mirror of https://github.com/teamchong/pxpipe.git synced 2026-07-22 02:02:51 +02:00

Files

T

teamchong 81550f690e feat(render): pack reflow across newlines + grayscale atlas + L1/L2 eval harness

- wrapLines packs continuously; NL_SENTINEL stays as an inline marker, never breaks a row
- atlas-gray.ts: anti-aliased grayscale variant (gated via RenderStyle.aa, unused in production)
- gen-atlas.ts: optional ATLAS_GRAY=1 emission
- eval/: standalone Anthropic-client L1 OCR fidelity + L2 session replay harness, claude -p backed
- README: reflect packed reflow + measured Opus 4.7 numbers

2026-05-22 20:48:40 -04:00

1.1 KiB

Raw Permalink Blame History

Reflow Eval — Combined Summary Report

Generated: 2026-05-22T03:18:39.382Z
Model: claude-sonnet-4-5
Levels run: L1, L2

Overview

L1: OCR Fidelity

	Baseline	Reflow	Δ
Mean char accuracy	97.73%	80.64%	-17.09pp
Macro accuracy	97.18%	75.65%	-21.54pp
Image savings	—	0.0%

Full L1 report: l1-report.md

L2: Session Replay

	Value
Mean judge score	51.0%
Pass rate (≥ 0.75)	30.0%
Image savings	50.0%

Full L2 report: l2-report.md

Shipping Gate

Reflow is safe to ship if ALL of the following hold:

L1 mean accuracy delta ≥ −2pp (reflow OCR not materially worse)
L1 macro accuracy ≥ 95% (overall character fidelity high)
L2 mean judge score ≥ 0.80 (task comprehension preserved)
L2 pass rate ≥ 80% (failures are rare outliers)

If any gate fails, investigate the failing sessions/blocks before shipping.

How to Interpret

See README.md for full guidance on running and interpreting each level.

1.1 KiB Raw Permalink Blame History Unescape Escape