The raw files behind the numbers
These files support the performance figures on the engineering page. The August files are direct benchmark harness output. The September thumbnail record combines four measured service runs and includes the sample, timing boundaries, output checks, and hardware.
The August scan and query benchmarks cover 141,696 indexed files. The September thumbnail benchmark uses a 768-asset sample and a database containing 152,955 records. These are different workloads; the thumbnail result is not a whole-library scan time.
The thumbnail comparison alternated before/after/after/before with warm NAS and OS caches. It includes worker startup, file reads, rendering, image writes, database updates, and queue completion. Scanning, tagging, STEP models, and the interface were excluded. All runs read source files only and wrote to scratch profiles. Timings are in milliseconds; the personal library and its file paths are not published.