Roadmap

The committed sequence from spec v0.2 to a fully machine-checked ecosystem. Phases 0–1 are this repo’s current state; 2–4 are committed next steps; 5 is parked with explicit triggers.

Phase 0+1 — Spec v0.2 + contracts (done — this repo)

Contract repairs (native-speed/eager, endpointing ruling, family replaces fallback_of), two-plane replaceability, threat model, weight-update gate fix, license composition rule, maturity tags, ADR log; Runtime ABI v0.1 (cartridge-abi/include/cartridge_abi.h), descriptor + manifest schemas v0.2 with re-signed worked examples, task profile asr/v1 drafted as the cartridge-task-profiles KMP module.

Phase 2 — Tooling

  • cartridge-gradle-plugin — the staging step: per target resolve → verify manifest → stage artifacts → generate typed bindings from the descriptor (typed facade). Precedents to copy from SKaiNET's build logic: its cacheable JSON→generated-output tasks (CI-diffed), its schema-validation task, and the opt-in property gating of the Minerva host-verification task.

  • ctg-conform (OQ5) — schema-validate descriptor; verify digests + signatures (productizing verify_manifest.py); dlopen + ctg_abi_version check; smoke inference against test vectors bundled in the package; machine-readable report.

  • License gate — effective license must satisfy every per-artifact upstream license (SPDX expressions + entitlement-restricted markers); generates the NOTICE file including display duties (e.g. "Powered by Moonshine AI"); fails on undeclared licenses — the current Whisper-on-NPU cartridge state fails this gate by design.

Phase 3 — Retrofit the two real cartridges

  • Whisper-on-NPU cartridge (asr-whisper-tiny-npu): thin ABI wrapper over the existing C API, plus what a wrapper can’t hide — -fvisibility=hidden + version script + SOVERSION (2,144 exported symbols → the intended ~15), compile-time API version macro / struct-size field, normalized error model (no 0-with-stop_reason=3), the whisper_set_mel_filter_path process-global documented or fixed. Licensing remediation: LICENSE + NOTICE, per-artifact licenses in the manifest (Whisper weights → MIT attribution; FLEURS calibration → CC-BY; vendored NPU sample code → Apache-2.0 with modification notices), patched-driver requirement moved into requirements.driver + ctg_preflight. Descriptor v0.2 with measured WER.

  • ORT Moonshine cartridge (asr-moonshine-ort-cpu): ABI wrapper over moonshine-c-api.h; descriptor v0.2 (execution_mode: eager, family, endpointing: signal, honest per-ABI RTF); LICENSE
    NOTICE (MIT English models; the non-English Community-License constraint documented as the OQ10 worked case); Silero-blob artifact granularity decided (extract to file vs. accept whole-binary granularity, per the granularity rule in index.adoc#integrity).

  • Adapters onto asr/v1: both of the first app’s ASR adapters implement the profile; the duplicated session/channel/PCM plumbing (~250–300 LOC each) hoists into a shared support library; adapters shrink toward their genuine ~120 LOC of app logic (NLU, command assembly, context). The app’s command parity tests must stay green.

Phase 4 — Prove the claims

  1. WER harness first (zero ABI work — shells out to the existing CLI): supplies the quality.wer numbers Phase 3 descriptors require.

  2. Consumer #2 = a Kotlin/Native CLI: an --engine npu runner via cinterop against cartridge_abi.h (linuxArm64 target already declared; device-only engine, stated honestly). The first app stops being the only real consumer.

  3. Binary-plane swap demo: two same-io Whisper-on-NPU variants (en vs. de model set) swapped without recompiling the consumer.

  4. Profile-plane swap demo: Whisper-on-NPU ↔ a Moonshine engine behind asr/v1; the capability delta (languages, streaming, endpointing) surfaced from the descriptor diff.

  5. Weight-update end-to-end: one signed weights-update passing the v0.2 regression gate, applied and verified.

Phase 5 — Parked, with build-when triggers

Item Build when

Cartridge hub (Mode A)

there is an external or multi-tenant audience — not before (OQ4)

Revocation (OQ8)

shape-B dynamic delivery ships, or the first commercial cartridge does

Dynamic on-device delivery (threat-model shape B)

the first weight-update pipeline is real

Fleet telemetry (correcting measured performance/quality from the field)

≥ 2 products consume the same cartridge

A/B weight rollout

the first OTA weight update ships

NPU tenancy arbitration

two accelerator cartridges must coexist on one device