Ir al contenido

24b — Open Glucose Stream draft 3: a clean-sheet standard

Esta página aún no está disponible en tu idioma.

2026-10-19 · investigation report · follows 24 · spec: poc/spec/draft3/ · reference: poc/spec/reference/ogs/d3 · evidence: evidence/24b-draft3-e2e.log · reviewed in 24c

Updated after the independent review (24c). The review found two critical holes (a removed writer or a lost phone could keep writing into the old epoch; strangers got envelopes over Bluetooth) and several smaller ones. They are fixed in the spec, the vectors and the reference; the figures below are the current ones. 24c lists every change and what was rejected.

2026-11-18: the standalone repository gets C§9.5 Leaving: a leave that is hard to lose, kept a year, sent until two relays have it, final for its key. Rationale D17 and two reference tests come with it; the vectors didn’t change (pull request #1). It is what Sukrara built on draft 1 (docs/44 Appendix G).

The user, 2026-10-19: “Do a draft 3 that defines the spec without being biased to what Sukrara had implemented. It is fine if it includes breaking changes and deep mental-model changes. Make sure to check the state of the art again when making it, to not forget anything important.” The privacy rule stands: no public identity linked to a glucose stream.

Result
Spec set OGS draft 3 in poc/spec/draft3/, written in RFC 2119 style, with vendor-neutral constants. Documents: core, data model, Nostr transport, Bluetooth transport, security and privacy, conformance, rationale (requirements and decisions), and a migration note from draft 1.
Mental model Draft 1: an owner phone broadcasts with one read key per epoch. Draft 3: a set of records, each signed by its author. Members hold grants that are ranges of days, cut from a key tree. Relays see keys that change every day.
State of the art Re-checked against current NIPs and open PRs, Marmot (MLS over Nostr), the HL7 CGM IG 1.0.0, Tidepool, Nightscout API v3, Open mHealth / IEEE 1752, HealthKit, Health Connect, Bluetooth CGMS, the consent models of Dexcom Follow, LibreLinkUp and Nightscout, MLS / Signal / Megolm, and BLE platform limits. What changed the design: below.
Reference poc/spec/reference/ogs/d3, Python with standard-library crypto (the draft-1 primitives plus HMAC trees and secp256k1 key tweaking). MLS was not chosen, so no heavy library is needed. Parts: owner/administrator, member, BLE identifiers, bridges, vectors, live run.
Test vectors poc/spec/draft3/test-vectors/ogs-draft3.json, byte-for-byte reproducible:
• the tree: 2 epochs, 5 covers, day material, blinded authors and a blinded secret, d tags, a punctured cover;
• the invite in 3 text forms, 3 invite checks;
• 44 events, with the plaintext of every wrap;
• 3 administrator checks;
• a 42-step acceptance scenario with its final state;
• BLE UUIDs, both proofs, and frames at 3 MTUs;
• bridge examples (Nightscout, FHIR, Health Connect, Tidepool).
Tests 36 offline tests: the 17 from draft 1, 12 from this draft (cover exactness and minimality on 60 random ranges, bounded grants, puncturing, blinded secret ↔ blinded public key, invite forms, validation, padding, BLE UUIDs, vectors and determinism) and 7 from the review (removed writer, grant-only cut, ended ordering, history cap, read-only emergency invites, merge ties, BLE AUTH). All pass.
Live run uv run ogs d3 e2e against nos.lol, relay.primal.net and nostr.mom, throwaway keys and synthetic glucose, everything expiring within the hour. The flow:
1. an owner invites a reader F1 (one day back) and a writer F2 (today only);
2. both join by typed code and backfill exactly their windows;
3. a live update reaches both;
4. F2’s carbs are accepted by the owner and by F1;
5. the owner removes F1: rotation, F1 ended, F2 granted the new epoch;
6. new-epoch data reaches F2, and F1 can’t read it;
7. (review) the owner removes F2 too, and F2’s carbs signed in its old epoch are not recognised.
9/9 checks pass, in about 50 s (24c log).
Live run found bugs The first run found real protocol issues, now fixed in the spec and the reference. See Lessons from the live run.
Publishing Nothing was published. The package is ready: the site page, a standalone-repo README, and the steps. Below. No PR to nostr-protocol/nips.

Fifteen requirements, R1–R15, are listed in full in rationale.md §1:

  • R1 Actors: the wearer, or a parent as administrator; several wearer devices; family; carers who write and acknowledge; time-limited school staff; clinicians who need history only; emergency responders; watch apps; integrations.
  • R2 Privacy: no identity, no identifiers, unlinkable over time, readers hidden from each other.
  • R3 Informed, bounded consent.
  • R4 End-to-end security over untrusted transports.
  • R5 Offline and nearby.
  • R6 History, corrections, 90 days for clinicians.
  • R7 Multi-device on both sides.
  • R8 Revocation that also stops recognition.
  • R9 Emergency access.
  • R10 A full CGM data model; pump and AID left out.
  • R11 Units and time.
  • R12 Extensibility.
  • R13 Bridges.
  • R14 Implementable on phones, watches and in Python.
  • R15 No server.

