sceptre-tuning-ocr

<!-- AI-RULEZ :: GENERATED FILE — DO NOT EDIT Content-Hash: blake3:ad5a2fb97e28a0deefac8858dd0fb2b325625476f642d91546b5f997b79b4c15 Source-Hash: blake3:5c8442fa8b2f9f7d5d68801c4562f6bf03cc6e52237900bbcf720e4b5bedfa5e Schema-Version: v1 --> # Tuning Ocr Record a baseline image, language configuration and JSON output before changing settings. Use `--timings` for CLI stage timings and `--threads` for the worker budget. For large pages, lower `--canvas-size` (default 2560) or cap `--max-megapixels`; this trades recognition quality against detection memory. Change one setting at a time. For rotated scans, enable `--detect-orientation`; adjust `--orientation-probe-canvas-size` and `--orientation-margin` only after inspecting probe results. `--text-threshold` and `--link-threshold` are probabilities between zero and one. Lower detection thresholds may admit noise; higher thresholds can discard real text. Compare absolute correct line/word counts against the source image, confidence and boxes. Do not judge improvements by a ratio over shortened output. Preserve low-confidence lines for review. MCP runtime settings belong to `sceptre mcp` startup; they are not per-call tool arguments.