01. Sealed production run
The sealed LFM block-pair S2 artifact had two declared outputs and an input fixture with shape [1,2,2048]. The native runner loaded it under .cpuOnly and .cpuAndNeuralEngine, rejected non-finite values, and compared both outputs against the same pinned FP16 fixture.
MLComputePlan assigned 743/743 compute operations to ANE in the production graph. Both outputs met the established ANE tolerance: cosine similarity ≥ 0.999 and relative L2 ≤ 0.03.
| Output | cos(ANE, fixture) | relative L2 | Verdict |
|---|---|---|---|
lfm_h_after_liv | 0.999886220 | 0.017295585 | PASS |
lfm_h_after_gqa | 0.999761005 | 0.027863104 | PASS |
02. Debug variant: a different compiled artifact
Recompiling with LFM_PARITY_DEBUG=1 introduced 46 output taps instead of two. All 46 outputs passed the CPU-versus-fixture gate, but only 34/46 passed the ANE gate. The first reported ANE-gate failure was liv_14_mlp_output.
| Diagnostic output | ANE cosine | ANE relative L2 |
|---|---|---|
liv_14_mlp_output | 0.997936767 | 0.064227031 |
gqa_15_head_merge | 0.098921468 | 1.342520581 |
It would be incorrect to treat that diagnostic result as a failing production model. Additional outputs change the graph's compiled interface and can alter graph partitioning or transformations. We retain the negative diagnostic result but bind it to its own model package and compiled identity.
03. The native parity ladder
The runner reads the binary FP16 fixture directly, reconstructs logical tensor order from MLMultiArray strides, normalizes documented output layout permutations, and checks for NaN/Inf before computing max absolute error, mean absolute error, cosine similarity and relative L2.
production:
outputs = 2
assigned_ANE_compute_ops = 743 / 743
parity_ANE = PASS / PASS
diagnostic:
outputs = 46
CPU_parity = 46 / 46
ANE_parity = 34 / 46
verdict = ANE_ONLY_DIVERGENCE
these are DIFFERENT compiled artifacts04. Reproducibility boundary
The receipt pins SHA-256 identities for source, runner, fixture tree, emitted model package, compiled graph and report. The sealed production model.mlmodel SHA-256 is 3935a7b804d4a9a852faa7fe2ce0457cc2e23789547d22fd3cbe840efe5fd429.
Numerical parity is not a claim about every intermediate value. Compare the exact outputs of the exact compiled program you plan to execute. Use debug taps to localize divergence, not to reclassify another artifact.
Source receipt: ane-re/docs/receipts/LFM_S2_NATIVE_PARITY_2026-07-18.md. Referenced source artifacts are not public downloads.