Maintenance status (2026-09): passive. This repository is kept available as a reference implementation; CI runs on pushes and pull requests only, Dependabot security alerts remain enabled, and no scheduled jobs or hosted services consume ongoing resources. No active development is planned.
Nurse-first medical reasoning — trained and evaluated on NVIDIA's open-component stack.
Architecture status (2026-08). The training/eval program below ran entirely on self-hosted NVIDIA GPUs ("sovereign by construction") and its results stand as published. The live demo no longer runs on that stack: the dedicated GPU serving pod was decommissioned in June 2026 and demo inference is migrating to the Anthropic Claude API via the web BFF (branch
feat/claude-opus-migration). Queries to the demo are processed by a third-party AI service — the demo is not sovereign. The demo is degraded/unavailable while the migration completes. Any "no cloud LLM in the inference path" claim in historical documents in this repo describes the pre-June-2026 deployment only.
Headline (2026-05-07, paired-bootstrap N=600 per benchmark, gpt-4.1 grader, thinking=True):
| Benchmark | V0 (Nemotron-3-Nano-Omni-30B-Reasoning, FP8) |
|---|---|
| MedQA-USMLE | 83.50% |
| PubMedQA-L | 67.33% |
| MedXpertQA-Text | 33.00% |
| HealthBench-Hard | 12.52% |
These are the base-model-with-thinking-enabled numbers. The V2.5 reasoning-SFT we trained on top did NOT improve them — see PR #122 and findings/2026-05-05-v2.5-eval/DISPOSITION.md for the full A5 ablation (thinking=False vs thinking=True, 0/4 ship-rule criteria PASS in either polarity). The negative result + reproducibility audit is the rigor story; the strong V0 baseline is the deployment story.
Demo status: degraded during migration. The demo backend is mid-way through the move from the decommissioned self-hosted GPU service to the Anthropic Claude API and has been failing since 2026-06-26 pending the production API key configuration. Until the migration lands, treat the URL as a UI preview, not a working agent. The deployed page still shows an outdated "no third-party AI APIs called" banner from the pre-migration build; that claim is wrong for the Claude-backed architecture and is corrected in this repository — do not rely on it.
Public demo URL, no login, no account. Routes via Vercel edge rewrite from
the www.thegoatnote.com apex (owned by v0-goat-note-landing-page-3c)
to medomni.vercel.app/4UWHAt. The page is the Records OS dashboard:
- Three-column patient cockpit — left nav rail (Overview · Timeline · Labs · Meds · Conditions · Vitals · Imaging · Wearables · Visit notes · Care team · Genome · Sharing), main canvas, right rail.
- Ask your record command bar — voice in (Web Audio + Nemotron-Omni native ASR), voice out (Kokoro-82M TTS on H200 via Cloudflare quick-tunnel; browser
speechSynthesisfallback), image input (camera / file → multimodal Omni reasoning), text in. - Imaging gallery — three click-to-view DetailDrawer cards (X-ray · MRI · panoramic) backed by real CC0 / CC-BY reference films from Wikimedia Commons. Track D's FHIR
ImagingStudyingestion populates the same shape from a real EHR. - Five-tool agent —
pubmed_search,primekg_lookup,guideline_currency_check,clinical_calculate,get_patient_context(FHIR R4). Tool call provenance + verification badge under every assistant turn. - FHIR R4 Bundle export — Share button produces a valid Bundle with Patient + Conditions + Observations + MedicationRequests + AllergyIntolerances + DiagnosticReports + ImagingStudies, downloadable as JSON.
- Demo banner — "For evaluation only; do not enter PHI." The banner also states which backend serves inference; the pre-migration "no third-party AI APIs called" wording is retired (see the inference path note below).
Inference path (current architecture): the browser talks to the web BFF
(web/app/api/agent), which forwards the conversation to the model backend.
- Pre-June-2026 (decommissioned): Nemotron-3-Nano-Omni-30B-A3B-Reasoning (FP8) on a dedicated NVIDIA H100 via vllm; no cloud LLM API keys in the serving path. This is the deployment the published eval numbers describe.
- Current (migration in progress, branch
feat/claude-opus-migration): the BFF calls the Anthropic Claude API (Claude Opus). User queries — including any text, image, or audio submitted to the demo — are sent to Anthropic for processing. Do not enter PHI. The dedicated GPU service is decommissioned.