The threat model is in ogs3-security S§1.

The full reasoning, with what was considered and why each option was rejected, is in rationale.md §2. In short:

  1. The signed record is the authority (D1).
    • Every record is a NIP-01 event signed by its author inside the encryption, on every copy.
    • Contributors publish directly. There is no echo and no re-signing, which removes draft 1’s hardest rule (P4 writer scope).
  2. A GGM day-key tree instead of one epoch key (D2).
    • A grant is the minimal cover of exactly the agreed days. That enforces the history start and the end of access with keys: a 3-day school-trip grant ends by itself.
    • Removal rotates the epoch.
    • Members puncture old days for forward secrecy.
    • A 30-day rotation gives post-compromise security.
    • Rejected: MLS/Marmot (ordered commits over lossy relays; a heavy stack; account-bound KeyPackages), hash ratchets (no end), per-recipient ciphertexts (a p tag per follower).
  3. Blinded per-day authors (D3).
    • Each publisher key is tweaked by a secret per day and lane, as in Tor v3 key blinding. Relays see unrelated keys every day.
    • Only the publisher can sign, and every member can compute the key to subscribe. So replacement and spam filtering keep working, and members can’t overwrite each other.
    • Rejected: Marmot’s one-time key per event.
  4. One addressable kind (31078) and no ephemeral events (D4).
    • The live channel is the current 15-minute block, republished on each reading.
    • Lanes (g, r, m) let a member subscribe to glucose, other records or membership separately.
  5. Days as the access unit (D5): hours would multiply the number of authors.
  6. Rotating inboxes (D6).
    • A fresh inbox per invite and per epoch; members choose their own.
    • I and member keys never appear outside encryption.
  7. Invites that say what they offer (D7).
    • The ogs: text has a version byte and optional relay hints, and works from any host.
    • The offer inside gives scopes, history days and duration.
    • The match code is 6 digits, typed for nearby joins.
  8. Entries keyed by (creator, id) (D8): a writer can’t even address someone else’s entry.
  9. The membership lists only contributors (D9): readers stay hidden from each other.
  10. Data model (D10).
    • Units: integer mg/dL, converted at 18.01559 (Nightscout/Tidepool/xDrip+), with the 18.0 users documented.
    • Readings: row flags for lo/hi/warmup/unreliable/calibrated/backfilled/interpolated, and a tz offset.
    • Records: sensor-session records with a lifecycle and still no serial; device status; settings (unit, target, zone); typed entries (insulin, carbs, bg, exercise, note).
    • Alerts: an open kind, a Tidepool-style prio, a fallback text, and acknowledgements between carers.
    • Open enumerations, and no v.
  11. Bluetooth with no constant identifier (D11).
    • Advertised UUIDs come from the day key and rotate every 15 minutes; GATT UUIDs change per day.
    • A constant pairing service exists only while pairing in person.
    • No bonding.
  12. Emergency access as a standing invite (D12): auto-approved, read-only, short, always notifying, OPTIONAL.
  13. Short relay retention (D13): 2 days for live blocks, 30 days for archives and entries.
  14. Publishing (D14): a standalone repository; no nips PR.
Draft 1 Draft 3
What it describes Sukrara 0.11’s wire format a protocol from requirements; no app yet
Authority the visible event author; the owner echoes writers’ entries the record signature inside the encryption; no echo
Encryption key one secp256k1 G_n per epoch, nsec handed out, NIP-44 by ECDH 32-byte epoch root → day-key tree → symmetric NIP-44 key per day
Access granularity all of an epoch any range of days: history start and end enforced
Time-limited access no (needs a rotation, with the owner online) yes, without any action
Forward secrecy none members puncture past days
Post-compromise security on rotation only scheduled rotation (30 d), plus on removal
Authors on relays D, I, writers: constant, linkable forever blinded per day and lane: no content link across days (traffic patterns still link them, S§2)
Opaque tags d/c/gk from a never-rotating sid d from the day key; nothing constant
Removed follower can still recognise the stream (knows sid, BLE m) can’t recognise anything after rotation
Requests addressed to I (a constant p tag) a per-invite and per-epoch inbox
Kinds 21078, 21079, 30078 (NIP-78: relays may refuse to serve), rumors 30500–30502 31078 only; 4078 and 4079 never published
Live an ephemeral event per reading plus hour chunks the current block republished
Invite https://sukrara.mauri.app/join#…, no offer, 4-digit code ogs:… with a version, a 96-bit fingerprint and hints, any host, an offer, a 6-digit code
Consent approve = everything, forever approve = the offer: scopes, history days, duration
Writers visible keys, echo, first-seen creator rules contributors under blinded keys; entries keyed by creator
Multi-device one owner device; each follower device a separate join several publisher devices; link for a member’s other devices
Delegated admin no manage scope (a second parent)
Emergency access no standing invites (OPTIONAL)
Readers visible to members writers listed only contributors listed
HI/LO, quality no row flags
Calibration, fingersticks no bg
Sensor ss/se/sw fields sensor session records with state and reason
Device status, settings no device, settings
Treatments 4 kinds in one shape typed entries; exercise added
Alerts closed set; fire/clear open kinds, priorities, fallback text, acks
Time zone none tz offset, zone name
Enumerations, versions closed; v per object open; caps; kind and labels are the version
Corrections first copy wins newest object wins
BLE discovery one constant service UUID: a public health inference day UUIDs, 15-minute advertised UUID, pairing UUID only while pairing; both sides prove the day key before any envelope
Bridges Nightscout, Tidepool (informative) Nightscout, FHIR CGM IG, Health Connect, HealthKit, Tidepool, Open mHealth, CGMS; FHIR summary LOINCs

