The recorded run is the saved evidence. The numbers below come from the local CLI run of 27 October 2026, with the full Playwright trail. You can also run a fresh Beam repair from this page through a server-side proxy, or open the bundled two-page demo. Those runs are verified in your browser and do not change the recorded result. The interactive app still runs locally with Node.js and Chromium.

Verified accessibility repair

Six findings in. Zero out. Every edit checked.

Patchproof sent two HTML pages and their axe-core findings to Beam. Beam returned exact search/replace edits. Playwright applied them to a copy and rescanned it, and the copy kept only the edits that measurably removed findings without adding new ones.

Run a fresh Beam repair

Choose one HTML file to repair, or review the recorded five-mode benchmark and run the bundled demo at a chosen effort. Axe-core runs in this tab. A candidate is kept only if the rescanned copy removes targeted findings without adding new ones. The Reflection key stays in the server-side proxy and is never sent to the browser.

Upload one complete .html file. Scripts, iframes, event handlers, script URLs, and meta refreshes are removed before scanning. Only the sanitized HTML and its findings for html-has-lang, button-name, and label are sent through the proxy to Beam.

The uploaded page gets at most two Beam attempts. A rejected candidate is sent back to Beam with the deterministic failure message. The sanitized copy is retained if both attempts fail. The proxy rate-limits requests per network to slow abuse.

Recorded comparison

Five reasoning efforts, measured

All five efforts ran the same two-page fixture through the CLI on 28 October 2026, one attempt per page. Playwright and axe-core verified every candidate. All five repaired both pages. This is a single combined evidence file, not a live five-mode run, so it adds no Beam calls when you read it.

Recorded CLI benchmark: one attempt per page, no retry
Effort	Pages repaired	Targeted	Total tokens	Prompt	Completion	Reasoning	Elapsed	Requests
Low	2 / 2	6 → 0	3,767	2,990	777	362	11.2 s	2
Medium	2 / 2	6 → 0	4,104	2,992	1,112	737	9.6 s	2
High	2 / 2	6 → 0	4,413	2,994	1,419	1,013	11.2 s	2
X-high	2 / 2	6 → 0	4,628	2,994	1,634	1,224	11.7 s	2
Max	2 / 2	6 → 0	4,719	2,994	1,725	1,314	13.6 s	2

Download combined evidence JSON The file contains all five full CLI reports: request IDs, usage, edits, acceptance results, and the limitation statement. It does not contain the API key.

Elapsed time is the CLI wall clock for each mode, including Playwright scans. Low was slower than medium in this sample; that is run-to-run variance, not a monotonic latency curve. Token counts increase with effort.

Run the bundled demo

Reasoning effort

Loading the live repair runner…

A single run makes up to four Beam requests, two pages with at most two attempts each. The demo sends only the fixed page name, the selected effort, and, on a retry, the rejection message. The recorded CLI run above remains the primary evidence.

The recorded result

Counts come from Playwright and axe-core before and after the accepted edits.

Run 2026-10-27 16:59 UTC · Beam-1.0-501B-A23B · reasoning: medium · 4,105 tokens · CLI output run-20261027-165821-808

Targeted findings

6 → 0

All violation nodes

6 → 0

Pages repaired

2 / 2

Elapsed

55.6 s

Tokens

4,105

Before and after results by page and rule
Page	Outcome	Targeted	Rules before	Rules after	Beam attempts
index.html	Repaired	3 → 0	`button-name: 2` · `html-has-lang: 1`	None	1 of 2
contact.html	Repaired	3 → 0	`html-has-lang: 1` · `label: 2`	None	1 of 2

“Repaired” means the candidate passed the acceptance rule: the targeted count decreased, the total did not increase, no new rule appeared, and no existing rule increased. It does not mean the page is fully accessible.

The fixture, before and after

Printworks Studio is bundled demo data, not a real organization.

Preview page Home Contact

Screenshot of the original Home page: a Printworks Studio search page with two icon-only buttons. The layout is unchanged by the repair; the changes are in the HTML source. — Before Original fixture

Screenshot of the repaired Home page. It looks the same as the original because the accepted changes are attributes in the HTML source. — After Accepted repair

Screenshots of the saved fixture and the accepted copy. The layout barely changes; the accessibility repair is in the HTML source shown below. Form submissions are not part of the fixture.

What Beam changed

Six accepted edits. Each one is an exact string match applied once.

`index.html` — 3 edits

html-has-lang <html> became <html lang="en">.
button-name The search button gained aria-label="Search".
button-name The clear button gained aria-label="Clear search".

`contact.html` — 3 edits

html-has-lang <html> became <html lang="en">.
label Added <label for="email">Email</label>.
label Added <label for="message">Message</label>.

