How close is vellora to Chrome?

We printed the same 9 documents with Chrome and with vellora and measured where every word landed. Drag the handle to compare one of them yourself.

Chromevellora

An open-source invoice template, page 1. The same HTML printed by Chrome for Testing 155 (left) and vellora 0.3.0 (right).

The short answer

On 8 of 9 documents, every word lands within 1 mm of where Chrome puts it. All 9 have the same number of pages as Chrome's PDF. On the ninth, the Postmark receipt, nearly all of its words are within 1 mm.

These are invoices, receipts, bank slips, a letter and a long text, the kind of documents vellora is built for. They do not tell you how an arbitrary web page will print. For templates that need a real browser, see When you need Chrome.

See every document

Drag the line on each one. Left is Chrome, right is vellora 0.3.0, and both come from the same HTML file.

Chromevellora

Postmark’s open-source receipt template, page 1. This is the one document that is not perfect yet; see below.

Chromevellora

A three-page Portuguese invoice with made-up data, page 1.

Chromevellora

A DANFE, the printed form of a Brazilian electronic invoice, with made-up data. Its many small bordered boxes make any drift easy to see.

All nine documents, with our previous release for comparison

Each image shows the same HTML three times. Left: Chrome. Middle: our previous release, vellora 0.2.0. Right: vellora 0.3.0. The document data is made up. Where one PDF has a page the others don't, you'll see a "no page" placeholder.

Invoice

Receipt

Boleto (bank slip)

Boleto, FEBRABAN layout

DANFE (NF-e invoice)

Notification letter

Long pt-BR text (stress)

Postmark receipt (OSS template)

_{Postmark receipt (OSS template) is adapted from ActiveCampaign/postmark-templates (commit fa73527), MIT licence, Copyright (c) 2015 Wildbit; see NOTICE.}

Sparksuite simple invoice (OSS template)

_{Sparksuite simple invoice (OSS template) is adapted from sparksuite/simple-html-invoice-template (commit 7d74fe3), MIT licence, Copyright (c) 2021 Sparksuite; see NOTICE.}

Where it is not perfect yet

On the Postmark receipt, nearly all words land within 1 mm of Chrome's. A couple of words have no match, because Chrome draws a text shadow on the header and vellora does not draw text shadows. A few other words land more than 1 mm away. It is the first slider under See every document.

Test-suite and release-check details

• oss-postmark-receipt: 97.2% of words within 1 mm (vellora 0.2.0 (published): 40.0%); p90 error 0.14 mm.
• WPT: 4,553 of 8,047 reftests pass (56.58%); 4,028 (50.06%) pass without any suspect-pass flag. 2,608 still fail and 877 are rejected as outside the strict subset.
• Merge gate: verdict PASS on this run (strict contract, --contract), 0 gating items. Every condition is met. What each condition still lets through is listed with it.
  • (1) pin view (--main): no item worse than the published package unless the gate classifies it (class a: the old pass was false; class b: carried in the debt ledger): met. Passed with the published engine and not matched in 0.3.0: 6 WPT tests. Class a (the old pass was false, or Chrome fails it too): css/CSS2/floats/float-nowrap-8.html, css/css-page/page-name-flex-001-print.html, css/css-page/page-name-flex-002-print.html, css/css-page/page-name-flex-004-print.html, css/css-text/hanging-punctuation/hanging-punctuation-first-002.html. Class b (recorded in the debt ledger, not fixed): pass in 0.3.0 with a suspect-pass flag, which the gate scores as worse than the published package's clean pass: css/css-break/flexbox/flex-fragmented-with-float-descendant-001.html (flag ink-displaced-vs-chrome; raw Blitz: fail). 10 tests that failed before are now rejected by the strict subset.
  • (2) every fixture's median error and share within 1 mm at least as good as the published package's: met.
  • (3) every fixture's page count equal to Chrome's: met.
  • (4) step-to-step ratchet: no regression against the previous gated step, and the head view (against the chain's first step) passes: met. Head view: PASS; 1 test carried in the debt ledger (css/css-break/flexbox/flex-fragmented-with-float-descendant-001.html); 5 items accepted as class a, with the evidence in the acceptance file.

The numbers

The headline figures, then the charts. Lower is better on the first chart, higher is better on the second. The table has every document.

