Lithi Apple ML

Choose the result you need

First decide whether you need validation, typed predictions or benchmark measurements. A prediction request for model digest MODEL-A can return one typed prediction or explicit exception for each of CM-01–CM-03; an incompatible "unknown" value stays rejected. A benchmark needs a workload, profile, device scope and measurement state that can be reviewed separately.

Your agent chooses the qualified model operation. It prepares the permitted model, inputs and review rule for the selected operation. Lithi returns the declared result with its model and input identities. Your agent checks types, coverage and evidence before using it.

Your agent keeps the model question and acceptance rule. The Kit prepares the supported Apple ML request; participating Macs process the bounded inputs, and your agent reviews typed predictions or benchmark evidence.

Model assets, test inputs, evaluation sets, and rubrics belong to the Models & Evaluation information family. A model asset is the input; the operation determines whether the result is validation, prediction, or a benchmark.

The operation is the exact supported capability. Agent guidance does not install dependencies or grant access. The Apple ML pack covers coreml-validate, coreml-predict, coreml-batch-predict, and coreml-benchmark. Resolve the operation's current release, model format, input and output contract, resource profile, and supported execution scope before quoting. A file extension alone does not establish compatibility.

Conceptual task brief: Request typed predictions from model digest MODEL-A for feature rows CM-01–CM-03. The permitted inputs are the qualified model package, its declared feature schema and those rows. Reject incompatible types; do not coerce or guess a prediction. Return one typed prediction or explicit exception per sample, with model and sample identity. Accept only when every expected sample is reconciled and the output types match the declared contract.

The selected Apple ML operation and supported integration map that brief to the permitted request and returned evidence. The brief does not install dependencies or grant access. For a benchmark, include the workload, profile, device scope, measurement state, and evidence needed for acceptance.

Provide an immutable model asset and a feature dataset with stable sample IDs. A Compute job may stage admitted model and data assets, but it does not install arbitrary packages, frameworks, dylibs, shell installers, conversion tools, or customer dependencies on a provider Mac. Model conversion and quantization remain separate tool decisions.

Bind typed inputs to actual evidence

1. Record the model digest, package structure, feature names, tensor shapes, layouts, dtypes, and preprocessing and postprocessing versions.
2. For validation, state the checks and scope. Validation evidence does not become a prediction-quality report.
3. For prediction, keep each sample ID with its typed prediction or explicit exception. Preserve rejected inputs rather than coercing them.
4. For batch prediction, reconcile expected and returned sample IDs. Missing, failed, and excluded items remain separate.
5. For benchmarks, bind workload, model and source digest, runtime/profile, permitted device cohort, cold or warm state, measurement window, counts, and failures.

The fixture helps plan the review; it is not a Core ML package or a recorded native run. An actual result identifies model, profile, sample, prediction or exception, and observed evidence. A confidence value exists only when the model or operation returns one.

Keep validation, prediction, and benchmark claims separate

An allowed compute-unit setting does not tell you whether ANE or GPU hardware was used. Report measured timing with units, population, device cohort, cold or warm state, and preparation scope. Missing or incomparable measurements stay missing. Two correct predictions do not establish accuracy on a larger dataset.

If a model shape, runtime profile, memory requirement, or package member is unsupported, keep the refusal and choose a qualified alternative. If delivery fails after acceptance, recover the existing artifact. A changed model, feature dataset, preprocessing rule, or benchmark profile creates a new request with its own review.

Worked example

Check typed model inputs and prediction evidence

Bind a small qualified model to typed input features using the declared relationship y = 2x + 1: CM-01 (x = 1.0) expects y = 3.0, and CM-02 (x = 2.0) expects y = 5.0. CM-03 (x = "unknown") is rejected as an incompatible type. Record the model and release identity and inspect the prediction structure. These are mathematical expectations, not measured Core ML results.

Use Code & Tests when source or qualified test evidence is the subject and Data & Evaluation when the question is candidate comparison. Continue with model and build asset guidance before preparing a real model request.