Architecture decision behind the patient slice: Pattern B (dual lookup) — the agent hits FHIR (Medplum, self-hosted) and PrimeKG independently and merges in-prompt. Measured p95 = 11 ms for the FHIR fetch across 12 patients, 60 samples in a dev environment (findings/2026-05-04-pattern-b-spike/RESULTS.md); the same file's production caveat measured 150-400 ms end-to-end.
The training / eval / methodology stack below is the proof-of-quality behind that demo. Read on if you care how the model got there.
For the full planned trajectory (V2.5 reasoning-SFT → V2.7 tool-call SFT → V3 GRPO → V3.5 DPO refusal → V_final HF release as Apache-2.0; eval gauntlet with MedAgentBench as the headline target), see the navigable index at findings/INDEX.md. 16+ pre-registered documents covering every stage; ~32-41 H200-hrs total.
Status (2026-08): V2.5 reasoning-SFT eval landed (PR #122) — A5 ablation closed with both arms FAILing the pre-registered ship rule; the durable result is the V0 baseline strength (numbers above) with enable_thinking=True canonical. The single-H100 serving pod that carried the demo from 2026-05-07 was decommissioned in June 2026; demo inference is migrating to the Anthropic Claude API (see the architecture status note at the top). The Karpathy-autoresearcher training loop (3-GPU fleet) was decommissioned 2026-05-07; future SFT runs wait for the next training-budget cycle (Issue #130).
V0 baseline with thinking=True is the medical reasoner. Pre-registered paired-bootstrap eval (PR #122, N=600 per benchmark, 4 benchmarks × 3 seeds, gpt-4.1 grader, sha256-verified manifest):
- MedQA-USMLE 83.50% — competitive with claimed-SOTA medical LLMs on the same harness
- HealthBench-Hard 12.52% — vs published OpenAI-paper baseline on the same rubric
- PubMedQA-L 67.33%, MedXpertQA-Text 33.00%
The V2.5 reasoning-SFT we trained on top (MedReason 32K + medical-o1-reasoning-SFT 25K + R1-distill-USMLE) did not add value — both thinking=False (1/4 PASS, deniable) and thinking=True (0/4 PASS, canonical) FAILed the pre-registered ship-rule criteria. The A5 ablation is documented in
findings/2026-05-05-v2.5-eval/DISPOSITION.md.The negative-result-with-rigor is the publication-grade story; the strong V0 baseline is the deployment story. Both are reproducible end-to-end via
findings/2026-05-05-v2.5-eval-thinking/REPRO.shon a fresh GPU.
Held-out mean 0.385 ± 0.000 across 6 chemoprevention fixtures, N=3 seeded trials, manifest byte-stable, score progression v0 0.273 → v1.0 0.385 (+41%). HPV +0.22, 5ARI +0.14, bisphosphonate +0.12, smoking +0.11, aspirin +0.08. Cross-family judge: Qwen2.5-7B-Instruct. Showcase artifact:
results/canonical-2026-05-01-hb-hard-n1000/CARD.md. Superseded by the 4-benchmark paired-bootstrap eval above; preserved here for historical audit.
This is the v0.1.0 public scaffold. The held-out mean is honest, the methodology is documented
in docs/methodology.md, and the
roadmap to ≥ 0.55 (open-source SOTA on HealthBench Hard) is in docs/SPEC.md §6.
A sovereign, multi-modal, persona-aware medical reasoning system that reads the same clinical case (drug + device + image + auscultation audio + family history + prior labs) and returns four register-shaped answers — one for the physician (full diagnostic depth + literature citations), one for the nurse (clinical depth + early-warning escalation cues + teaching scaffold), one for the family (caregiver register + plain-language analogies + when-to-call-911), one for the patient (FKGL ≤ 8 + shared-decision-making tone) — each with an auditable cited graph path beneath the answer.
Irreducible core feature: the persona-tagged grounded explanation. Same case, same evidence, four registers, one auditable subgraph, manifest-locked reproducibility. We are not aware of another open-source system that combines these on one stack with provenance.
The literature gap that justifies the wedge: only 6% of 67 medical-KG studies addressed nursing-specific applications (JMIR AI 2025). The four-persona stack with persona-tagged graph edges is genuinely under-served.
