# Source and modification notice

Source: john-rocky/apple-silicon-llm-bench, revision bd5feaf9fb1a5539046bb30789223d628d39d622, session 2026-08-17-mac-litert-v0160, original MLX JSONL lines 2–3. Retrieval: 18 October 2026. Copyright (c) 2026 john-rocky and ios-llm-benchmark contributors. Preserve the full unchanged MIT notice in public-assets/SOURCE-LICENSE.txt and the direct source citations in the selection.

Modifications: Codex for Nathan Fennel, 18 October 2026. Newly authored minimal factual selection of two source-reported warm MLX token counts/rates and declared controls with original explanatory qualifications. No source execution, remeasurement, full-dataset copy, generated text, source code, model weights, tokenizers, binaries, images, logos or raw source captures are distributed.

The root MIT notice expressly covers software and associated documentation. It does not independently establish a separate benchmark-database grant or upstream/model/generated-output rights. Acceptance is confined to this tiny attributed factual commentary and the supplied notice/license; do not use it to export the whole source or normalized dataset.

Both records report 128 generated tokens and a length finish. The rate is source-reported runtime decode, with raw native generateTime and exact upstream timer/EOS internals unavailable. Never present count/rate inversion as an observed duration, or chunk-gap/end-to-end timing as the native denominator. Keep all mandatory_qualifications with any view or standalone export. Exact executed input, loaded model/runtime bytes, full output parity, answer quality, sustained thermal behavior and statistical independence are unverified.

No LiteRT rate, cross-arm ratio, hardware rank, energy/per-GPU/correct-answer allocation, new count credit or deployment permission follows. Source-content acceptance and exact UI/export/application/release review remain separate.