Ourobion

An agentic research system for health evidence. Ourobion reads open-access scientific literature and turns it into relationships between health measures — each one traceable back to the paper it came from and to an independent review by a model from a different company. Two products consume that evidence: Biotope, the phone app, and Nao, the research workbench.

Biotope is a One Health personal ecological health monitor for the ASEAN market. It connects daily observations about a person's body and behaviour with environmental context, then presents descriptive patterns without making diagnostic claims. One Health is the principle that the health of people, animals, and the environment is interconnected; Biotope applies the part an individual can observe and act on — personal physiology, daily behaviour, and environmental exposure — and does not claim that every One Health domain is implemented. Environmental collection is a placeholder in the current build.

This README is the human front door: where to go, what the two products do, and how to run them from source. AI coding agents start with AGENTS.md.

Where to go

You want to…	Start here
Judge the Launchpad 2026 submission	`docs/hackathon/the_launchpad_challenge/submission/` — write-up, appendix, references
Install Biotope on Android	Install Biotope on Android below — sideloadable demo APK, no toolchain needed
Open Nao without installing anything	nao.ourobion.com — live; sign in with the shared test account
Sign in and look around	Shared test account below — `[email protected]` / `test123`, one credential for both apps
See the project at a glance	www.ourobion.com — the showcase site, or `apps/site/` for its source
Take a guided reviewer route through the repository	`docs/repository-guide.md`
Understand the product, origin, and two-system design	`docs/project-overview.md`
Review what is actually implemented and measured	`docs/implemented/system-truth.md`
Read the research-model results (Zebra, Viceroy)	Research models below, then `model-training/evidence/publication-results/`
See what we got wrong along the way	`docs/development/what-we-got-wrong.md`
Build and run Biotope locally (not needed to review it — use the APK)	`apps/biotope/README.md`
Build and run Nao locally (not needed to review it — use the hosted app)	`apps/nao/README.md`
Understand the full insight-engine architecture	`docs/implemented/shared/insight-engine-architecture.md`
Navigate all active documentation	`docs/INDEX.md`
Understand the engineering practice	`docs/engineering-practice.md`
Work as an AI coding agent	`AGENTS.md`
Review third-party credits and licences	`ATTRIBUTION.md`

Launchpad 2026 submission

Document	What it holds
`writeup.txt`	The 1,000-word write-up, five pillars, plain text as the portal requires
`project-summary.md`	A short prose overview — what Ourobion is, why it is built this way, and what it does not claim
`appendix.md`	Evidence table, the prebuild-versus-delta commit boundary, and what is explicitly not claimed
`references.md`	External works cited by the write-up

Prior work versus the challenge delta is stated in the appendix and checkable in commit dates: 115 commits predate 3 July 2026, and 737 at revision 7b5a064 are the delta. The prior-work figure is permanent; the delta keeps growing, so the appendix pins the exact command and revision.

Two products, different users

Biotope — the personal mobile app

Biotope is the Flutter app used by the person whose health is being observed. It keeps daily logging brief, records signals such as digestion, hydration, and wellbeing, and presents trends and descriptive insight cards in non-diagnostic language.

The insights are important because logging alone only returns a person's data to them. Ourobion is designed to connect observations across time, show when a pattern has enough support to be worth surfacing, and say plainly when it does not.

Nao — the research and operations interface

Nao is the Next.js operator dashboard used by the people maintaining Ourobion's evidence layer. It exposes the paper corpus, relationship claims, independent verification, ingestion gaps, pipeline controls, and model status. It is not the consumer health app.

Nao makes the research process inspectable: operators can see which evidence produced a relationship, how it was reviewed, and why it is accepted, held, or rejected before it is allowed to influence the product. Nao is intended for authorised operators, not as a general interface for browsing private health histories; its production-grade role and row-level-security boundary remains pending verification.

How they connect

scientific literature
        │
        ▼
paper ingestion → relationship synthesis → independent verification
        │                                      │
        └──────────── inspected through Nao ───┘
                                               │
                                               ▼
                                   governed evidence and rules
                                               │
daily observations → personal baselines ───────┤
                                               ▼
                              descriptive insights in Biotope

The brain is the research and reasoning layer behind this flow, not a third user-facing product. One model proposes relationships from scientific literature; a verifier from a different vendor family checks them against independently retrieved evidence. The resulting graph is graded and rebuildable rather than treated as unquestionable truth.

Not every verified research relationship is automatically serving a Biotope card today. The current measured implementation and explicit gaps are recorded in docs/implemented/system-truth.md.

Brain design: brain-synthesis-design.md · brain-ingestion-design.md · shared/brain/.

Research models

Two small research checkpoints were trained during the challenge. Neither serves the product, and that is a deliberate decision rather than an unfinished one — evidence is not serving permission. Both were trained on local Apple Silicon after the requested GPU container did not arrive.

Model	What it classifies	Result	Verdict
Zebra v1	whether evidence supports a claim	macro-F1 0.599 ± 0.008 against a pre-registered bar of 0.70	Failed its own bar. The code refuses to promote it.
Viceroy v0	causal versus correlational wording	macro-F1 0.866 against a keyword baseline of 0.507	Beat its baseline, on one frozen holdout — not completed cross-validation.

