{ "target_version": "0.4.12", "target_commit": "ffc142012af63d015fa19f05f29f9c5ff329a2cc", "status": "complete", "completed_this_stage": [ "Three-way port of existing Core ML host modules, preserving native device/artifact boundaries", "Added upstream _agreement, evals_tournament, integrations/agno and integrations/tools", "Exported OptionStabilityHook and jev_confidence", "Ported common ECE/temperature diagnostics and encode-state option-count reuse", "Pinned CI and local official reference to v0.4.12", "Ported shared shortlist/tournament and confidence fixes to MLX", "MLX now includes official batch/long/structured/hooks/calibration/evaluation/router/HTTP/MCP/integrations host modules with a native tensor bridge", "Preserved native MLX precision, compile, row limits and prefix-cache controls; cache covers option order, head budget and tokenizer identity", "Validated cached and uncached MLX batches against official output; updated native tensor fixtures without changing upstream MCP cache assertions", "Exposed native config metadata to OptionStabilityHook on both backends and verified sequential/parallel detection", "MLX service metadata reports native Device and DeviceType correctly", "Regenerated Core ML golden reference from official v0.4.12 and source checkpoint hashes", "Validated three MLX checkpoints in FP32/FP16 and four existing Core ML/ANE bundles against v0.4.12", "Pinned MLX build backend after reproducing metadata 2.5/Twine incompatibility; rebuilt artifacts pass strict checks", "Added MLX optional extras, service entrypoints, migration documentation and provenance manifest", "Added missing Agno and LangGraph optional dependency entrypoints to both distributions", "Added isolated feature performance/differential matrix with explicit upstream automatic, FP32 and FP16 autocast modes", "Corrected legacy benchmark FP32 label by disabling upstream row-gated autocast explicitly", "Validated unchanged standalone SDK live HTTP assertions against MLX and Core ML native models, extended/tournament and strict modes", "Avoided repeated Core ML sequence-shape changes inside logical batches while retaining token masks and row order", "Added native adaptive Core ML bundle assembly, identity checks, lazy graph selection and resource cleanup; existing single-graph bundles retain their behavior", "Exported B=1/2/4/16 graph sets for all three checkpoints, preserving full context capacities and source weights", "Validated all fifteen adaptive feature workload outputs against upstream; repeated the batch comparison in fresh processes", "Validated complete public outputs and 100 mixed repeated calls on each adaptive checkpoint", "Added native warmup API to both backends and honored MLX compile_warmup with upstream best-effort automatic/explicit-error behavior", "Rebuilt both distributions and exercised final inference-only wheels offline, including warmup and adaptive large/tail batches", "Audited 341 public surface entries against pinned upstream; native backend introspection explicitly implemented", "Completed exact official calibration temperature/binning differential and save/load roundtrip on real native logits for all three models per backend; repeated pseudo-label records are compatibility evidence only", "Fixed MLX local_files_only support and removed unsupported Core ML Router keyword", "Corrected pure FP32 public reference by disabling upstream automatic MPS autocast; retained historical mixed public reference for traceability", "Passed full public-output and 63-question numerical checks plus 100 repeated calls for all three native models in both precisions against pure FP32; decimal comparator preserves original numeric tolerance" ], "verification": { "coreml_full_pytest": { "passed": 334, "skipped": 1, "warnings": 7, "seconds": 29.55, "log": "/tmp/coreml-v0412-reuse-final-tests.log" }, "coreml_lint_and_format": "passed", "coreml_build": "sdist and wheel built, current unreleased package version remains 0.4.0", "coreml_twine_strict": "passed", "mlx_targeted_pytest": { "passed": 