• 8 of 9 fixtures have every word within 1 mm of its position in Chrome for Testing 155.0.8059.85's PDF (share of word positions within 1 mm = 100%). vellora 0.2.0 (published): 0 of 9.
• 0.21 mm or less median word-position error on every one of the 9 fixtures (largest: danfe). vellora 0.2.0 (published): up to 106.22 mm (text-stress-ptbr).
• 9 of 9 fixtures print the same number of pages as Chrome. vellora 0.2.0 (published): 6 of 9.

Each cell reads vellora 0.2.0 (published) → vellora 0.3.0. Reference: Chrome for Testing 155.0.8059.85.

Fixture	Chrome words	Median error (mm)	p90 error (mm)	Words within 1 mm	Chrome words unpaired	Pages: 0.2.0 / 0.3.0 / Chrome
Invoice (invoice)	541	1.46 → 0.10	1.73 → 0.28	36.0% → 100.0%	0 → 0	3 / 3 / 3
Receipt (receipt)	151	5.18 → 0.05	77.00 → 0.11	7.9% → 100.0%	0 → 0	2 / 2 / 2
Boleto (bank slip) (boleto)	297	1.77 → 0.12	14.09 → 0.20	32.0% → 100.0%	0 → 0	1 / 1 / 1
Boleto, FEBRABAN layout (boleto-febraban)	315	80.70 → 0.08	82.32 → 0.28	15.6% → 100.0%	8 → 0	1 / 1 / 1
DANFE (NF-e invoice) (danfe)	630	17.68 → 0.21	37.17 → 0.27	8.9% → 100.0%	1 → 0	2 / 1 / 1
Notification letter (notification)	248	0.41 → 0.01	1.02 → 0.09	86.7% → 100.0%	4 → 0	1 / 1 / 1
Long pt-BR text (stress) (text-stress-ptbr)	1,833	106.22 → 0.02	221.05 → 0.27	1.4% → 100.0%	7 → 0	8 / 7 / 7
Postmark receipt (OSS template) (oss-postmark-receipt)	180	0.66 → 0.13	18.19 → 0.14	40.0% → 97.2%	39 → 2	1 / 2 / 2
Sparksuite simple invoice (OSS template) (oss-sparksuite-invoice)	46	9.87 → 0.00	10.95 → 0.00	0.0% → 100.0%	0 → 0	1 / 1 / 1

Check your own template

Our documents are not your documents. To see how your template does, keep one PDF from your current pipeline and let vellora doctor compare it with vellora's output, page by page:

vellora doctor invoice.html --pixel-diff --reference-pdf old-output.pdf --out artifacts --json

It writes page images and red diff overlays to artifacts/visual. It needs poppler's pdftoppm and @vellora/cli 0.3.2 or newer. All the options, and how to send a template to Chrome instead, are in When you need Chrome.

How we measured

Chrome is the reference, so its own error is zero. Both engines get the same HTML file for each document, and we compare where each word starts on the page. The full story of what we changed to get here is in the engineering note.

Method

The 0.3.0 numbers were measured on commit aa41259, whose crates/, packages/, Cargo.toml and Cargo.lock are identical to the v0.3.0 release commit c40b3c2 (git diff --stat aa41259 c40b3c2 -- crates packages Cargo.toml Cargo.lock is empty). Every number on this page is generated from one manifest, docs/assets/fidelity/manifest.json. The generator rewrites the tables, so they cannot drift from it.

• Reference renderer. Chrome for Testing, headless print-to-PDF, with the harness's single reference setup (fidelity/lib/chrome-reference.mjs). It maps the fixtures' font families to the same bundled Liberation font files vellora uses, and every fixture sets its own @page box. Chrome is the reference by definition: its own error is 0.
• Inputs. Both engines get the same native.html for each fixture, byte for byte. The manifest records that both sides were scored against the same Chrome word geometry (sameChromeGeometry).
• Word positions. PyMuPDF extracts every word of each PDF with its pen origin (x) and baseline (y). Text is normalised (NFC, ligatures folded, whitespace removed), and pages are stacked into one document coordinate.
• Pairing. The k-th occurrence of a word in reading order is paired with the k-th occurrence of the same word in Chrome's PDF. A Chrome word with no counterpart is unpaired.
• Median error (mm). The median distance between paired words. Lower is better.
• Share within 1 mm. Paired words at most 1 mm from Chrome's position, divided by all of Chrome's words. An unpaired word counts as a miss, so a lost word cannot improve this number.
• Page count. Each PDF's page count, compared with Chrome's.
• WPT. WPT reftests are run through the same pipeline, printed to PDF and compared with the reference page. The pass rate measures coverage of the CSS test subset the harness runs. It is not a document-quality score, and Chrome itself does not pass every test through this harness.
• Oracle. fidelity/oracles/doc-aggregate.mjs is the committed document-level oracle. The same numbers feed the merge gate's strict exit contract.
• Paint probe (report-only). fidelity/oracles/paint-probe.mjs reads the vector paint of both PDFs: dashed and dotted border segments, text-decoration underlines, and the colour of the page canvas. It never gates. Word positions cannot see paint, so this is where a solid border that should be dashed, or a missing link underline, shows up.