During the model-training phase (2026-04 to 2026-05-07) a 3-GPU fleet ran the data → train → evaluate flywheel below — each GPU with a distinct role. Once the V0 baseline was established and the V2.5 trajectory measured, the fleet had completed its purpose, and serving was deliberately consolidated to a single right-sized H100. This section is retired history; the live demo runs on one dedicated NVIDIA H100.
| Pod | Hardware | Role | What's running |
|---|---|---|---|
| catfish | NVIDIA B300 (Blackwell, 288 GB HBM3E) | Inference + agent surface | vllm-omni-b300 serving nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4 with native structured tool-calling (--enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser nemotron_v3). Powers medomni.vercel.app/agent (PubMed tool, visible reasoning). NVFP4 quantization is Blackwell-only. |
| lobster | NVIDIA H200 (Hopper, 143 GB HBM) | Training + sovereign judge | Path D Megatron-Bridge LoRA training (V1 shipped 2026-05-03 in 11.3 hr at 2.6 s/step — 12.4× faster than HF/PEFT eager). Then judge-qwen serving Qwen/Qwen2.5-7B-Instruct for sovereign corpus filtering. |
| narwhal | NVIDIA H200 (Hopper, 143 GB HBM) | Data factory | vllm serving nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 + factory_loop.py generating quality-judged medical reasoning chains for the next training round. ~21k+ items / day. |
The autonomous loop (Karpathy autoresearcher pattern, applied to a vertical):
┌─────────────────────────────────────────────────────────────────┐
│ │
▼ │
[narwhal + catfish: factory_loop.py generates clinical reasoning items] │
│ │
│ 47k+ raw items / week │
▼ │
[laptop: scripts/judge_reasoning_sovereign.py — gpt-4.1 + Qwen ensemble] │
│ │
│ ~45% pass rate, structured rejection reasons, $1.20 / 1k │
▼ │
[curated jsonl → lobster: scripts/deploy_v2_corpus.sh] │
│ │
▼ │
[lobster: Path D Megatron-Bridge LoRA on Nemotron-3-Nano-30B (≤5 epochs)] │
│ │
│ ~28 hr wall, $120, kill-switch on val PPL > 10 │
▼ │
[lobster: HF PEFT export → catfish: vllm --enable-lora] │
│ │
▼ │
[catfish: serve V_n adapter → /api/agent renders new behavior live] │
│ │
│ (eval loop: HealthBench Hard N=1000 paired V_{n-1} vs V_n, │
│ gpt-4.1 graded, paired-bootstrap CI, ship rule applied │
│ literally per PREREG.yaml) │
│ │
└─────────────────────────────────────────────────────────────────┘
While it ran, the loop operated continuously, with PRs as the human-inspected checkpoints between autonomous work. It was decommissioned alongside the 2026-05-07 consolidation.
Sovereignty floor (historical): every box above ran on hardware under our control, and that property held for the live demo through June 2026 while it served Nemotron from the dedicated H100. It does NOT hold for the current demo architecture, which routes inference to the Anthropic Claude API (see the "Cloud API usage" section below).
Target architecture. The live demo on a single H100 implements a subset of this pipeline today; the remaining stages are the build plan.
[user query, modality bundle (text + optional image / audio / video)]
│
1. NemoGuard INPUT RAIL (NeMo Guardrails 0.21 Colang 2.0)
▼
2. cuVS DENSE RECALL (top-50) ◄┐
│ embed: llama-nemotron-embed-1b-v2
│ IVF-PQ over node descriptions + corpus chunks
│ │ parallel
3. BM25 SPARSE RECALL (top-50) │
▼ │
4. RRF FUSION (top-50 → top-25) ◄────┘
▼
5. CROSS-ENCODER RERANK (top-25 → top-8)
│ llama-3.2-nv-rerankqa-1b-v2 (TensorRT-LLM-FP8 on H100)
▼
6. nx-cugraph SUBGRAPH SLICE
│ FACTUAL: PrimeKG (Harvard, 129K nodes / 4M edges, MIT license)
│ REGISTER: persona-tagged graph (NIPDS, evidence-currency, pedagogical-intent)
▼
7. SUBGRAPH SERIALIZE (5–15K tokens)
▼
8. OMNI INFERENCE (B300, NVFP4)
│ nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4
│ 23 Mamba2 + 23 MoE + 6 GQA, 256K context, multimodal in / text out
▼
9. CONSTRAINED DECODING + GROUNDING CITE RAIL
