Performance / correctness / evidence

How fast is a frame?

Measure the work between rendered pixels and usable data. Compare image size, exact regions, codecs, first reads, and repeat reads. Every extraction is checked against known values.

Correctness checks

Awaiting run

Correct values, rejected stale positions, and expected decode/OCR work. A fast wrong value fails.

Identical frame · reader p95

—

Matching bytes and recipe can bypass PNG decode, crop encoding, and recognition.

Real Chrome screenshot time

Measure in extension

Open “Measure capture performance” in the sidebar. This web suite measures image processing and OCR separately.

Open Capture setup →

The same data, different paths

Five measured samples per case, plus an excluded warmup. Cold OCR is one first read. p95 uses nearest rank; these are exploratory measurements.

render → locate → capture → decode → read → validate

Test each boundary. Skip unchanged work. Keep source and identity checks when pixels move or disappear.

  • Full viewport and exact region; 1× and 2× pixels.
  • PNG compression, JPEG, and lossless WebP.
  • Unchanged, missing, covered, moved, and changed fields.
  • Actual browser codecs and local OCR; zero remote model calls.

Browser PNG overhead

Runs the extension’s actual image module on a fixed synthetic PNG in this browser. Full-frame processing uses the PNG header; region processing decodes and encodes once.

Delivery and display

Server work, response delivery, JSON parsing, and UI settling are separate from a single field read.

This suite starts from synthetic rendered pixels. It does not establish performance or accuracy on arbitrary websites, authenticated sessions, or remote models. The live extension report measures already-open-page capture; initial website loading and region discovery still need a site-specific run. CLI: pnpm bench --samples 15 exports JSON, CSV, and a readable report.