VRAM Lab MAC-CONTEXT evidence, 2026-09-10.
Nine primary runs: 2K/4K/8K x three. Scouts are separate, including a censored 16K attempt.
Redaction replaces the exact private package/repository path (both Unicode forms) and pmset battery registry IDs.
All numeric observations, generated synthetic content and timestamps are retained.
file_sha256 inside result.json binds ORIGINAL PRIVATE raw bytes. manifest.json binds each public derivative separately.
Shared fixture/tokenizer files are in inputs/ instead of duplicated inside each run.
The interrupted 16K attempt has no completed TTFT/throughput or time(1) peak. Its zero-byte time.txt is retained as missing data.
No claim of cold OS cache, fresh macOS installation, second computer, model quality or dedicated VRAM.