Three further models — Giraffe, Salmon and Leafcutter — remain planned and untrained. Both checkpoints carry validated=false, serving_ready=false, public_weights_cleared=false. Weights are never committed; only manifests, hashes and evaluation artifacts are tracked here.

To read	Go to
The canonical reports, aggregate metrics, provenance hashes	`model-training/evidence/publication-results/`
Why the models exist and why they are not wired in	`research-models.md` — Part 2
The caveat we most want read — same-paper leakage is unsolved, so Viceroy's number carries an unquantified optimistic bias	`LEAKAGE.md`
What we chose to train, what we did not, and why	`model-roster.md`
Approvals still gating any real training run	`human-gates.md`
The training/evaluation workspace itself	`model-training/README.md`

Both models were also compared against Claude Haiku 4.5 on 96 real papers. That comparison has no adjudicated ground truth, so it measures disagreement, not accuracy.

Install Biotope on Android

You need an Android phone or tablet running Android 8.0 or newer. An APK is an Android package — it does not run on Windows, macOS or Linux. There is no desktop or web build of Biotope, and iOS is out of scope (it needs a Mac and a paid Apple Developer account).

The quickest route is to open this link on the phone itself:

ourobion-biotope-demo.apk

Android will ask the browser or file manager for permission to Install unknown apps before it opens the download. No Flutter, Android Studio, or local backend is required — the build is universal (every CPU architecture in one file) and connects to Ourobion's hosted demo Supabase project. Sign in with the shared test account below.

Optional integrity check. Only relevant if you download to a computer first and copy the file across. The release asset was built from c9ea97b:

SHA-256: 861824d1ecd6be50faf10f4aa21b10fe4a5e5ecae358ac731695894bd52cbd29

# run on the computer holding the downloaded file, not on the phone
shasum -a 256 ourobion-biotope-demo.apk              # macOS / Linux
Get-FileHash ourobion-biotope-demo.apk -Algorithm SHA256   # Windows PowerShell

This hackathon APK is debug-signed for sideloading and is not a Play Store artifact. Demo APK updates must be built on the same Windows host; if Android reports an incompatible signature, uninstall the older demo first (which deletes its local app data) and then install the new APK. The source-build and release process is in apps/biotope/README.md.

Open Nao in a browser

Nao is live at nao.ourobion.com — nothing to install. Sign in with the shared test account below. It is the fastest way to see the brain: the paper corpus, every drafted relationship, the verification verdict behind each one, and the loader.

Shared test account

Use the following account for the Android APK and for the hosted Nao:

Email:    [email protected]
Password: test123

In Nao, the account has viewer authority. It can inspect the read-only research and evidence surfaces but does not grant curator or administrator operations.
In Biotope, the same account is also view-only. It opens a preseeded demonstration health profile so reviewers can inspect trends and existing insight cards without first entering days of observations. The seeded data is demonstration data, not a real person's health record.
A Biotope user may instead register a separate account and build their own private profile.

This is a public shared account. Do not enter personal or sensitive information into it. A fully local Supabase stack has a separate authentication database, so the account is available only when the app points to the configured demo environment unless it is provisioned locally as well.

Building from source

Nothing here has to be built to review it — install the APK, or open the hosted Nao. If you do want to build, configure, run or test either product locally, those instructions live with the product they belong to, which is the authority for prerequisites, environment variables, troubleshooting, verification and deployment:

Product	Local build, run and test
Biotope (Flutter)	`apps/biotope/README.md`
Nao (Next.js)	`apps/nao/README.md`

Repository-wide documentation and context checks are the exception, since they belong to no single app and run from the root:

npm run context:check

Repository guide

For a map of every top-level directory and a guided reviewer route, see docs/repository-guide.md. The shortcuts are in Where to go at the top of this file.

Team and contact

Member	Role	Responsibility
Jayden	Project Lead & Systems Architect	Leads product direction, system architecture, technical strategy, project planning, and final decision-making.
Alton	Product Design & Submission Lead	Leads UI/UX design and prepares the public-facing product experience, submission materials, and presentation deliverables.
Janson	Development Enablement & Technical Support	Supports implementation, provides technical assistance, and provisions access to development tooling, AI services, and project accounts.

Project contact: [email protected]
Technical questions and reproducible defects: use the repository's GitHub Issues.
Want a populated account? The shared demo account [email protected] is view-only, so nothing one visitor does changes what the next one sees. If you would prefer a separate account pre-loaded with seeded health data — enough history for baselines, trends and insight cards to appear straight away, instead of logging for weeks — email [email protected] and we will set one up. That data is seeded and marked as simulated in the database, exactly as the shared demo account's is.

Third-party services, models, datasets, frameworks, fonts, and generated assets are credited in ATTRIBUTION.md. Team membership is recorded here rather than mixed into third-party attribution.

Ourobion is non-diagnostic by construction. It describes patterns in the available data; it does not provide medical diagnosis or replace professional care.

Reviewed and re-signed by Jayden

2 August 2026