55, "command": ".venv/bin/pytest -q tests/test_shortlist.py tests/test_decision_helpers.py", "log": "/tmp/mlx-0412-shared.log" }, "mlx_full_pytest": { "passed": 319, "skipped": 1, "warnings": 2, "seconds": 28.86, "log": "/tmp/mlx-v0412-last-tests.log" }, "mlx_real_model": { "report": "laya-mlx/benchmarks/results/upstream-v0412-validation.json", "upstream_revision": "ffc142012af63d015fa19f05f29f9c5ff329a2cc", "rows": [ { "model": "laya", "dtype": "float32", "questions": 63, "argmax_agreements": 63, "probability_max_abs_error": 5.170394474396378e-06, "stable_calls": 100, "active_growth_bytes": 0 }, { "model": "laya", "dtype": "float16", "questions": 63, "argmax_agreements": 63, "probability_max_abs_error": 0.005444271459395356, "stable_calls": 100, "active_growth_bytes": 0 }, { "model": "laya-multilingual", "dtype": "float32", "questions": 63, "argmax_agreements": 63, "probability_max_abs_error": 1.1730551289779179e-06, "stable_calls": 100, "active_growth_bytes": 0 }, { "model": "laya-multilingual", "dtype": "float16", "questions": 63, "argmax_agreements": 63, "probability_max_abs_error": 0.0012887104780422431, "stable_calls": 100, "active_growth_bytes": 0 }, { "model": "laya-typed-decisions", "dtype": "float32", "questions": 63, "argmax_agreements": 63, "probability_max_abs_error": 2.6919645247369672e-06, "stable_calls": 100, "active_growth_bytes": 0 }, { "model": "laya-typed-decisions", "dtype": "float16", "questions": 63, "argmax_agreements": 63, "probability_max_abs_error": 0.001684852828996053, "stable_calls": 100, "active_growth_bytes": 0 } ] }, "coreml_real_model": { "report": "laya-coreml/benchmarks/results/upstream-v0412-validation.json", "reference": "laya-coreml/benchmarks/results/reference-v0412.json", "rows": [ { "repository": "aac6fef/laya-coreml", "questions": 63, "argmax_agreements": 63, "probability_max_abs_error": 0.0033076258324740704, "stability": { "calls": 100, "identical_rounded_results": true }, "passed": true }, { "repository": "aac6fef/laya-multilingual-coreml", "questions": 63, "argmax_agreements": 63, "probability_max_abs_error": 0.0015776376887714216, "stability": { "calls": 100, "identical_rounded_results": true }, "passed": true }, { "repository": "aac6fef/laya-typed-decisions-coreml", "questions": 63, "argmax_agreements": 63, "probability_max_abs_error": 0.0027266347551877512, "stability": { "calls": 100, "identical_rounded_results": true }, "passed": true }, { "repository": "aac6fef/laya-multilingual-coreml-ane", "questions": 59, "argmax_agreements": 59, "probability_max_abs_error": 0.0029248531754093587, "stability": { "calls": 100, "identical_rounded_results": true }, "passed": true } ], "note": "ANE includes only 59 fixture questions within its 96-token export; all three general graphs cover 63." }, "coreml_layout_regression": { "passed": 15, "log": "/tmp/coreml-v0412-layout.log" }, "mlx_build_and_twine_strict": "passed with hatchling 1.31.0", "mlx_clean_wheel_import": "passed; neither torch nor transformers installed", "mlx_parallel_target": { "passed": 1, "log": "/tmp/mlx-v0412-parallel.log" }, "mlx_clean_wheel_smoke": { "report": "laya-mlx/benchmarks/results/wheel-v0412-smoke.json", "passed": true, "models": 3, "network_attempts": 0, "without": [ "torch", "transformers" ], "paths": [ "single", "batch", "long", "structured", "OptionStabilityHook" ], "repeat_calls_per_model": 3 }, "feature_matrix_initial": { "report": "laya-mlx/benchmarks/results/v0412-features/summary.json", "workloads": [ "single", "batch8 heterogeneous states", "long", "structured", "OptionStabilityHook" ], "models": 3, "mlx_fp32_speedup_vs_upstream_fp32_range": [ 1.1045, 1.8081 ], "native_max_numeric_error": 0.0027, "note": "All 75 completed upstream/MLX workload rows plus one Core ML row agree on non-float fields; float tolerance 0.02. Fixed 20 iterations, 3 warmup for MLX/upstream. Core ML old hub bundle is a performance blocker (single p50 799.668 ms) and full run was stopped after single workload; no completion claim." }, "python_sdk_interop": { "mlx": "passed extended/tournament