The full list with sources is in rationale.md §3. These findings changed the design:

  • NIP-78 and registry-of-kinds: relays SHOULD serve 30078 only to its author → no 30078 anywhere.
  • NIP-DA (nips#2411): a symmetric key used directly as the NIP-44 conversation key → day keys, and no nsec of a group key.
  • Marmot: one-time signing keys per event, and pages of rules for out-of-order commits → blinded keys instead, and confirmation that MLS doesn’t fit lossy relays and one writer.
  • Signal sender keys, Megolm, NIP-171 epoch tickets: ratchets give a start but no end → the tree.
  • NIP-TPLD and NIP-101h: a p tag per recipient, and health kinds with plaintext tags → what not to do; those kinds are avoided.
  • CGMS, Tidepool and Nightscout:
    • HI/LO as status, not a magic number → flags;
    • Tidepool’s alert priority and acknowledged time → prio and ack;
    • backfilled and time-zone offsets everywhere → kept.
  • Conversion constants differ between projects (18.01559 / 18.0156 / 18.0) → one is fixed, and the others are documented.
  • HL7 CGM IG 1.0.0:
    • LOINC 99504-3 for readings, and summary LOINCs;
    • no treatments, trend or quality;
    • → the FHIR bridge, with summaries computed by members.
  • Health Connect has no insulin record → the bridge table says so.
  • Dexcom, LibreLinkUp and Nightscout roles: invite expiry, pause, a careportal write role, and nothing for history limits, time limits or emergencies → scopes, covers, standing invites.
  • iOS overflow-area advertising; no phone API for Encrypted Advertising Data → derived 128-bit UUIDs, and documented platform limits.
  • NIP-44 now allows over 64 KiB → still capped at 65 535 bytes for older libraries.
  • NIP-59 kind 21059 → not used. NIP-70 / NIP-77 / NIP-62 → OPTIONAL hooks.

The first run against public relays failed 2 of 7 checks, and every failure was a real protocol gap that the offline scenario hadn’t covered:

  1. Replayed joins re-admitted a removed member. Relays delivered the same join wrap again after a resubscription. The administrator approved it again, and the removed F1 got a grant for the new epoch.
    • Spec (C§8.6, C§9.4): deduplicate wraps by id. A join from an existing member never changes anything. A replayed join MUST NOT re-admit.
    • Members (C§10.2): ignore a grant older than the last ended.
  2. Same-second replacements were refused (“replaced: have newer event”). Two membership versions were published in one second → created_at MUST strictly increase per object; publishers bump it (C§7.1).
  3. Rekey storm. An epoch_end processed before the grant triggered a rekey. Duplicates multiplied it → members wait up to 2 minutes for the grant before asking (C§10.2).
  4. History scope worked as designed. F2, offered “today only”, correctly got nothing from before UTC midnight. The check in the test was wrong, not the protocol.

Estimated in agent-days, with the repo’s parity rules: Android and Mac in the same round. Today’s protocol code is about 4 900 lines on Android (core/Proto.kt, Derive.kt, Ble*.kt, owner/, follower/, wearer/, treatments/, alerts/ wire parts) and 2 600 on the Mac (NostrV2.swift, BLEV2Central.swift, Treatments.swift, the alert wire parts).

