use it for
Portfolio (`docs/demos/index.md`)	D1–D7	one patient, many capabilities	proving the factory to a room
Catalogue (this document)	E1–E8, A1–A8, P1	one capability, shown properly	proving a single subject works

They compose: portfolio D3 (TSC flagship) is the same program as catalogue P1, shown from the patient's side rather than the capability's. Portfolio D1 draws on catalogue E1 → E2 → E3.

Do not renumber D1–D7. It is referenced by docs/honesty/maturity-matrix.md, the site, and the demo-foundation alignment. This catalogue was re-keyed to E/A/P precisely to avoid colliding with it.