and strict", "coreml": "passed extended/tournament and strict", "commands": [ "scripts/sdk_interop.py --model models/laya-multilingual (MLX)", "scripts/sdk_interop.py --model models/hub/laya-multilingual-coreml (Core ML)" ], "logs": [ "/tmp/mlx-v0412-sdk.log", "/tmp/coreml-v0412-sdk.log" ] }, "coreml_uniform_padding": { "official_contract_and_regressions": 73, "input_tests": 19, "report": "laya-mlx/benchmarks/results/v0412-features/laya-coreml-uniform-float16.json", "single_p50_ms": 32.48, "old_single_p50_ms": 799.668, "batch8_p50_ms": 238.324, "long_p50_ms": 124.595, "structured_p50_ms": 9.252, "stability_p50_ms": 41.489, "all_five_nested_output_differentials_passed": true, "note": "Same old hub bundle before/after. Fresh v2 export alone was still ~805 ms; CPU-only ~448 ms. Batch and stability remain slower than upstream; no overall acceptance claim." }, "coreml_adaptive_real_models": [ { "model": "laya", "report": "../laya-coreml/benchmarks/results/v0412-laya-adaptive-validation.json", "passed": true, "questions": 63, "argmax_agreements": 63, "max_probability_error": 0.0033076258324740704, "stability": { "calls": 100, "seconds": 21.50356466700032, "rounded_public_results_identical": true, "rss_before": 3200614400, "rss_after": 3088285696, "rss_samples": [ 3200614400, 3200532480, 3201499136, 3201581056, 3201597440, 3201859584, 3201957888, 3201990656, 3202088960, 3202121728, 3088285696 ], "memory_note": "Process RSS includes Core ML/framework caches; not an active-tensor allocator measurement" } }, { "model": "laya-multilingual", "report": "../laya-coreml/benchmarks/results/v0412-laya-multilingual-adaptive-validation.json", "passed": true, "questions": 63, "argmax_agreements": 63, "max_probability_error": 0.0015776376887714216, "stability": { "calls": 100, "seconds": 18.68126870800006, "rounded_public_results_identical": true, "rss_before": 3252371456, "rss_after": 2797240320, "rss_samples": [ 3252371456, 3252355072, 2861334528, 2861957120, 2861957120, 2862022656, 2862055424, 2862055424, 2797191168, 2797174784, 2797240320 ], "memory_note": "Process RSS includes Core ML/framework caches; not an active-tensor allocator measurement" } }, { "model": "laya-typed-decisions", "report": "../laya-coreml/benchmarks/results/v0412-laya-typed-decisions-adaptive-validation.json", "passed": true, "questions": 63, "argmax_agreements": 63, "max_probability_error": 0.0027266347551877512, "stability": { "calls": 100, "seconds": 22.83919779200005, "rounded_public_results_identical": true, "rss_before": 3269214208, "rss_after": 3270230016, "rss_samples": [ 3269214208, 3269148672, 3269689344, 3269738496, 3269754880, 3269754880, 3269771264, 3269787648, 3270213632, 3270180864, 3270230016 ], "memory_note": "Process RSS includes Core ML/framework caches; not an active-tensor allocator measurement" } } ], "coreml_adaptive_performance": { "report": "laya-mlx/benchmarks/results/v0412-features/summary.json", "paired_batch_repeat": "laya-mlx/benchmarks/results/v0412-batch-repeat/paired-summary.json", "models": 3, "workloads_per_model": 5, "warmup": 3, "iterations": 20, "precision_note": "Core ML FP16 export; upstream FP32 and MPS FP16 autocast reported separately; mixed arithmetic differs", "coreml_over_upstream_fp32_p50_range": [ 1.07, 2.07 ], "repeated_batch_speedups_over_upstream_fp32": [ 1.073, 1.201, 1.082 ], "repeated_batch_speedups_over_upstream_fp16_autocast": [ 1.113, 1.345, 1.121 ], "full_nested_output_comparison": "all passed at 0.02 float tolerance, categorical and structural values exact" }, "final_build_twine": { "both": "sdist/wheel built and strict Twine check passed", "log": "/tmp/laya-v0412-final-twine.log", "version": "0.4.0 unpublished", "includes": "Final base-instance reuse shape cache, dense conversion defaults, LlamaIndex repair, MLX offline model resolution" }, "final_wheel_smoke": { "mlx": "laya-mlx/benchmarks/results/wheel-v0412-smoke.json", "coreml": "laya-coreml/benchmarks/results/wheel-v0412-adaptive-smoke.json", "models_per_backend": 3, "offline_network_attempts": 