Sentlog

A universal, device-first evidence and presence SDK.

Captures a short voice interaction (optionally augmented with motion and facial-affect signals), produces a multi-dimensional presence assessment, and emits the result as a signed, W3C-Verifiable-Credentials envelope. Runs entirely on-device. Inert by default. MIT.

⚠️ Status: pre-v1.0 — NOT production ready. The architecture is in place and CI is green, but the SDK has not been through its ≥72 h on-device soak, has not been cut as v1.0, and the face-affect MLP is a placeholder pending the labelled corpus (gated to months 7–12 per docs/rollout-plan.md). The public API surface and envelope wire format may shift before v1.0; the deprecation cycle in docs/deprecation-policy.md only kicks in from v1.0 onward. Use on main at your own risk.

What it is

Sentlog is a mobile SDK for iOS and Android with first-class bridges for React Native and Capacitor. It is not a trust SDK, not an authentication SDK, not a fraud SDK — it is the evidence-producing layer on which integrators build those things.

Four design pillars:

• Universal — native iOS, native Android, Capacitor-wrapped PWAs, React Native apps. One shared C++ core; thin per-platform and per-framework bridges. @sentlog/web (browser WASM) is a defined post-MVP deliverable.
• Device-first — capture and inference happen on the device. Defaults are inert: nothing transmits, nothing runs in the background, no cloud is implied by integration.
• Evidence — every result is wrapped in a W3C Verifiable Credentials 2.0 envelope with Ed25519 or ECDSA-P256 proofs (the latter via Secure Enclave / Android KeyStore when available), model versions, timestamps, and a homegrown MIT perceptual audio fingerprint (sentlog-pcp-v1). Optional RFC 3161 TSA timestamps when an integrator configures a TSA endpoint.
• Presence — the subject of the evidence. Liveness, engagement, direction, presence proper, and optionally facial affect — each exposed separately so integrators compose their own verdicts.

What the SDK ships, in one call:

val result: SentlogResult = Sentlog.evaluate()
// → transcript, liveness, engagement, direction, presence, affect,
//   confidence, fingerprint, model versions, diagnostics, evidence_grade

What leaves the device: nothing, unless the integrator configures a forwarder. Forwarders are pluggable sinks (HTTP, webhook, local queue, callback) with per-sink projections (full result, scores only, transcript only, fingerprint, clip, custom). The SDK has no opinion on destinations.

Status

Pre-Phase-0. The repository currently contains only the planning documents. No source code, build system, models, or sample apps exist yet. Phase 0 (foundations) and the 16-week MVP to v1.0 are documented in docs/rollout-plan.md.

Documentation

Three documents cover what Sentlog is, how it works, and how it ships:

• docs/whitepaper.md — vision, foundational principles, positioning, multi-modal signal rationale, data strategy. Start here if you're evaluating Sentlog as a building block.
• docs/technical-specification.md — architectural decisions, repository layout, the full API surface, composable forwarding, security model, versioning policy. Start here if you're integrating or contributing.
• docs/rollout-plan.md — 16-week MVP phases, Android benchmark gate, testing and observability, risk register, open items. Start here if you're executing.

CLAUDE.md covers conventions for AI-coded contributions.

Planned repository structure

Full layout in docs/technical-specification.md §2. Summary:

sentlog/
├── LICENSE                # MIT
├── README.md              # this file
├── SECURITY.md            # vulnerability disclosure policy
├── CLAUDE.md              # AI-coded contribution guidelines
├── shared/                # C++17 core — narrow C ABI, heavily tested
├── android/               # AAR (Kotlin + JNI)
├── ios/                   # XCFramework (Swift + C interop)
├── rn/                    # @sentlog/react-native
├── capacitor/             # @sentlog/capacitor
├── models/                # training + quantisation pipelines (not bundled)
├── tools/
│   ├── benchmark-harness/
│   ├── reference-consumer/   # HTTP server for HttpSink testing
│   ├── evidence-verifier/    # single-file Python verifier
│   ├── sample-capacitor/     # Svelte/Ionic + Capacitor sample app
│   └── sample-react-native/  # React Native sample app
└── docs/
    ├── whitepaper.md
    ├── technical-specification.md
    └── rollout-plan.md

Quick comparison

Concern	Sentlog
Runtime	On-device only (iOS, Android, Capacitor, React Native; WASM post-MVP)
Output	W3C Verifiable Credentials 2.0 envelope (signed, durable)
Signing	Ed25519 (software) / ECDSA-P256 (Secure Enclave / KeyStore)
Timestamping	Device timestamp; optional RFC 3161 TSA
Networking	None by default. Pluggable forwarders, named by the integrator.
State	No user model, no sessions, no accounts. Opaque session_id passthrough.
Personal data	Voice and video never leave the device unless explicitly configured. Fingerprint forwarding ships ~120 bytes of non-invertible perceptual hash.
Licence	MIT
Foundational dependencies	whisper.cpp (MIT), Silero VAD (MIT), AASIST-L (MIT), MediaPipe Face Mesh (Apache 2.0). Audio fingerprint and prosody features are homegrown MIT — see THIRD_PARTY_NOTICES.md.

Roadmap

16-week MVP to v1.0 internal release; details in docs/rollout-plan.md §1.

Phase	Weeks	Focus
0	0	Foundations — ADRs, repo scaffolding, AASIST-L licence verification
1	1–2	Liveness path + Android benchmark gate
2	3–4	Transcription, transcript exposure, direction scoring
3	5–6	Multi-modal trust (motion) + prosody features
4	7–8	Composable forwarding + W3C Verifiable Credentials envelope
5	9–10	Video affect (sentiment enrichment) + permissions
6	11–12	React Native + Capacitor bridges
7	13–16	OTA + hardware-backed signing + hardening + soak + v1.0
8	M4+	Public release readiness (gated by corpus milestones)
9	post-MVP	@sentlog/web (browser WebAssembly)

Status snapshot

• Documentation phase complete.
• Phase 0 not yet started.
• Open items for product and Phase 0 engineering gates listed in docs/rollout-plan.md §5.

Support

If Sentlog is useful to you and you'd like to support the work:

☕ Buy me a coffee

License

MIT. See LICENSE.

Third-party attributions (upstream models, datasets, and libraries the SDK builds on) are listed in THIRD_PARTY_NOTICES.md.