Machine, versions and commands

Released in vellora 0.3.0

The measured run passed the strict merge gate (verdict PASS, --contract). INDICATIVE: local macOS measurement. The authoritative run is pinned Linux CI, which is not wired yet; these numbers will be regenerated from it when it runs. Before is vellora@0.2.0, the previous release.

• Status: PASS (gate report verdict PASS (--contract)). INDICATIVE: local macOS measurement. The authoritative run is pinned Linux CI, which is not wired yet; these numbers will be regenerated from it when it runs.
• Reference: Google Chrome for Testing 155.0.8059.85, headless print-to-PDF (fidelity/lib/chrome-reference.mjs).
• Before: vellora 0.2.0 (published), tag v0.2.0, commit 5810062, Blitz 0.3.0-alpha.5 line, git 2003ba5. Scored 2026-10-27T22:36:58.971Z.
• After: vellora 0.3.0, measured on phasec/07-tip, commit aa41259, Blitz 0.3.0-beta.4, crates.io. Run 20261028T110750Z (2026-10-28T11:07:50Z).
• Environment: darwin 27.0.0 arm64 (macOS 27.0.1), Apple M5 Pro, Node v22.23.2, rustc 1.96.0 (ac68faa20 2026-05-25), PyMuPDF 1.28.2, pdftoppm version 26.06.0.
• Word positions: fidelity/oracles/py/words.py (PyMuPDF rawdict: pen origin x, baseline y); pairing: k-th occurrence of a text key in reading order (page, baseline, x) ↔ k-th in Chrome's; quantiles: nearest rank over paired words (ceil(q*n)-th smallest).
• Run command: node fidelity/run-all.mjs --gate --main --contract --raw --chrome '<repo>/fidelity/.cache/chrome/chrome/mac_arm-155.0.8059.85/chrome-mac-arm64/Google Chrome for Testing.app/Contents/MacOS/Google Chrome for Testing' --baseline fidelity/baselines/port/11-phase-c --chrome-from ~/Repositories/vellora-fidelity-runs/run-all-10-tip-20261023T004536Z-blitz-0.3.0-beta.4-vellora-a293cf5/wpt/results.json --tools-image vellora-fidelity-tools:baseline-beta4 --accept fidelity/gate/snapshots/2026-10-28/c07/accepted.json --step 16-phase-c-07-tip --out fidelity/results/run-all/c07-tip
• Gate: node fidelity/gate/run.mjs --baseline fidelity/baselines/port/11-phase-c --run ~/Repositories/vellora-fidelity-runs/run-all-c07-tip-20261028T110750Z-blitz-0.3.0-beta.4-vellora-aa41259 --head fidelity/baselines/port/00-start --pin fidelity/baselines/pin --layout-from ~/Repositories/vellora-fidelity-runs/run-all-c07-tip-20261028T110750Z-blitz-0.3.0-beta.4-vellora-aa41259 --debt fidelity/baselines/port/DEBT.json --step 16-phase-c-07-tip --accept fidelity/gate/snapshots/2026-10-28/c07/accepted.json --doc-aggregate ~/Repositories/vellora-fidelity-runs/run-all-c07-tip-20261028T110750Z-blitz-0.3.0-beta.4-vellora-aa41259/doc-aggregate/summary.json --contract --out ~/Repositories/vellora-fidelity-runs/c07-tip/regate-contract → PASS (2026-10-28T11:19:27.811Z).
• Paint probe (report-only): node fidelity/oracles/paint-probe.mjs --run <repo>/fidelity/results/run-all/c07-tip --label run-all; published package: node fidelity/oracles/paint-probe.mjs --run ~/Repositories/vellora-fidelity-runs/run-all-c07-tip-20261028T110750Z-blitz-0.3.0-beta.4-vellora-aa41259 --subject ~/Repositories/vellora-fidelity-runs/doc-aggregate-20261027/pin-pdfs-v0.2.0 --label pin-v0.2.0 --out ~/Repositories/vellora-fidelity-runs/docs-c07/paint-probe-pin-v0.2.0.json.
• Page images: node fidelity/report/visual-compare.mjs --run <RUN> --pin-pdfs <doc-aggregate-pin.mjs --pdf-out dir> --out <compare dir> (pdftoppm -r 55), composed by fidelity/report/py/compose.py
• This page: node fidelity/report/build-site-evidence.mjs --compare ~/Repositories/vellora-fidelity-runs/docs-c07/visual-compare --run ~/Repositories/vellora-fidelity-runs/run-all-c07-tip-20261028T110750Z-blitz-0.3.0-beta.4-vellora-aa41259 --gate ~/Repositories/vellora-fidelity-runs/run-all-c07-tip-20261028T110750Z-blitz-0.3.0-beta.4-vellora-aa41259/gate-contract/gate-report.json --paint-before <repo>/fidelity/gate/snapshots/2026-10-28/c07-docs/paint-probe-pin-v0.2.0.json --pin-key pin --new-key c07 --release 0.3.0 --update-docs
• Raw values: manifest.json.