0, "without_training_frameworks": true, "includes": [ "single", "batch", "long", "structured", "OptionStabilityHook", "native warmup" ], "coreml_batch_states": 17, "coreml_loaded_extra_capacities": [ 2, 4, 16 ], "final_cache_source_hash_verified": true, "coreml_models": "models/v0412/MODEL-fp16-dense-adaptive" }, "mlx_real_compile_warmup": { "report": "laya-mlx/benchmarks/results/v0412-real-warmup.json", "automatic_flag_true_false": "passed", "explicit_warmup_preserves_output": "passed" }, "coreml_adaptive_sdk_interop": { "log": "/tmp/coreml-v0412-adaptive-sdk.log", "extended_tournament_and_strict": "passed" }, "public_api_audit": { "report": "docs/UPSTREAM_0412_API_AUDIT.json", "entries": 341, "unclassified_differences": 0 }, "real_calibration": { "mlx": "laya-mlx/benchmarks/results/v0412-mlx-calibration.json", "coreml": "laya-coreml/benchmarks/results/v0412-coreml-calibration.json", "note": "Exact same native logits supplied to official fitter; pseudo-label repetitions are not independent calibration-quality data" }, "pure_fp32_reference": { "report": "laya-coreml/benchmarks/results/reference-v0412-fp32-torch214.json", "sha256": "efc2dc32ce042142210510cef0121c619bcfe0abb541119ea1ccb17563d13de4", "autocast_enabled": false, "historical_reference_note": "reference-v0412.json raw logits are FP32; public large-batch outputs used row-gated autocast. Superseded for strict public FP32 comparisons.", "torch": "2.14.0" }, "pure_reference_native_validation": { "mlx": "laya-mlx/benchmarks/results/upstream-v0412-pure-fp32-validation.json", "coreml": "laya-coreml/benchmarks/results/v0412-*-pure-reference-validation.json", "models_per_backend": 3, "precisions": [ "float32", "float16" ], "questions_per_model_precision": 63, "stable_calls_per_model_precision": 100, "passed": true }, "mixed_shape_audit": { "report_directory": "laya-mlx/benchmarks/results/v0412-mixed", "model": "laya", "precision": "float32", "requests_per_iteration": 4, "upstream_p50_ms": 202.538, "mlx_p50_ms": 71.997, "coreml_uncached_p50_ms": 1117.43, "coreml_bounded_cache_p50_ms": 78.853, "public_output_max_decimal_difference": 0, "coreml_cached_process_peak_rss_gb": 9.18, "upstream_process_peak_rss_gb": 2.91, "note": "Initial four-shape experiment passed, but broader campaign exposed LRU reload thrashing. Replaced with aged-frequency admission and base-graph fallback; final campaign in progress. Do not treat initial bounded-LRU timings as final production acceptance." }, "shape_cache_targeted_tests": { "passed": 24, "log": "/tmp/coreml-v0412-shape-cache-tests.log" }, "real_optional_integrations": { "report": "benchmarks/results/v0412-real-integrations.json", "passed": 8, "versions": { "laya-coreml": "0.4.0", "laya-mlx": "0.4.0", "agno": "1.8.4", "langchain-core": "1.6.12", "langgraph": "1.2.16", "llama-index-core": "0.14.25", "crewai": "1.15.28" }, "scope": "Real framework types; deterministic native decision runner; no external model/tool execution", "fix": "LlamaIndex SingleSelection import and abstract prompt mixin support; both single and multi selectors instantiate real BaseSelector", "wheel_report": "benchmarks/results/v0412-real-integrations-wheel.json", "wheel_passed": 8 }, "lockfiles": "Regenerated both uv.lock files for all declared extras; uv lock --check passed", "frequency_cache_real_models": [ { "report": "laya-coreml/benchmarks/results/v0412-laya-float16-frequency-v1-validation.json", "passed": true, "questions": 63, "agreements": 63, "max_probability_error": 0.0033076258324740704, "cache_capacity": 4, "maximum_cached_instances": 4, "stable_calls": 100, "repeat_seconds": 15.045183374999397, "rss_before": 6648692736, "rss_after": 6654377984, "cache_released": true }, { "report": "laya-coreml/benchmarks/results/v0412-laya-float32-frequency-v1-validation.json", "passed": true, "questions": 63, "agreements": 63, "max_probability_error": 6.4322300447061664e-06, "cache_capacity": 4, "maximum_cached_instances": 4, "stable_calls": 100, "repeat_seconds": 28.96876499999962, "rss_before": 12689719296, "rss_after": 12679118848, "cache_released": true }, { "report": "laya-coreml/benchmarks/results/v0412-laya-multilingual-float16-frequency-v1-validation.json", "passed": true, "questions": 63, "agreements": 63, "max_probability_error": 0.0015776376887714216, "cache_capacity": 4, "maximum_cached_instances": 4, "stable_calls": 100, "repeat_seconds": 4.619535499999984, "rss_before": 5897306112, "rss_after": 5903499264, "cache_released": true }, { "report": "laya-coreml/benchmarks/results/v0412-laya-multilingual-float32-frequency-v1-validation.json", "passed": true, "questions": 63, "agreements": 63, "max_probability_error": 6.319790438052975e-07, "cache_capacity": 4, "maximum_cached_instances": 4, "stable_calls": 100, "repeat_seconds": 18.36190175000047, "rss_before": 10874011648, "rss_after": 10891083776, "cache_released": true }, { "report": "laya-coreml/benchmarks/results/v0412-laya-typed-decisions-float16-frequency-v1-validation.json", "passed": true, "questions": 63, "agreements": 63, "max_probability_error": 0.0027266347551877512, "cache_capacity": 4, "maximum_cached_instances": 4, "stable_calls": 100, "repeat_seconds": 16.410943332999523, "rss_before": 6661718016, "rss_after": 6707200000, "cache_released": true }, { "report": "laya-coreml/benchmarks/results/v0412-laya-typed-decisions-float32-frequency-v1-validation.json", "passed": true, "questions": 63, "agreements": 63, "max_probability_error": 1.699359256368549e-06, "cache_capacity": 4, "maximum_cached_instances": 4, "stable_calls": 100, "repeat_seconds": 30.68189058399912, "rss_before": 12811059200, "rss_after": 12814876672, "cache_released": true } ], "real_llamaindex_regression_tests": { "per_backend": 2, "passed": true, "includes": [ "SingleSelection result types", "BaseSelector inheritance", "single/multi selectors", "sync/async methods", "prompt mixin contract" ], "ci": "Actual llamaindex extra now installed in both CI workflows" }, "phase_chain_matrix": { "directory": "laya-mlx/benchmarks/results/v0412-phase-chain", "public_output_comparison": "54/54 rows passed (including 18 upstream reference rows), FP32 tolerance 0.0001", "status": "performance not accepted", "issues": [ "English Core ML batch P50 169.1 ms vs upstream 138.0 ms; fresh isolated cache-disabled/enabled repeats 111.6/105.7 ms indicate workload/machine-state sensitivity", "Core ML mixed-shape phase-change mean 435\u2013715 ms vs upstream 145\u2013204 ms due deferred cache admission/model creation" ] }, "baseline_environment_correction": { "affected_directories": [ "v0412-phase-chain", "v0412-torch27-matrix", "v0412-abba-batch-torch27" ], "reason": "These performance runs used PyTorch 2.7 from the conversion environment. Final qualification uses upstream PyTorch 2.14 from the MLX reference environment.", "golden_note": "Pure FP32 correctness golden already used PyTorch 2.14; regenerated explicit-version reference is byte-identical." }, "final_wheel_source_hashes": { "mlx": "f622efca650c0f832884264d2acff292cb6a97d536767e339e7f6b231f8cf3f4", "coreml": "9f16a897b0f5ed642ac3c4919ef899f9375771863cd60822d60148b3537913f3", "installed_source_matches_checkout": true }, "final_dense_coreml_models": [ { "report": "laya-coreml/benchmarks/results/v0412-laya-float16-cached-validation.json", "questions": 63, "agreements": 63, "max_probability_error": 0.0033076258324740704, "precision": "float16", "stable_calls": 100, "maximum_extra_instances": 4, "cache_released": true, "reference_sha256": "efc2dc32ce042142210510cef0121c619bcfe0abb541119ea1ccb17563d13de4", "passed": true }, { "report": "laya-coreml/benchmarks/results/v0412-laya-float32-cached-validation.json", "questions": 63, "agreements": 63, "max_probability_error": 6.4322300447061664e-06, "precision": "float32", "stable_calls": 100, "maximum_extra_instances": 4, "cache_released": true, "reference_sha256": "efc2dc32ce042142210510cef0121c619bcfe0abb541119ea1ccb17563d13de4", "passed": true }, { "report": "laya-coreml/benchmarks/results/v0412-laya-multilingual-float16-cached-validation.json", "questions": 63, "agreements": 63, "max_probability_error": 