Work Android Mac
Crypto core: tree, covers, puncturing, day material, key tweaking. Needs secp256k1 tweak-add: rust-nostr’s bindings may not expose it; secp256k1-kmp (ACINQ) / secp256k1.swift do. 1.5 1.5
Records and envelopes: control messages, invite parser, match code; driven by the draft-3 vectors 1.5 1
Owner/administrator: membership, grants, invites with offers, rotation, epoch roots in the backup, emergency invites (OwnerEngine rewrite) 4 — (the Mac isn’t an owner)
Member/follower engine: acceptance pipeline, merge rules, author index and resubscription, tick 2.5 2.5
Data model: flags (HI/LO display in one graph everywhere: the graph parity group), sensor records, device, settings unit/target, typed entries, open alert kinds 2.5 2
Bluetooth: derived UUIDs, midnight re-registration, scan-filter refresh, pairing service 2 1.5
UI: offer picker (history, duration, scopes), member list with windows, emergency invite, ack button; strings en/es 3 1.5
Tests and specs: vectors in InviteTest/BleTest/new Ogs3VectorsTest, -v3selftest; replace PROTOCOL-v2.md/PROTOCOL-identity.md by a pointer to poc/spec/draft3; parity groups; FEATURES.md, QA.md 2 1.5
Total ≈ 19 ≈ 11.5
  • About 30 agent-days in all. With two agents per platform, that’s roughly three calendar weeks, plus device QA by the user.
  • No compatibility: every follower must be invited again after the update, and the release notes must say so.
  • The alert rules and their scenarios don’t change. Only the alert wire format does.
  • The Nightscout, Health Connect and xDrip+ bridges get the flags for free.
  • No app speaks draft 3. The proof is the reference against its own vectors and the live run. A second, independent implementation (Kotlin or Swift) is the next real test.
  • Bluetooth over the air was not exercised, only identifiers, proof and framing. The iOS background behaviour (B§3.3) is from Apple’s documentation, not measured.
  • The owner backup (C§13) has no vectors and no implementation.
  • Push for iOS followers (N§9), pump/AID records, cover traffic and kind registration remain open.
  • The CGMS bridge octet order is still unchecked (carried over from draft 1).
  • Some state-of-the-art facts were marked unverified by the research and are worded as such in the spec: the NIP-77 relay share, HealthKit’s exact molar mass, and the CGMS bit positions.

Nothing has been published, and no PR to nostr-protocol/nips will be made (the user’s decision). The steps that remain are the maintainer’s.

Material:

Licence: CC0-1.0, chosen by the user on 2026-10-19, for the text, the test vectors and the reference implementation. The trade-off that was weighed:

  • CC0 lets anyone, including closed apps and other standards bodies, copy it with no conditions, and gives up the right to require credit.
  • CC-BY-4.0 would require attribution, which is a little more friction for vendors and for inclusion in other specs.

The reference code is CC0 too in the standalone repository: nothing requires GPL there, because the author owns all of it and it imports no GPL code. In this repository the same files keep their GPL header, as the repo rule asks.

What is done (2026-10-19):

  1. The standalone repository is prepared locally, not pushed, in /Users/blue/Projects/sukrara/tmp/open-glucose-stream/ (git-ignored here), committed and tagged draft-3. It contains:
    • poc/spec/ from poc/spec/draft3, with links into this repository pinned to a commit;
    • test-vectors/;
    • reference/: the draft-3 package and the shared crypto, CC0 headers, its own crypto tests;
    • README.md from poc/spec/drafts/standalone-README.md;
    • LICENSE: the official CC0-1.0 legal code;
    • a CI workflow that runs the vector check and the tests.
  2. The website page is Open Glucose Stream (website/src/content/docs/open-glucose-stream.md, English and Spanish), linked from the home page.

Left for the maintainer:

  1. Create github.com/mauriciabad/open-glucose-stream, then push the local repository and its tag.
  2. Optionally, once a second implementation exists, register kinds 31078, 4078 and 4079 in nostr-protocol/registry-of-kinds. That is a registry entry, not a NIP.
  3. Send the outreach texts from a project account (never from stream keys), when the user decides.

Nothing was added to the app. This is a standalone investigation. Moving to draft 3 is a separate decision; the cost is above. Draft 1 stays the description of what Sukrara 0.11 speaks.

Path What
poc/spec/draft3/ README, core, data, Nostr, BLE, security, conformance, rationale, migration
poc/spec/draft3/test-vectors/ogs-draft3.json vectors (throwaway keys from public strings, synthetic data)
poc/spec/reference/ogs/d3/ core.py (tree, day material, blinding, records, envelopes), data.py, control.py, member.py, owner.py, ble.py, bridges.py, vectors.py, e2e.py
poc/spec/reference/tests/test_ogs3.py 12 offline tests
docs/evidence/24b-draft3-e2e.log the live run on public relays
poc/spec/drafts/site-page.md, standalone-README.md publishing drafts