← product evidence + the ladder to state of the art
Nishi browser vs Chrome — visual parity benchmark (first honest run)
workstream: S21 experiential / S18 graphics — browser chrome-parity ratchetrun: browser-parity-2026-07-162026-07-16
Verdict: NOT YET MATCHED — and that is the honest baseline. Layout height: /experiential 64% of Chrome, Wikipedia 111%. Paint parity: 95.8%/81.7% coarse, 56.5%/45.1% tight. Two root bugs found by this benchmark were fixed and gate-proven today (3x row bloat from a fractional font-size parse; white void below dark pages). The instrument itself is gate-proven (identical=100%, half=50%, oracle-vs-itself=100.0% ceiling).



Measured
| what | result | meaning |
|---|---|---|
| /experiential layout height | 1418px vs Chrome 2227px (64%) | We now render DENSER than Chrome (missing padding/margin shorthands + line-height); before today's font-size fix we were 206% (4580px) |
| /experiential paint parity | 95.8% coarse (16x12, tol 32) · 56.5% tight (64x48, tol 8) | Coarse = structure/colors match; tight = font raster + spacing differences dominate |
| Wikipedia layout height | 8428px vs Chrome 7582px (111%) | Within 11% of Chrome on a real complex article |
| Wikipedia paint parity | 81.7% coarse · 45.1% tight | Dominant delta = inline link flow (measured: layout reserves bitmap-font widths, paint draws narrower vector font) |
| Calibration ceiling | Chrome-vs-itself = 100.0% | The instrument is capture-stable; every point below 100 is real difference |
| Sensitivity liar-killer | First-screen parity scored the PRE-fix browser HIGHER (96.3 vs 95.8) | Proved single-axis parity lies; the benchmark is multi-axis by law (height + paint) |
| Root fix 1 (today) | font-size "13.5px" parsed as 135px -> 3x-tall rows | Fixed at the parser; nx_css_fontsize_frac_gate GREEN 4/4; page height 4580 -> 1418 |
| Root fix 2 (today) | white void below dark pages | CSS background propagation; frame-7 luminance 236 -> 15; no regression (bg-cascade gate GREEN) |
| Usability study grade | FAIL 25% -> CONDITIONAL 75%, 0 major findings | NN/g task-based grading (nx_ux_study_gate); PASS still refused while the inline-flow overlap remains |
| Oracles | Chrome + Edge captured; Firefox NOT INSTALLED on this machine | Edge output byte-class identical to Chrome (same engine); Firefox pending install |
Machine detail: api.json · related: /experiential