Paint: dashed borders, underlines and page colour

Vector paint read from each PDF (PyMuPDF get_drawings and text spans): dashed/dotted border segments, text-decoration underlines, and whether the page canvas colour matches Chrome's. Report-only: no gate rule reads it. Each cell reads vellora 0.2.0 (published) → vellora 0.3.0. Both are probed against this run's Chrome PDFs (the published package's PDFs through paint-probe.mjs --subject; the generator checks that Chrome's side of both probes agrees).

Fixture	Dashed/dotted border segments, vellora / Chrome	Underlines, vellora / Chrome	Pages whose canvas colour differs from Chrome's	Probe flags
Boleto (bank slip) (boleto)	0 / 135 → 135 / 135	0 / 0 → 0 / 0	none → none	dashes-missing → none
Boleto, FEBRABAN layout (boleto-febraban)	0 / 152 → 152 / 152	0 / 0 → 0 / 0	none → none	dashes-missing → none
DANFE (NF-e invoice) (danfe)	0 / 2,052 → 2,111 / 2,052 (59 more than Chrome)	0 / 0 → 0 / 0	none → none	dashes-missing → none
Postmark receipt (OSS template) (oss-postmark-receipt)	0 / 136 → 136 / 136	0 / 5 → 5 / 5	pages 1, 2 → none	canvas, dashes-missing, underlines-missing → none

The other 5 fixtures have no dashed borders, underlines or canvas colour in either PDF. A count equal to Chrome's says the segments exist, not that each one sits where Chrome's does. Method: fidelity/oracles/py/paint.py (PyMuPDF get_drawings + text spans; docstring); pymupdf 1.28.2.

CSS test-suite coverage (WPT)

WPT reftests passed through the vellora pipeline, printed to PDF and compared with the reference page: coverage of the CSS test subset the harness runs, not a document-quality score.

	Blitz 0.3.0-alpha.5 engine of vellora 0.2.0 (baseline commit with layout instrumentation)	vellora 0.3.0 (Blitz 0.3.0-beta.4, crates.io)	Chrome/155.0.8059.85 through the same harness
Reftests	8,047	8,047	8,046
Pass	2,578 (32.04%)	4,553 (56.58%)	7,258 (90.21%)
Pass without suspect-pass flags	1,988 (24.70%)	4,028 (50.06%)	7,258 (90.21%)
Unsupported (rejected by the strict subset)	869	877	0

Commits: 2c66544 (2026-10-21T05:41:54Z), aa41259 (2026-10-28T11:07:50Z); WPT 9d8aceec. Chrome's outcomes were reused from an earlier run with the same WPT commit and comparison settings (--chrome-from). The 0.2.0 column was measured on a baseline commit that adds layout instrumentation to the 0.2.0 engine; its render-path equivalence to the tag is not separately verified.

Does it match on Linux?

