robbiemu/coreml-numerics-evidence

Core ML Numerics Evidence

This collection contains the experimental evidence for the paper From Numerical Differences to Bitwise Prediction in Core ML Arithmetic. The motivating controlled and pretrained-classifier results were reported in Backend-Sensitive Numerical Differences and Discrete Decisions in Core ML, with released records in the original artifact repository. Start with the eight study designs below. Each explains the question, controls, progression and limits, and links directly to its experiment reports, data folders and available reproduction instructions. The ImageNetV2 images used in Study 2 are obtained separately. See the image setup instructions.

The eight studies are the scientific organization of the work. Each experiment has one primary study, determined by the question it was designed to answer. Other studies may cite its results or controls as dependencies, without duplicating its membership or counting its evidence again. A physical record is a report or folder that may contain several experiments. Its location does not determine their study assignments. Records are stored once, with links to the relevant sections.

Studies

No.	Study design	Primary experiments	Purpose
1	Controlled setting-sensitive decisions	Ex. 1	A fixed synthetic model shows stable decision differences on the same machine.
2	Decision differences in a pretrained classifier	Ex. 2	A frozen ResNet-18 evaluation tests individual predictions beyond the constructed example.
3	Localization and robustness of the controlled difference	Ex. 3a, 3b, 4–7	Quantization, execution, localization and downstream controls establish the scope of the controlled difference.
4	Recovering and diagnosing CPU convolution arithmetic	Ex. 8–10, 11a, 11b, 12–13	Exact recurrences, order-changing interventions and a model-decision test characterize the comparator.
5	Recovering ANE-associated contribution arithmetic	Ex. 14–15	Grid-rounding probes predict constrained arrays, while ordinary-weight failures motivate refinement in Study 7.
6	Recovering phase-dependent GPU convolution arithmetic	Ex. 16a, 16b, 16c, 17, 18a, 18b, 19a, 19b, 19c, 19d, 20	A sequence of discriminating probes leads from dense residuals to complete reconstruction expressions.
7	Refinement and transfer of the ANE and GPU results	Ex. 21–26, 27a, 27b, 28–30	Operand-exponent and transform-order refinements complete the initial accounts, followed by operand, stack and configuration tests.
8	Reconciling ANE rounding observations across operations	Ex. 31a, 31b, 31c, 31d	Public-probe replay and matched arithmetic challenges separate operation/configuration effects from an unrecovered historical setup.

Experiment reference points

Experiment	Primary study	Design
Ex. 1	Study 1	Controlled decision comparison
Ex. 2	Study 2	Pretrained classifier comparison
Ex. 3a	Study 3	Quantization controls
Ex. 3b	Study 3	Execution robustness
Ex. 4	Study 3	Gain sensitivity
Ex. 5	Study 3	Source-reference comparisons
Ex. 6	Study 3	Boundary localization and fidelity
Ex. 7	Study 3	Downstream controlled-model reconciliation
Ex. 8	Study 4	CPU recurrence identification
Ex. 9	Study 4	Fresh CPU prediction
Ex. 10	Study 4	CPU channel-order intervention
Ex. 11a	Study 4	Nearby CPU configuration transfer
Ex. 11b	Study 4	CPU regime factorial
Ex. 12	Study 4	CPU applicability diagnostic
Ex. 13	Study 4	CPU state intervention
Ex. 14	Study 5	ANE contribution-rounding probes
Ex. 15	Study 5	ANE grid-rule development tests
Ex. 16a	Study 6	GPU cancellation and bias tests
Ex. 16b	Study 6	GPU filter and inactive-coefficient controls
Ex. 16c	Study 6	GPU single inactive-coefficient tests
Ex. 17	Study 6	GPU spatial phase and support
Ex. 18a	Study 6	GPU numerical-stage selectors
Ex. 18b	Study 6	GPU reconstruction interactions
Ex. 19a	Study 6	GPU terminal-rounding test
Ex. 19b	Study 6	GPU four-cell reconstruction tests
Ex. 19c	Study 6	GPU complete-tree constraints
Ex. 19d	Study 6	GPU dense reconstruction test and analysis
Ex. 20	Study 6	Initial GPU dense validation
Ex. 21	Study 7	ANE operand-exponent refinement
Ex. 22	Study 7	Equal-product factorization
Ex. 23	Study 7	Encoded-subnormal discrimination
Ex. 24	Study 7	ANE operand and stack transfer
Ex. 25	Study 7	ANE pretrained-boundary transfer
Ex. 26	Study 7	Longer ANE convolution transfer
Ex. 27a	Study 7	GPU ordinary-weight transfer
Ex. 27b	Study 7	GPU input-axis discrimination
Ex. 28	Study 7	Fresh GPU dense validation
Ex. 29	Study 7	Fresh GPU discriminating pairs
Ex. 30	Study 7	GPU pretrained-weight and configuration tests
Ex. 31a	Study 8	Public ANE reduction replay
Ex. 31b	Study 8	Short ANE reduction characterization
Ex. 31c	Study 8	Longer ANE reduction transfer
Ex. 31d	Study 8	Operation and placement comparisons