│ guided_json against citation schema
│ each cited passage: cosine ≥ 0.8 to retrieved node description
▼
10. NemoGuard OUTPUT RAIL (Nemotron-Content-Safety-Reasoning-4B)
▼
[persona-shaped grounded answer + cited subgraph]
Bill of materials (every layer NVIDIA-shipped open):
| Layer | Component | Version |
|---|---|---|
| Driver / CUDA | NVIDIA driver 580+ / CUDA 13.2 Update 1 (April 12, 2026) | 13.2 |
| Math | cuDNN 9.x · cuBLAS 13.2 · NCCL 2.25.x | bundled |
| Kernels | CUTLASS 4.x + CuTeDSL · FlashAttention-4 (Mar 5, 2026) | latest |
| Compute libs | RAPIDS 26.04 (cuDF, cuML, cuVS, cuGraph, nx-cugraph) | 26.04 |
| Vector | cuVS (CAGRA + IVF-PQ + IVF-FLAT) | 26.04 |
| Graph engine | nx-cugraph | 26.04 |
| Graph data | PrimeKG (Harvard mims-harvard, MIT) + persona-tagged graph (this repo) | v2.0 |
| Frame | NeMo Framework | 2.7.0 |
| Curate | NeMo Curator 1.1.0 (Feb 23, 2026) | 1.1.0 |
| Guard framework | NeMo Guardrails 0.21.0 + Colang 2.0 | 0.21.0 |
| Guard models | NemoGuard JailbreakDetect + Nemotron-Content-Safety-Reasoning-4B | revision-pinned |
| Engine | TensorRT-LLM + vLLM 0.20.0 | 0.17+ |
| Embed / rerank | llama-nemotron-embed-1b-v2 + llama-3.2-nv-rerankqa-1b-v2 | revision-pinned |
| Brain | nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4 | revision-pinned |
| Frontend serving | Triton Inference Server with OpenAI-compat | 26.03 |
| Judge (cross-family, sovereign) | Qwen/Qwen2.5-7B-Instruct | FP8 TRT-LLM engine |
The only non-NVIDIA weights in the target pipeline are Qwen2.5-7B-Instruct, which acts as the cross-family judge per NVIDIA's own published Nemotron-3 reproducibility recipe. Note this describes the self-hosted target architecture, not every code path in the repo: the ship-rule and HealthBench graders call gpt-4.1 / Claude, CI review calls the Claude API, and the current demo routes inference to the Claude API (see "Cloud API usage" below).
For the full architecture spec, see findings/research/2026-04-29-medomni-v1-northstar/SPEC.md.
# 1. Clone + create venv
git clone https://github.com/GOATnote-Inc/medomni.git
cd medomni
# 2. Install (uv recommended; pip + venv supported)
uv sync --extra dev # OR: python3 -m venv .venv && .venv/bin/pip install -e ".[dev]"
# 3. Configure sovereignty-only env
cp .env.example .env
# Edit .env to set HF_TOKEN, BREV_PEM_PATH (no cloud LLM keys — see CLAUDE.md §2)
# 4. Lint + test (laptop, no GPU required)
make lint
make test # pytest -m "not integration and not gpu"
# 5. Verify the headline manifest is byte-deterministic
make manifest-verify ARTIFACT=results/ci-medomni-heldout-consolidated-20260429-173557/heldout.json
# 6. (Optional, requires Brev pod access) full bench
make ci-medomni # captures snapshot + sovereign_bench + write_card + verdictThe make manifest-verify step in particular is the central reproducibility claim: it re-emits
the 9-layer manifest from the same artifact twice and asserts byte-equal output. If you can run
that and see "manifest emitter is byte-deterministic," you've reproduced the v0.1.0 result.
Every demo run emits a MANIFEST.yaml with byte-identical fingerprint covering 9 layers:
- Container image digests (vLLM, TRT-LLM, NeMo Guardrails, NeMo Curator)
- Weight SHAs (Omni NVFP4, NemoGuard models, embed-1b-v2, rerank-1b-v2, Qwen2.5-7B)
- Corpus SHAs (OpenEM 370, PubMed-OA shard, fixture set)
- Config files (Colang rails, persona prompts, retrieval params)
- Random seeds (decoding seed, retrieval tie-break seed)
- Hardware-foot-gun flags (
--no-async-scheduling,--max-model-len,--kv-cache-dtype fp8) - Benchmark fixtures (chemoprevention 6 + tamoxifen 1 = 7 cases)
- Judge model digest (Qwen2.5-7B-Instruct FP8 TRT-LLM engine)
- Git SHA of harness + SPEC.md SHA
The 9-layer manifest is the rigor that OE / Hippocratic / ChatGPT-for-Clinicians do not demo
because their stack isn't built for it. See docs/SPEC.md §5.6
and scripts/emit_manifest.py.
