[techmore]

Summary

Add chip- and OS-labelled dot plots to the local AI scoreboard and retain the model revision, weight and prompt hashes, kernel, run times and per-run evidence. Add M1 Pro macOS MLX results for the four pinned scoreboard models at 32- and 256-token limits; preserve the installed TensorFold Vulkan 2B results and their long-output divergence.

Show matched-workload ratios of Aurora to macOS median decode and first-token times, and report available historical Vulkan submission, dispatch, barrier and system-monitoring counters. Ratios come from separate OS runs, text differences are flagged, and the counters do not identify per-kernel bottlenecks.

Record the earlier exploratory Metal profiles separately, and document the public, closed and unmerged TensorFold batching candidate and its failed clean-apply check against the installed revision. No Linux runtime, driver or service changes.

Validation

• 24 Mac runs completed with three fresh-process trials per model and workload. Receipt prompt, text and token hashes were independently checked; output differences against available Linux MLX receipts remain visible.
• Seven scoreboard platform, provenance, ratio and counter regression tests pass; regenerated Markdown and SVG are deterministic and both SVGs parse.
• git diff --check passes. Both charts were rendered for a visual check.
• Multi-gigabyte trace bundles remain local because they contain model buffers or system process metadata; synthetic benchmark logs and receipt data are committed.

[techmore]

Superseded by the archived ElementaryOS experiment and the active Aurora M4 repositories.