A technically exceptional and genuinely original native HarmonyOS entry whose strong engineering and broad platform use are undercut by missing required deliverables, above all the recorded demo and the final .hap, leaving its core claims unverifiable as submitted.
All members agree the capsule idea is genuinely novel: plain-language requests become schema-validated JSON mini-apps with a per-permission consent gatekeeper and on-device-first AI routing. The median follows B and C over A, because with no demo and no deck the novelty rests on code and docs rather than a shown working product.
Members disagree here: scores range by 2.0 points.Real everyday use cases are implemented and a judge-reported failure was diagnosed and fixed, but the cloud path needs a user-supplied API key, the on-device path needs a separate 480 MB model push, and no run is verifiable in the evidence. B's middle position fits best: C's 5 undersells the offline rule parser that delivers many capsules in code, and A's 7 ignores how narrow the submitted experience is.
Members disagree here: scores range by 2.0 points.The engineering is substantial and real: about 45,000 source lines, 420 tests in 56 files, 366 commits during the event, a strict validator, a hand-written safe expression parser instead of eval, and a Node-API C++ engine port. The score is held back because the team's own audit lists claim/code gaps and features marked built but never verified.
Usage goes far beyond a webview: Calendar, Notification, Form Kit widgets, Scan, Vision OCR, Sensor, Network and Preferences kits plus the self-ported Cactus inference engine. The median credits that breadth while accepting B's caveats that the .so is arm64-only and the model weights are not bundled, so the on-device path is unverified on an x86 emulator.
No recorded demo exists anywhere: the team's own compliance audit records it as missing, the 15 screenshots resolve to placeholders and sample eval images rather than the app running, and the only checkable demo link points to esbuild.github.io. The members differ only in wording, not substance, on this total absence.
Transparency is excellent: AI_WORKFLOW.md, AI integration docs, a compliance audit that names its own gaps, third-party attribution with hashes, and exact test commands. Reproduction fails though, because there is no final signed .hap, the model must be pushed separately, the form's setup field is blank, 8 commits were unpushed at check time, and no deck exists.
claude:glm-5.3-flash75.070% agreedots-studio/dots-3-note-preview:free63.090% agreeinclusionai/ling-3.0-flash-sante:free62.090% agreeclaude:glm-5.3–judgeThe council, the evidence pack, the prompts and the queue are all on GitHub. If a review helped you, a star helps other teams find it.