← product evidence + the ladder to state of the art

Nishi browser vs Chrome — the TOP-20 real-site suite (first honest baseline)

workstream: S21 experiential / S18 graphics — browser chrome-parity ratchetrun: browser-top20-2026-07-162026-07-16

Verdict: The benchmark is now 20 real sites we do not control (operator: one self-authored page proves nothing), re-runnable as one command (bench/nx_top20_parity.ps1) — and its FIRST HONEST LESSON is about the metric itself: the eyeball pass caught Nishi rendering google.com as a BLANK page that still scored 97.3% coarse, because google is mostly white space and a blank white page matches white space. Region-mean parity is GAMEABLE BY EMPTINESS on sparse pages — so google/reddit/ebay/example's high scores are VACUOUS until an ink-weight check lands (the next instrument rung: compare content-pixel mass per side; a side with a fraction of the oracle's ink cannot score parity). What stands after that discipline: wikipedia 90.6% coarse and hackernews 88.5% are REAL (dense text pages, visually confirmed); stackoverflow 81.7 / netflix 80.2 / linkedin 79.6 render real content with real gaps. The suite still earned its keep twice over: run 1 caught the canvas background-propagation bug (a banner div painted google's whole canvas dark; fixed to spec root-chain html->body, gate + /experiential void re-verified), and run 2's eyeball caught the blank-parity metric hole. Named gaps, each a ladder rung: google renders blank despite fetching (layout has 3378px of content — paint/visibility diagnosis queued); 5 sites FETCH-EMPTY (youtube/facebook/instagram/duckduckgo/craigslist); bing + github 0.0% (background images + attribute-selector themes); x/reddit tiny JS shells; amazon blocks headless Chrome (no oracle); ebay's oracle was a bot page.
evidence
BEST REAL RENDER — Wikipedia: 90.6% coarse. Body text, links, infobox image all render; the skin sidebar column collapses; height 115% (margin-collapse). open full size
evidence
hackernews: every story, link, point-count and comment-count renders — but the TABLE MODEL scatters them (ranks isolated, subtexts shoved right by flex-ified table rows). (A human first read the dark region as a wrong canvas; the organ eye measured canvas-delta 23 and was right — it is the post-content area, not the page canvas.) open full size
evidence
stackoverflow: real content in readable flow (questions, sidebar text, survey banner) — but the multi-column layout linearizes, AND raw HTML attributes leak into painted text (the visible `data-s-popover-placement=...` string) — a tokenizer bug with a clean reproducer. open full size
evidence
linkedin: text content renders fully and readably (headings, links, agreements — even menu items Chrome hides). Missing: the two-column split, the illustration (SVG), button shapes. Content parity without layout parity. open full size
evidence
x.com: same class as linkedin — the login shell's full text renders in clean linear flow (97% of the words are there); no giant X logo (SVG), no buttons, no two-column split. open full size
evidence
bing — THE COLUMN-COLLAPSE SPECIMEN: Nishi paints a one-character-per-line vertical strip on gray where Chrome shows a full-bleed photo. Two rungs in one image: a container laid out at near-zero width (every line breaks after one glyph), and background-image unpainted. open full size
evidence
netflix: the same 1-char vertical collapse ("MLB Home Run Derby 2026" spelled downward) on the correctly-propagated near-black canvas; the hero collage is background imagery we don't paint. open full size
evidence
yahoo: the closest to a real portal render — nav, links, Trending, headlines AND photos paint — but images overlap text (no float/position), an ad column collapses to 2-chars-per-line, and the card grid linearizes. open full size
evidence
google — Chrome (left) vs Nishi (right, BLANK): fetched, laid out 3378px, painted nothing visible — and STILL scored 97.3% coarse because google is mostly white space. The image that exposed the metric hole (ink-weight rung) and, earlier, the canvas-propagation bug. Two bugs from one site. open full size
evidence
reddit — BOTH SIDES JUNK, honestly labeled: Chrome's oracle is itself a bot-block page ("You've been blocked by network security"), Nishi shows one orange image block. The 92.1% score is two near-empty pages agreeing; the CHROME-BLOCKED class exists for exactly this. open full size
evidence
github: 0.0% coarse, honestly — its dark theme lives behind attribute selectors (html[data-color-mode]) and hero imagery, neither matched nor painted yet. open full size
evidence
ebay: Chrome got a 354px bot-challenge stub while Nishi fetched real content — the oracle failed, not the engine; scored ratio 347% is meaningless and flagged. open full size
evidence
example.com control: renders correctly; the tall-window Chrome capture centers content vertically so the naive tight score is a capture artifact, noted. open full size