Beam proposed the edits; application code validated the response, found each find string exactly once, applied it to a fresh copy, and rescanned. The original fixture was never edited.

Accepted changes

The exact unified diff saved with the run.

changes.patch 10 changed lines

===================================================================
--- a/index.html
+++ b/index.html
@@ -1,6 +1,6 @@
 <!DOCTYPE html>
-<html>
+<html lang="en">
 <head>
   <meta charset="utf-8">
   <meta name="viewport" content="width=device-width, initial-scale=1">
   <title>Printworks Studio — Home</title>
@@ -30,10 +30,10 @@
       <form class="search-form" action="index.html" method="get" role="search" aria-labelledby="search-heading">
         <label for="search-query">Search our print catalogue</label>
         <div class="search-controls">
           <input id="search-query" name="q" type="search">
-          <button type="submit" id="search-button"><span aria-hidden="true">&#128269;</span></button>
-          <button type="reset" id="clear-button"><span aria-hidden="true">&times;</span></button>
+          <button type="submit" id="search-button" aria-label="Search"><span aria-hidden="true">&#128269;</span></button>
+          <button type="reset" id="clear-button" aria-label="Clear search"><span aria-hidden="true">&times;</span></button>
         </div>
       </form>
     </section>
 

===================================================================
--- a/contact.html
+++ b/contact.html
@@ -1,6 +1,6 @@
 <!DOCTYPE html>
-<html>
+<html lang="en">
 <head>
   <meta charset="utf-8">
   <meta name="viewport" content="width=device-width, initial-scale=1">
   <title>Contact — Printworks Studio</title>
@@ -28,12 +28,14 @@
       <h2 id="form-heading">Send a message</h2>
       <form class="contact-form" action="contact.html" method="post">
         <div class="field">
           <p class="field-hint" id="email-hint">We only use this to reply.</p>
+          <label for="email">Email</label>
           <input id="email" name="email" type="email" autocomplete="email">
         </div>
         <div class="field">
           <p class="field-hint" id="message-hint">Include format, quantity, and timing if you can.</p>
+          <label for="message">Message</label>
           <textarea id="message" name="message" rows="6"></textarea>
         </div>
         <button class="primary-button" type="submit">Send message</button>
       </form>

Diff and patch are also available as changes.patch.

Verification trail

Request IDs, finish reasons, and token usage recorded by the CLI.

Beam attempts

index.html — repaired, 1 attempt

Attempt 1: accepted

request 196d425b-761e-4541-9561-162700ca2f39 · finish stop · 1,981 tokens · 3 edits

1,520 prompt · 461 completion · 248 reasoning

Rescan: 3 targeted → 0; 3 total → 0; no validation or rejection errors.

contact.html — repaired, 1 attempt

Attempt 1: accepted

request e33fdbd4-b2e5-48e3-879a-c4df974e9db9 · finish stop · 2,124 tokens · 3 edits

1,426 prompt · 698 completion · 486 reasoning

Rescan: 3 targeted → 0; 3 total → 0; no validation or rejection errors.

Aggregate usage: 2,946 prompt · 1,159 completion · 734 reasoning · 4,105 total. The maximum is two attempts per page.

Saved run artifacts

report.json Full scans, page outcomes, edits, request IDs, usage, and limitations.
changes.patch The unified diff between the fixture and the repaired copy.
patchproof-a11y-source.zip Source code, tests, fixture, and the recorded run.
Repaired index.html The accepted copy, opened directly.
Repaired contact.html The accepted copy, opened directly.

No residual findings were reported for either page in this run. The report’s own limitation statement still applies.

Run it yourself

The same repairSite pipeline backs the CLI and the local web demo.

CLI

npm ci
npx playwright install chromium
$env:REFLECTION_API_KEY = "<your Reflection key>"

npm run scan -- ./fixtures/demo-site
npm run fix  -- ./fixtures/demo-site

Local web demo

npm run build
npm start
# open http://localhost:7860

The server accepts no URLs, uploads, prompts, or visitor HTML. It only repairs the bundled fixture, serializes requests, and serves the latest run.

The Reflection API key stays in the server process. It is never sent to the browser or written to report.json. The hosted live runner uses the Cloudflare Worker proxy in worker/ for the same reason; the static page performs network requests only when you press Run live Beam repair.

What this result does not prove

Automated checks with axe-core cover only a subset of accessibility requirements. They do not establish full WCAG compliance. Manual review by people with disabilities is still required.

Only index.html and contact.html were scanned.
Only html-has-lang, button-name, and label were targeted.
CSS, JavaScript, focus order, reading experience, and content quality are outside this demo.
The recorded run is a saved artifact. The live runner performs a fresh, rate-limited repair with one browser-side axe-core pass per candidate; the recorded CLI run remains the stronger evidence because it used Playwright on served copies.