medomni/
├── README.md # this file (results-first → architecture → quickstart)
├── LICENSE # Apache-2.0
├── CONTRIBUTING.md # style + tests + manifest discipline
├── SECURITY.md # disclosure email + scope + durable doctrine
├── CODE_OF_CONDUCT.md # Contributor Covenant 2.1
├── CHANGELOG.md # Keep-a-Changelog
├── CLAUDE.md # AI-agent operating charter (isolation rules, current architecture)
├── CITATION.cff # citation metadata
├── pyproject.toml # PEP 621 + uv-compatible
├── Makefile # health, manifest, ci-medomni, demo-pre-flight
├── DEMO.md # 12-minute on-stage demo runbook
├── .env.example # env template (HF_TOKEN, NGC_API_KEY, BREV_PEM_PATH, tunnel URLs)
├── .pre-commit-config.yaml # ruff + detect-secrets + custom hooks
├── .secrets.baseline # detect-secrets baseline (no secrets in baseline; verified)
├── .github/ # workflows, ISSUE_TEMPLATE, PR template, CODEOWNERS
├── corpus/ # clinical fixtures + held-out + golden cases + medical guidelines
├── data/ # seed_kg/ (illustrative graph), embeddings (gitignored)
├── configs/ # guardrails rails, persona prompts, retrieval params
├── scripts/ # bench, retrieval, manifest, serve_*, _runpod_ssh.sh
├── tests/ # unit (laptop), integration / gpu (Brev pod required)
├── mla/ # medical-LLM agent harness (legacy from prism42 lift)
├── findings/ # research provenance — north-star SPEC + POSITIONING + DEMO + PITCH
│ └── research/
│ ├── 2026-04-29-medomni-v0/ # methodology audit
│ └── 2026-04-29-medomni-v1-northstar/ # the SPEC + the moat
├── results/ # per-run CARDs + MANIFEST.yaml (showcase entries committed)
└── reproducibility/ # manifest emitter design + captured templates
Earlier revisions of this README claimed "no cloud LLM API keys exist in any code path." That was true of the self-hosted serving path while it ran; it was never true of the whole repository, and it is not true of the current demo. The actual ledger:
| Path | Backend | Why |
|---|---|---|
| Demo inference (current) | Anthropic Claude API via the web BFF | The dedicated GPU pod was decommissioned in June 2026; migration branch feat/claude-opus-migration |
| Demo inference (pre-June-2026, decommissioned) | Self-hosted Nemotron FP8 on H100 via vllm | The published "sovereign" deployment |
| Ship-rule / HealthBench rubric grading | OpenAI gpt-4.1 (laptop-side), claude-opus-4-7 (HealthBench runner) |
Pre-registered eval protocol graders |
| CI PR review + clinical-content gate | Claude API (ANTHROPIC_API_KEY_PR_REVIEW repo secret) |
.github/workflows/agent-pr-review.yml, clinical-skill-review.yml |
| Sovereign cross-check judge | Qwen2.5-7B-Instruct, self-hosted | Cross-family judge of the eval recipe |
"Sovereign" in the historical documents under findings/ describes the
pre-June-2026 self-hosted inference deployment only.
The 2026-04-29 HF_TOKEN PTY-echo incident and its durable mitigation (the _runpod_ssh.sh secret-grep
guard) are summarized in SECURITY.md; the full postmortem lives in the private
research repo. Operational maturity is shipping the postmortem alongside the fix.
- v0.1.0 (this release) — initial public scaffold; held-out 0.385; reproducibility manifest shipped; 9-layer byte-deterministic emitter.
- v0.2 (Phase 2.4 — in flight) — PrimeKG (Harvard mims-harvard, 129K nodes / 4M edges) wired into Stage 6 nx-cugraph subgraph slice; projected lift +0.05–0.10 closes the SPEC §6 ≥ 0.45 gate.
- v0.3 (Phase 1.6 / 1.7) — broader verbatim PMC corpus + N=30 fixture scale.
- v0.4 (Phase 2.3) — ensemble cross-judge agreement scoring.
- v1.0 — pre-launch SOTA gate: HealthBench Hard ≥ 0.55, MedAgentBench ≥ 0.70, MedQA-USMLE ≥ 0.85. If we miss the gate, we don't ship — accuracy degradation = revenue collapse.