Measured

whatresultmeaning
Suite verdict (renderable 13 of 20, ink-disciplined)REAL parity: wikipedia 90.6 · hackernews 88.5 · stackoverflow 81.7 · netflix 80.2 · linkedin 79.6 · bbc 71.3VACUOUS (blank/sparse-matching, pending ink-weight): google 97.3 · ebay 99.4 · example 98.9 · reddit 92.1 — high numbers on near-empty renders count for NOTHING
Metric hole found by the eyeball passblank Nishi google scored 97.3% coarseregion-mean parity is gameable by emptiness on sparse pages -> NEXT instrument rung: ink-weight (content-pixel mass per side; ink ratio << oracle => BLANK-NISHI class, no parity awarded)
google renders BLANK despite fetchingnishi_h=3378 laid out; zero visible paintpaint/visibility diagnosis = the top engine rung from this run (text color? visibility rules? fold?)
The suite caught a real engine bug on run 1google/reddit coarse 0.0% -> 97.3%/92.1%canvas bg propagated from ANY first non-white block (a banner div); spec fix = root chain only (html->body); /experiential void re-verified dark (luma 17) after the fix
FETCH-EMPTY (5)youtube, facebook, instagram, duckduckgo, craigslistnishi_fetch returns nothing — the fetch-reliability rung (TLS fingerprinting / redirect chains / HTTP-2-only to diagnose)
Oracle failures (2)amazon (Chrome headless blocked), ebay (Chrome got a 354px bot page while Nishi got real content)the CHROME-BLOCKED class exists so two blank pages can never again score 100% parity (run 1 did exactly that)
Named paint gaps from the suitebackground-image (bing full-bleed photo), attribute-selector themes (github data-color-mode), JS shells (x, reddit skeletons)each a ladder rung with its measuring site
Height ratios (renderable)wikipedia 115 · bbc 112 · netflix 120 · hackernews 134 · stackoverflow 88 · linkedin 77 · github 51 · google 42over-100s = margin-collapse rung; under-50s = JS-populated content we honestly don't have
THE ORGAN EYE (nx_parity_judge) — the suite now sees these failures WITHOUT a humangate GREEN 6/6 against the human truth set from this very runautonomous classes per pair: BLANK-NISHI (google, netflix, ebay) · NO-ORACLE-CONTENT (bing) · CANVAS-OFF (github) · SPARSE-NISHI (linkedin, x, stackoverflow) · CONTENT-OK (wikipedia, hackernews, yahoo, example, reddit)
The organ out-judged the human oncehackernews "dark canvas" human read -> canvas-delta 23 (fine)the dark region is the post-content area, not the page canvas — the eyeball was wrong, the measurement was right; caption corrected
Common failures, machine-countable now1 blank-paint class (google) · 1-char column collapse (bing/netflix strips) · canvas themes (github) · decoration/column loss (linkedin/x/so) · table scatter (HN) · attr-leak (SO) · overlap (yahoo/wiki)judge v2 rungs: attr-leak + overlap need a layout-tree dump from the shot lane (pixel-only can't see them)
READING RECALL — the number that ends "it looks fine" (OCR tier: the same neutral instrument reads BOTH renders; sovereign scorer = truth-token recall, permille)suite mean ~384 of 1000: the Nishi browser preserves ~38 percent of what a Chrome reader can readper site: x 848 · example 789 (even the trivial control loses 21 percent to font quality) · linkedin 772 · wikipedia 766 (our best real page: "bad but legible") · bbc 543 · hackernews 400 (the table scatter halves readability) · stackoverflow 116 (the CONTENT-OK class hid an 88 percent reading loss) · github 108 · yahoo 84 · google/reddit/netflix/ebay 0 (blank) — instrument bench/nx_ocr_read.ps1 (Windows.Media.Ocr, typography-arc pattern) + organ _offc/nx_ocr_recall.elf (identity self-test 1000)
Precision tells its own storywikipedia precision 507 vs recall 766we paint MORE tokens than Chrome shows (duplicated/garbled runs + menu text Chrome hides) — badly-placed text is as measurable as missing text
RepeatabilityONE command: bench/nx_top20_parity.ps1 — self-judging + read-scoredChrome oracle + Nishi record + sovereign instruments + nx_parity_judge class + OCR reading-recall per site; results committed with the run

Machine detail: api.json · related: /evidence/nishi-browser