0.0015776376887714216, "precision": "float16", "stable_calls": 100, "maximum_extra_instances": 4, "cache_released": true, "reference_sha256": "efc2dc32ce042142210510cef0121c619bcfe0abb541119ea1ccb17563d13de4", "passed": true }, { "report": "laya-coreml/benchmarks/results/v0412-laya-multilingual-float32-cached-validation.json", "questions": 63, "agreements": 63, "max_probability_error": 6.319790438052975e-07, "precision": "float32", "stable_calls": 100, "maximum_extra_instances": 4, "cache_released": true, "reference_sha256": "efc2dc32ce042142210510cef0121c619bcfe0abb541119ea1ccb17563d13de4", "passed": true }, { "report": "laya-coreml/benchmarks/results/v0412-laya-typed-decisions-float16-cached-validation.json", "questions": 63, "agreements": 63, "max_probability_error": 0.0027266347551877512, "precision": "float16", "stable_calls": 100, "maximum_extra_instances": 4, "cache_released": true, "reference_sha256": "efc2dc32ce042142210510cef0121c619bcfe0abb541119ea1ccb17563d13de4", "passed": true }, { "report": "laya-coreml/benchmarks/results/v0412-laya-typed-decisions-float32-cached-validation.json", "questions": 63, "agreements": 63, "max_probability_error": 1.699359256368549e-06, "precision": "float32", "stable_calls": 100, "maximum_extra_instances": 4, "cache_released": true, "reference_sha256": "efc2dc32ce042142210510cef0121c619bcfe0abb541119ea1ccb17563d13de4", "passed": true } ], "final_sdk_interop": { "mlx_log": "/tmp/laya-v0412-final-mlx-sdk.log", "coreml_log": "/tmp/laya-v0412-final-coreml-sdk.log", "both": "PASS original real HTTP assertions: extended/tournament, strict, auth and limits" }, "final_fp32_matrix": { "directory": "laya-mlx/benchmarks/results/v0412-final", "upstream_torch": "2.14.0", "public_output_rows_passed": 54, "mlx_p50_wins": 18, "coreml_p50_wins": 14, "workloads_per_backend": 18, "status": "All output rows passed; four sequential Core ML P50 regressions retained and investigated in paired confirmation" }, "final_coreml_paired": { "report": "laya-mlx/benchmarks/results/v0412-final-paired/paired-summary.json", "upstream_torch": "2.14.0", "order": [ "upstream", "coreml", "coreml", "upstream" ], "workloads": 18, "output_comparisons": "all passed, FP32 tolerance 0.0001 and exact categories/structure", "p50_wins": 18, "p95_wins": 18, "mean_throughput_wins": 18, "p50_speedup_range": [ 1.0569071326363693, 1.9801773792143524 ], "all_runs_retained": true }, "final_package_receipt": "benchmarks/results/v0412-package-gates.json" }, "test_adapter_boundaries": [ "Namespace/import/mock rewriting only; expected host values unchanged", "Excluded ONNX-specific runner/constructor checks and PyTorch backend switching properties", "Training CLI assertion excluded by syntax instead of unstable line numbers; local explicit refusal test retained", "Core ML device/thread-pool semantics covered by local boundary tests", "Preserved original email slice spelling because upstream asserts source text" ], "remaining": [], "completion_claim": true, "completion_scope": "Pinned upstream v0.4.12 applicable host API, native backend boundaries, local tests, finite real-model fidelity/stability and measured FP32 workloads; unpublished, no universal accuracy/performance claim.", "qualification_report": "docs/QUALIFICATION_0_4_12.md", "limitations": [ "Finite fixture parity is not universal model accuracy or calibrated-confidence quality", "Core ML extra execution instances increase RSS; final FP32 feature run recorded 5.10\u201313.51 GiB process peak RSS, not a complete device-allocation measure", "Cold/unseen shapes, loading and long-tail misses are outside steady-state guarantees", "Shared desktop load varied; sequential Core ML regressions and alternating-order confirmation both retained", "FP16 and fixed-shape ANE have separately documented precision/capacity boundaries", "No remote CI, release, push or Hub artifact replacement performed" ], "delivery_status": "published_and_verified", "release_version": "0.5.0", "delivery_remaining": [], "release_receipt": "benchmarks/results/release-v050.json" }