jheapsandClaude Opus 5
6fa4532781
Contrast is the one check I've got wrong twice, both times by being clever about backgrounds. The first version read only background-color, so text on any of the gradient-painted sheets was measured against transparency and the dealer puck's dark "D" was reported at 1.2:1 against nothing. The second over-corrected and treated low-alpha gradient stops as opaque, scoring white-on-dark at 1.0:1. So this one is built to be incapable of false positives rather than exact. Backgrounds composite through alpha; where a gradient is involved every colour stop is tried and the most flattering one used. A finding means the text fails even under the reading most generous to it. It's checked against known answers before being believed: black on white at 21:1, white on white at 1:1, #767676 on white at the 4.54:1 AA boundary, white over a dark gradient, dark text on that same gradient, white over a 6% white wash on a near-black page. Nine cases, all correct — and building them found a real bug in the check, which walked past an opaque gradient and offered the page background as a candidate. That made it too lenient, which is as bad as the old false positives. What it found: Both blind badges failed against their own gradients, in 9px bold text a dealer reads at a glance across a table. At the worst stop: small blind 3.52:1, big blind 2.15:1. The small blind keeps its blue and darkens to 4.77:1. The big blind flips to dark text on brighter gold, 5.84:1 — how a gold chip is marked anyway, and it makes the two badges far easier to tell apart at speed. Quick Split's footnote names a control the dealer is told to reach for, and that name inherited the muted grey of the surrounding aside: 3.39:1. Emphasised text carrying an instruction gets its own colour now. The felt watermark came back at 1.19:1 and is meant to. It's decoration; holding it to a reading threshold would mean making it not a watermark. It's aria-hidden now, which says the same thing where it also does some good, and the audit skips aria-hidden subtrees. 32 contrast passes over the table, all five overlays and Quick Split at five viewports: zero findings, zero unmeasured. Layout audit still clean. 137 tests pass. Co-Authored-By: Claude Opus 5 <[email protected]> Claude-Session: https://claude.ai/code/session_015bnp7xKUPFFSpwkYNFaPgb