Live site · Two-phone demo · How it works · Proof · Counsellor desk (demo passcode in the submission) · Programme view · Governance · العربية

Two Keys lets a couple learn what they could pass on to their children together, and nothing else. Their premarital carrier results are compared inside their own phones with authorised private set intersection: each partner learns only the genes they both carry. Neither learns any other gene the other carries, the families learn nothing, and the relay in between can't read a thing. The lab, which already holds both results, confirms the match and names only the shared genes. Each result is explained in plain Arabic or English at three reading levels, and a counsellor is one tap away. Relatives can be told, anonymously, that they may want a test.

Built for the Future Health Digital Innovation Challenge (Abu Dhabi), statement: Understanding and Reducing Inherited Disease Risk.

What's inside

Part	Where	What it does
Protocol	src/lib/psi.ts, src/lib/lab.ts	Authorised PSI (De Cristofaro and Tsudik) with session-bound lab tokens. Diffie–Hellman PSI on ristretto255 over 64-entry padded sets, with a batched Chaum–Pedersen proof on every lock. Role tags make severity-aware matches. A woman's X-linked carrier result reaches her as a personal result
Pairing	src/lib/pairing.ts, src/app/api/relay, src/app/api/lab	A link secret in the URL fragment gives the session id, an AES-GCM relay key and three check words. Relay slots are write-once and expire after 10 minutes. Signed identities and atomic seats; the lab recomputes and confirms the match
Phone app	/phone	Carrier card (couple-only by default; own results on request), pairing, couple result with pedigree icon arrays, three reading levels, Listen, teach-back, options, counsellor booking, explainer, family notes
Explainer	src/lib/ask.ts	Amazon Nova Pro on Bedrock. Keyword guards and a model intent check run first (partner probing, decide-for-me, crisis in English, Arabic, Gulf dialect and Arabizi). Answers are grounded in vetted facts; code checks reject invented numbers, wrong odds and directive language. Distress offers a same-day counsellor call; anyone can ask a person instead
Counsellor desk	/desk	Named counsellor accounts, consent-scoped cases, priority clocks (same day, pregnancy, routine), referral and closure, HMAC-keyed audit chain, withdrawal that erases shared data and seals notes
Family notes	/r/[id]	Letters sealed on the sender's phone. The server stores ciphertext; the relative alone decides whether the sender learns it was read or a test was booked. Senders can take a note back
Programme view	/programme	Comprehension, uptake, time to referral, cascade testing and equity by language and literacy. Weekly releases on an append-only ledger, exact discrete Laplace noise from a keyed generator (each person counted once ever; ε ≤ 3 per person per year), small-cell suppression; live data and a labelled synthetic month kept apart. It never counts marriages
Governance	/governance	Clause-by-clause mapping to the DoH Responsible AI Standard, risk tier, model card, data card, DPIA table, risk register, live monitoring counts
Evidence	/proof, scripts/	Unit and end-to-end tests. Pedigree simulator against every published UAE target, with a sensitivity run for unmodelled genes. Explainer evaluation: guards on and off on the same model, 60 development questions and a new frozen held-out set for each version (178 in v6, written by a model that is neither product nor judge), three language-model judges, majority and strict counts, McNemar and judge agreement; every run in data/eval-history.json

Run it

npm install
npm run dev -- -p 3410        # http://localhost:3410
npm test                      # protocol and risk-engine tests
BASE=http://localhost:3410 npx vitest run tests/pairing.int.test.ts   # two phones over HTTP
npm run calibrate             # data/calibration.json
npm run eval:ask              # data/eval-ask.json (needs AWS Bedrock credentials)
npx tsx --env-file=.env.scripts.local scripts/heldout.mts eval/heldout-vN.json <author model>   # writes a new set once, then refuses
HELDOUT=eval/heldout-vN.json npm run eval:ask              # commit the set and its SHA-256 first

Environment (.env.local):

Variable	Meaning
LAB_KEY	64 hex chars
AWS_REGION, AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY	Bedrock access
ASK_MODEL	Default apac.amazon.nova-pro-v1:0
STORE_TABLE	DynamoDB table; without it, memory is used
DESK_PASSCODE	Counsellor desk passcode

Data and rights

• No real person's data is used: every person, result and case is synthetic.
• Carrier frequencies come from published Emirati exomes (Jain et al., Genet Med Open 2025), cited, and from gnomAD v4.1 (CC0).
• Variant classes come from ClinVar. Condition explanations link to MedlinePlus Genetics.
• The full table is on /proof.

Limits

• The protocol has not been independently audited. In the single-lab setting it trusts the lab, which already holds both results; private matching matters most when partners are tested by different labs.
• Gene-level matching: where the exact pair of variants matters (non-classic CAH, alpha-globin, HbS/HbC, SMN1), the result says a counsellor will confirm.
• The demo runs in Sydney on synthetic data. Production needs UAE hosting, a DoH licence (Decree-Law 49/2023 Art. 28) and a National Genomic Database feed (Art. 26).

The details are on /proof.