73,110 actual Swift calls complete with unchanged choices. Preserve raw model probabilities, conversion failures, every existing report and every model artifact. Add source-pinned runtime report and complete trace; clarify raw versus stable scores in the model card.
Add experimental INT4 weights and full matched benchmark
INT8 weights with FP16 compute: 813,422 bytes versus 1,511,163 bytes. All 24,370 choices match PyTorch/FP16, with 24,359 correct (99.9549%); matched median 0.990 ms for both. Numerical parity fails: 64 INT8 probability errors above 0.005, maximum 0.067738, no normalization failures. Default and ANE-gather artifacts unchanged. Publish exact settings, complete per-row trace, demo checks and compute-plan/fallback reports.
Compare the full published synthetic test and report failed numerical gates
All 24,370 decisions: PyTorch and both unchanged Core ML exports score 24,359 correct (99.9549%) with 100% selected-index agreement. Strict parity fails: 11 probability errors above 0.005 per export (max 0.0204874) and one sum outside the Swift guard. Original median 1.003 ms; ANE-gather 1.052 ms on M5 Pro. Publish exact manifest and complete per-row trace; no normalization or tolerance changes. Historical browser reproduction now pins its exact source commit.
Add live Swift browser demo and matched variant benchmark
Record actual DOM observe/decide/act/readback on all three original forms: 100/100 choices. Release Swift ABBA comparison: 200/200 correct per variant, original median 0.912 ms vs ANE gather 0.961 ms. Unchanged model artifacts. Public demo data only; animation excluded from the benchmark.
Add optional ANE-gather variant with 98.2 percent ANE placement
Preserve int32 inputs and all trained weights while moving embedding lookup onto ANE. 196/196 reference decisions pass in Python and Swift; maximum probability error 0.002336. Matched median latency is 0.970 ms versus 0.915 ms for the unchanged default, so this variant is opt-in.
Add real-input ANE profiling across four compute policies
Record scheduler placement and bounded three-form timing with unchanged model hashes. CPU + ANE: 149 ANE / 24 CPU operations, 0.929 ms median, 0.973 ms p95. All 120 timed predictions correct. Results are not utilization or power traces.
Add CUA-S1-FORMS FP16 Core ML conversion and verified artifacts
Add the 706,048-parameter CUA-S1-FORMS decision scorer as a portable FP16 Core ML package and compiled FluidAudio bundle, with encoding helper, pinned provenance, licenses, checksums, and conversion reports.
Validation uses the complete pinned 196-row upstream demo, not a new synthetic dataset. PyTorch, Core ML ALL, and Core ML CPU_AND_NE select all 196 labeled options correctly. Maximum probability differences are 0.003099 and 0.002336 respectively, below the preselected 0.005 tolerance. The unmodified Cua evaluator reports zero wrong actions, wrong targets, or unsafe actions on these saved predictions. Seventeen Python regression tests pass.
FluidAudio's Swift runtime independently matches all 196 decisions against the PyTorch probabilities (maximum error 0.002336), including local package loading, shared compiled-model cache loading, UTF-8 encoding, and concurrent/reordered-option checks. The Swift library builds; local XCTest execution is unavailable because this Mac has Command …
