Model assemblies

New-KokoroDecoderGraph.ps1 is the current parsed two-node decoder contract. The build does not execute it as a script. It lowers the AST to a graph hash and control schema persisted in the managed Android host assembly. Its inputs are precomputed acoustic tensors; it does not implement phoneme-to-PCM speech.

From a verified checkout with PowerShell 7.4 or newer, the existing host build and its Windows identity test are:

./setup-kokoro.ps1 -Step 4 -KeepIntermediates -AcceptWritePlan
./tools/Test-WindowsModelAssembly.ps1

The current internal controls are asr, F0_curve, N, style, gb, har8, mask, mask8, and capacity. They are coupled tensor boundaries, not a stable public prosody API. The test matches the saved host assembly's graph hash to an independent lowering of the checked-out source.

Separate validation artifacts now exist outside Git:

• Kokoro.Phonemes.dll persists the pinned phoneme ID and voice-row methods; the tested voice variant embeds the raw af_heart rows.
• Kokoro.Weights.FP32.dll and Kokoro.Weights.FP16.dll contain indexed, hash-checked tensor resources. All 548 resources load on Windows. The FP16 payload has only sampled weight-error evidence, not audio-quality evidence.

These three artifacts have not been joined into the final model DLL. They do not contain the full Kokoro computation, a text-to-phoneme adapter, a QNN-free Hexagon path, or live speech. The checkpoint reader is host-side build tooling; the released app must load only verified model resources and must not read the checkpoint or historical QNN contexts.