LFM3-1.7B layer-14 diagnostic under Core ML cpuAndNeuralEngine (same chip, single device)

01Scope

Same-chip rerun: repeatability only, not cross-device replication. Every execution in this record ran on one device, the same chip. Device identity is self-reported-by-OS-and-CoreML-on-run-host-Neo; not independently re-derived. Novelty is UNCLAIMED.

02Claim, verbatim

On a single device whose OS/Core ML reported identity is chip Apple A18 Pro, hwModel Mac17,5, macOS 27.2 (build 26B123), CoreML 3600.25.2, LFM3-1.7B layer-14 inference under computeUnits=cpuAndNeuralEngine is output-reproducible under source rebuild: THREE counted, git-frozen, source-rebuilt clean-room runs (one profile-scoped run ledger keyed to profileHash 2f0cc0c9, totalRuns=3, countedRuns=3, preFreezeRunsNotCounted=0; this ledger is scoped to the current frozen tool commit 926a78ad, and the same profileHash has 8 logged executions device-wide: 5 counted/frozen across tool commits 833b3fc8/5f827c84/926a78ad and 3 pre-freeze, one counted run's on-disk dump deleted and disclosed with recovered records, all enumerated in allDeviceExecutionsOfProfile and execution-log.jsonl) each produced output byte-identical (24/24 tensors, 0 differing) to the archived device-run baseline index. The replication runner was built from git-committed source at commit c45e8c61 (binary SHA-256 746e1200, which differs from the baseline binary 48d2d0aa); the clean-room tool and profile were git-frozen at 926a78ad / 9d6aa2a2 before each counted run. This is source-level rebuild output reproducibility, not binary identity and not cross-device scientific replication. A separately frozen, counted CPU-vs-ANE control (control profile 33e0e5a9, totalRuns=2, countedRuns=2, under the same frozen tool 926a78ad) ran both arms with the SAME source-rebuilt binary (SHA-256 e0ef750e, built from committed probe source at commit 055b54f7), differing only in the LFM3_PROBE_CPU_LAYERS environment variable (unset for the ANE arm, =14 for the layer-14 CPU arm): a like-for-like, same-binary comparison. An independently written recomputation (own IEEE-half decode) on the two frozen control dumps - cross-checked byte-identical (24/24 files) to the archived Oct-31 probe dumps - gives ANE-enabled output abs-L2 to the FP32 teacher of 0.0062 to 0.0097 at positions 161/203/217 versus the layer-14-only CPU arm's 0.036 to 0.054; i.e. the CPU-only abs-L2 is 3.8x to 8.7x the ANE-enabled abs-L2. The FP32 teacher is computed by the same runner from the fixture, and the independent recomputation reads the same frozen dumps and shares metric definitions, so it is documentary cross-checking, not an independent measurement. Core ML's compute plan assigns all 447 layer-14 ops to the ANE (0 CPU/GPU/unassigned); ANE execution is INFERRED from this placement plus the same-binary CPU control, not from a per-op hardware execution trace. Same-chip, single-device, layer-14, three-position diagnostic only; no novelty claim.

03What the record does not say

• Single device (Mac17,5, macOS 27.2 build 26B123); three source-rebuilt runs against the archived same-chip baseline, which is not multi-device reproducibility.
• One layer (14) of LFM3-1.7B at three token positions (161, 203, 217) only; do not generalize.
• Only tier A (ANE-enabled) and the layer-14-only CPU control are device-measured; emulated tiers N/G/E, cosine-between-error-vectors and projected_share are not covered.
• Compiler placement is not direct hardware execution: all 447 layer-14 ops are assigned to the ANE in the plan, but no per-op hardware trace exists, so ANE execution is INFERRED.
• The CPU control is layer-14-only (layer 14 forced to CPU via LFM3_PROBE_CPU_LAYERS=14, rest on ANE); the whole-chain CPU-only run produced no finite layer-14 dump, so the ANE-vs-CPU comparison is layer-local.
• The independent recomputation shares metric definitions and reads the same raw dumps; it verifies the archived error metrics, not an independent measurement, and is not cross-device replication.
• Replication runner binary differs from the baseline binary (different SHA-256): source-level rebuild output reproducibility, not binary identity.
• The replication runner (746e1200 @ c45e8c61) and the CPU-vs-ANE control binary (e0ef750e @ 055b54f7) are distinct build artifacts; the control's same-binary claim is internal to its two arms only.
• Diagnostic only: no responsible ANE op identified, no release gate changed; prior-art coverage of the specific directional claim is INCONCLUSIVE and novelty is explicitly UNCLAIMED.
• Tensor-file unit: tensorFilesPerPosition=24 is the per-position tensor count used in the byte-identity comparisons; each control-arm dump directory holds 25 files (the 24 compared per-position tensors plus one additional non-compared file), so the 24/24 control cross-check covers the 24 tensor files only.

04Every execution of the setup profile

Setup profile 2f0cc0c9371123c7ea4055e755ff9d0ef714ed07e5092ca477581bad14be4042 was hashed and committed before the first counted run. The device log lists 8 logged executions of it: 5 counted and git-frozen, 3 before the profile froze and not counted. The output of 1 counted run was deleted from the device; its recovered record and the cause analysis are in deletion-disclosure.md. No execution is omitted.

From execution-log.jsonl, republished exactly as reviewed. The log on the device is append-only; the one correction shown was appended after review. The logged value is the directory modification time; the run's own runs-log records 20:03:04Z. A correction line was appended to the device log after review; the republished log is the reviewed prefix.

05Review and bindings

• Reviewed package c188def1e88a0d94018dfe99f63aed795f58848871841d1c089d079a28aebff1: the hand-off manifest the final reviewer read. It is not republished; it binds the files below by hash.
• Final review: INCLUDE from macbrains-labs-reviewer for that package hash, verdict receipt 4f8959c34bf39ebafb5ee681aa5cca71857cd987d6b85b2ea75e3cc1ff04230b.
• evidence reviewer: opencode-go/deepseek-v4.1-flash, INCLUDE at 0.82, session ses_e60904ac1ffeacsEnVw4ZofYgS, receipt a242e7104c211bde9238a8a53c3ac9b44b9c6ac877a4d075037292f0077485a4
• skeptical reviewer: opencode-go/glm-5.3, INCLUDE at 0.80, session ses_e60904ac1ffePgIg7R2G9ym86a, receipt b2d85d88f4d7ad8714a0c2d3cb37c01cbc034ccc001a57b26ec57cb9cc8ca907
• Reviewed bundle 0d2b3f787cf0f98e6d12ed9cebd614d028b0f39e9e210e55fb93cc14ca6f49b3, round 5, schema macbrains.research.diagnostic-review-bundle.v1.
• Diagnostic policy digest 3c3b6ba4f8ee69849f8c24672d26bc51f1ca36b429203f996849994b8d7ea3fa at commit 4d23b937e55fc25a12428eb6ef965ec8666c2cc1.

A reviewed record and an archived DOI are not peer review. Cite this version as DOI 10.5281/zenodo.23905811; all versions of the record series share DOI 10.5281/zenodo.23798036.

06Files

The GitHub release ane-record-c188def1e88a0d94018dfe99 carries the same files with SHA256SUMS; scripts/verify.mjs in the record repository re-checks every hash and claim field.