Finding the evidence

studies/ contains the eight design documents. experiments/ contains the original experiment directories, with reports, protocols, predictors, saved operands, model packages, outputs, execution records and audits as available. The older controlled and pretrained work shares experiments/all-vs-cpu-and-ne/. Its relevant baseline and Phases 1–4 and 6 are retained. Phases 5 and 7 are outside this selection.

The primary study assignments follow the current manuscript's Appendix F. The controlled FP16 comparison belongs to Study 1, while the quantization controls indexed as Ex. 3a belong to Study 3, even though they share a historical record. Ex. 7 belongs to Study 3, and Ex. 21–24 belong to Study 7. Lettered identifiers distinguish distinct experimental designs within retained numbering slots. Discovery, confirmation, controls and unchanged replay stay with their experiment. Supporting arithmetic analyses and retrospective selection retain their stated roles. The experiment index above links each complete identifier to its question and scoped records. Historical phase and folder names are preserved.

The new links use relative paths. Historical reports and scripts are copied byte for byte and can contain the original absolute workspace paths, host aliases and commands. Those records are provenance, not a promise that every command runs unchanged from a new checkout. In particular, historical runners can write to their original experiment paths or contact remote machines. Use an experiment's reproduction instructions to plan a separate working run. This migration performed no new inference and did not retune any predictor.

Checking this copy

The original local migration contained 60,088 files totaling 36,686,540,505 bytes (34.17 GiB) before this README, the eight designs and migration metadata. This collection omits the ImageNetV2 archive and 10,000 extracted images. The remaining scientific payload contains 50,087 files totaling 34,177,830,672 bytes (31.83 GiB). SHA-256 comparisons verified the original copied bytes against the sources. A second source pass verified that the selected source bytes and modification times remained unchanged. The original local copy preserved file permission modes and empty package directories.

For a Git checkout, retrieve the large-file contents before checking:

git lfs pull
python3 tools/verify_integrity.py --restore-empty-dirs

For an already populated copy, run the portable integrity check from the repository root with Python 3:

python3 tools/verify_integrity.py

It checks the current payload and documentation against the integrity manifest, including required directories and executable status. Original permission bits remain recorded, while Git preserves only executable versus non-executable status. It also checks relative links in this README and the eight designs. It does not rerun Core ML, validate the scientific conclusions independently, or rewrite any experiment. Because Git does not retain empty directories, --restore-empty-dirs can recreate only the recorded missing directories after a future checkout, before checking.

The migration record describes the selection and verification. The study and experiment catalog gives each experiment identifier one primary study and links it to its physical records. Study references and record links express relevance or dependence without adding study membership. Records can support several identifiers, and an identifier can use several records. These links do not count the same observation as independent evidence again. Repeated arrays, archives, failed hypotheses and independent audits within the selected records are retained. Environment directories and generated Python caches were excluded as declared in the migration record. Historical inventories may mention those omitted caches or paths outside this selected collection. The new integrity manifest defines this copy's coverage.

Reading the claims

The CPU, ANE and GPU accounts characterize effective arithmetic in specified configurations and software/hardware stacks. Preferred placement is not a private instruction trace. Frozen prospective tests are distinguished from evidence used to select a rule. Array positions, duplicated filters and repeated executions are not interchangeable counts of independent configurations. The later refinement design identifies the current GPU axis-order rule and the tested ANE configuration extensions.

Environments, rights and release status

Saved-array inspection and this integrity check use ordinary local files. Numerical predictors and inference have experiment-specific dependencies described in the linked records. New Core ML execution requires a compatible Apple Silicon/macOS environment. Original and later environment strata must not be pooled as a controlled hardware-only experiment.

The migration record describes the local copy before its first upload. The original experiment's MIT license is preserved in its own scope. It does not license third-party images, pretrained weights, source material or every later contribution. The collection does not redistribute ImageNetV2 image files. Pretrained packages and retained literature material remain subject to their original terms. Their presence in this collection is not a redistribution clearance or an ACM artifact badge.