On a pre-release 0.3.0 tree, vellora wrote the same PDF bytes on Apple Silicon Macs and arm64 Linux. Chrome did not: on Linux it moved some words slightly and drew the footer in another font. So we don't publish Linux numbers yet. We have not tested x86_64 yet.

What we ran on Linux

The same gated run was repeated on Linux arm64 in a local Docker container (Debian 13, Chrome for Testing 155.0.8059.85 linux-arm64, 2026-10-28, run 20261028T113714Z, on the Phase C 06 tree: the strict contract PASS before the step-07 fixes). vellora's PDFs are byte-identical on macOS arm64 and Linux arm64: 94 of 98 PDFs outside WPT, including all 9 fixtures, the corpus and the PDF/A renders; the other 4 are the pdf-probe canary mutations, which also differ between two macOS runs. WPT on Linux: 4545 passes of 8047, 0 per-test changes against macOS. The Chrome reference is not the same on Linux: words move by up to 0.123 pt, and the @page footer is drawn in another fallback font (DejaVu Sans instead of Times in three of the four legacy fixtures). The committed baselines were measured with macOS Chrome, so the gate refuses the Linux comparison. Gating on Linux needs baselines measured with Linux Chrome (not done yet). x86_64 was not measured. Record: fidelity/gate/snapshots/2026-10-28/linux-arm64.

Reproduce the measurement

The harness is in the repository (fidelity/). Its setup is in fidelity/README.md ("Run everything"): Node 22, the native addon, a Python venv with PyMuPDF, poppler, and the oracle tools image. Then:

# 1. The reference browser: Chrome for Testing at the pinned version.
node fidelity/wpt/install-chrome.mjs

# 2. A full measured run. The doc-aggregate step writes <run>/doc-aggregate/summary.json.
#    --wpt-chrome runs Chrome's WPT outcomes live; --chrome-from <earlier run>/wpt/results.json
#    reuses them (same WPT commit and comparison settings). --contract makes the gate report the
#    strict merge contract, which is the only report that can turn the evidence from DRAFT to PASS.
#    The run's paint-probe step writes <run>/paint-probe/summary.json (report-only).
node fidelity/run-all.mjs --gate --main --contract --chrome <Chrome for Testing> \
  --baseline fidelity/baselines/port/12-phase-c-final --wpt-chrome \
  --tools-image vellora-fidelity-tools:local --out fidelity/results/run-all/<id>

# 3. The published engine, scored on the same inputs. Build 0.2.0 in a fresh directory: a copied
#    target/ is not rebuilt by cargo.
mkdir -p /tmp/vellora-0.2.0 && git archive 5810062 | tar -x -C /tmp/vellora-0.2.0
(cd /tmp/vellora-0.2.0 && npm ci && npm run build)
node fidelity/oracles/doc-aggregate-pin.mjs --engine /tmp/vellora-0.2.0 \
  --engine-commit 5810062cd8a30d3721d3ffbf17c08bf01e4dc364 \
  --engine-blitz 2003ba5fc2d8931da9213ef26a813b0d24c83ca6 \
  --run fidelity/results/run-all/<id> --pdf-out /tmp/pin-pdfs --out /tmp/pin-doc.json
# The published engine's paint, probed against the same run's Chrome PDFs.
node fidelity/oracles/paint-probe.mjs --run fidelity/results/run-all/<id> \
  --subject /tmp/pin-pdfs --label pin-v0.2.0 --out /tmp/pin-paint.json

# 4. Page images, then this page's charts, tables and manifest.
node fidelity/report/visual-compare.mjs --run fidelity/results/run-all/<id> \
  --pin-pdfs /tmp/pin-pdfs --pin-doc /tmp/pin-doc.json --out /tmp/visual-compare
node fidelity/report/build-site-evidence.mjs --compare /tmp/visual-compare \
  --run fidelity/results/run-all/<id> --pin-doc /tmp/pin-doc.json \
  --gate fidelity/results/run-all/<id>/gate/gate-report.json \
  --paint-before /tmp/pin-paint.json --update-docs
node fidelity/report/build-site-evidence.mjs --check

Step 3 scores the published engine on your run's inputs. Without it, the committed pin reference (fidelity/baselines/pin/doc-aggregate.json) is used, and the gate refuses the comparison under --main if any fixture's native.html differs from the one the pin was scored on. The generator stamps the evidence DRAFT unless the gate report it is given has the verdict PASS.