Image-aware MedOmni (training imminent — next 24h) — V_image-aware extends the Nemotron-Omni
reasoning surface with native radiology comprehension via continued pretraining on the
GOATnote-Inc/medimage-corpus registry: 134
open-source medical imaging datasets across CT, X-ray, MRI, ultrasound, and image-text paired
(VLM) collections, with download dispatchers and format converters wired for H200 ingest.
Manifests cover ~7.4 PB cataloged and ~700 TB pullable in the open + registration tiers; the
demo's Imaging rail (X-ray · MRI · panoramic) is the first surface that benefits.
Audit ledger — GOATnote-Inc/receipts — append-only
intent-vs-execution attestation layer that sits beside MedOmni's existing Medplum AuditEvent +
S3 Object Lock pipeline and turns it into audit-grade evidence. Every 5-tool /
4-persona inference becomes a clinical_drift_finding row plus a Merkle hash-chain attestation;
the κ-graded dual-judge agreement engine (claude-opus-4-7 + gpt-5.4-2026-03-05) scores each
output against the encounter contract; the FHIRConnector then writes an AttestationExtension
(https://goatnote.dev/receipts/attestation) back onto the committed Composition. PHI-aware
emitter has no Slack handle by design — patient text never leaves the FHIR + Markdown surface.
The clinical vertical (receipts-clin) audits MedOmni inferences end-to-end; the engineering
vertical (receipts-eng) audits the Linear/GitHub/Slack/Granola scope-drift on the training
- eval workflows themselves. Shared substrate: 9-table eng schema + 3-table clinical schema + Merkle log + CEIS L0/L1/L2 + Cohen's κ + Wilson CI + pass^k regression gate + 4 regulatory export formats (FHIR R4 Bundle / SARIF v2.1 / Markdown / CSV). Hermetic test suite: 441 tests, zero live API calls.
Detailed phasing and acceptance criteria: SPEC.md §8.
See CONTRIBUTING.md. The manifest discipline is the load-bearing rule: every PR that touches the inference path must attach a fresh manifest hash from a re-bench.
Bug reports + feature requests + reproducibility issues all have templates in .github/ISSUE_TEMPLATE/.
Every image/audio fixture under corpus/demo-fixtures/ ships a per-file
fixture-*-LICENSE.md with source URL, author, and license. They are NOT
all CC-BY-4.0; the share-alike items keep their original terms:
| Fixture | License |
|---|---|
auscultation/fixture-aud-001-crackles-pneumonia.{wav,mp3} |
CC BY-SA 3.0 / GFDL 1.2+ (dual) — share-alike; cannot be relicensed |
pill/fixture-pill-001-warfarin-tablets.png |
CC BY-SA 3.0 / GFDL 1.2+ (dual) — share-alike; cannot be relicensed |
cxr/fixture-cxr-001-rml-pneumonia.png |
CC0 1.0 (public domain) |
ecg/fixture-ecg-001-stemi-12lead.png |
CC BY 4.0 |
If you redistribute the CC BY-SA fixtures, your derivative of those files must carry CC BY-SA 3.0 terms regardless of this repo's Apache-2.0 / CC-BY-4.0 defaults.
@software{medomni_v0_1_2026,
author = {Dent, Brandon},
title = {{MedOmni: Sovereign nurse-first medical-LLM stack on NVIDIA's open-component stack}},
version = {0.1.0},
date = {2026-04-29},
url = {https://github.com/GOATnote-Inc/medomni},
organization = {GOATnote Inc.}
}
Full machine-readable citation metadata: CITATION.cff.
Built on NVIDIA's open-component stack — every layer carries an NVIDIA badge: CUDA 13.2, cuVS, nx-cugraph, NeMo Guardrails, NeMo Curator, TensorRT-LLM, NemoGuard, Llama-Nemotron-Embed/Rerank, Nemotron-3-Nano-Omni. Architecture derives from the NVIDIA RAG Blueprint reference, NVIDIA Nemotron-3 reproducibility recipe, and Stanford CRFM HELM reproducibility-first patterns.
Held-out clinical fixtures sourced from USPSTF, NCCN, ACOG, IBIS-II, NSABP P-1, and the Nurse.org RN-perspective corpus.
Provenance: derived from github.com/GOATnote-Inc/prism42 (private fork lineage); squash-imported
with no shared commit history. The medical-LLM eval harness was lifted from public prism42 with
zero prod-surface entanglement.