What actually ran?
Core ML compiles a model graph. Small changes in graph structure, shape, or diagnostics can change the compiled result. Name the precise artifact and configuration.
How the artifacts are documented ↗START HERE / RESEARCH FIELD GUIDE
MacBrains Labs / Start here
Before the number,
know the claim.
A one-minute map to MacBrains Labs research. You do not need to know Core ML or the Apple Neural Engine to follow the evidence.
FIRST PRINCIPLES / 01
The four questions.
Think of the Neural Engine as one possible destination for computational work, not a guarantee of correctness or speed.
Where was it assigned?
Apple's Core ML planner may report compute operations assigned to ANE, CPU, or GPU. This is useful placement evidence, but not a low-level trace proving each operation physically ran on ANE.
Inspect a placement receipt ↗Was the output correct?
A model can compile and receive an ANE assignment yet fail an output-parity test. Compare exact outputs against declared fixtures using defined numerical tolerances.
Read a passing and failing comparison ↗Did it get faster?
Throughput depends on batch shape, dispatch cost, memory, and what was timed. A speed measurement does not validate model quality or the claims made about deployment.
See a disproven performance hypothesis ↗TRUST BOUNDARIES / 02
Read the labels, not just the headline.
Measured means an observation under a named setup. Planner reported describes the compiler's view, not an ANE hardware trace.
Source-bound means the published claim links to its artifacts or documented receipts. Verified archive means published files and identifiers passed stated integrity checks, not peer review.
Negative results, incomplete chains, and changes to an earlier hypothesis belong in the research record.
Search all public entries ↗Independent Apple Silicon research. MacBrains Labs is not affiliated with Apple.