Open research · 2026-08-26

Evidence records 5/5. Automated screenshots 2/5.

This is not a success-only performance chart. v0.1 publishes successful evidence records alongside failed screenshots so readers can see what the product can—and cannot—guarantee.

5/5

exact quotes and locators

5/5

public evidence pages HTTP 200

2/5

automated screenshots succeeded

Case results

Failure states are results, too.

Each case requires a retrievable record, exact quote, locator, and HTTP-200 public evidence page. Screenshot capture is scored independently; source blocking and browser errors are not relabeled as success.

01

simple factual text

✓

Quote

✓

Locator

✓

Public page

Screenshot: done

The record uses the exact quotation itself as its text locator rather than a shorter explicit anchor.

Inspect public evidence ↗

02

context-dependent text

✓

Quote

✓

Locator

✓

Public page

Screenshot: failed · cloudflare_challenge

The W3C text record is public, but automated screenshot capture encountered a Cloudflare challenge.

Inspect public evidence ↗

03

table value

✓

Quote

✓

Locator

✓

Public page

Screenshot: done

This single NIST table row does not measure performance on merged cells, pagination, or interactive tables.

Inspect public evidence ↗

04

public PDF page evidence

✓

Quote

✓

Locator

✓

Public page

Screenshot: failed · browser_error

The PDF quotation and page are preserved, but the screenshot browser failed on this public NASA PDF.

Inspect public evidence ↗

05

source permits text but may block screenshot automation

✓

Quote

✓

Locator

✓

Public page

Screenshot: failed · access_denied

The SEC filing text was independently verified, while the source denied automated screenshot access.

Inspect public evidence ↗

Method and limits

We manually verified passages from five public primary sources, submitted those pre-verified passages to the production evidence API, and checked public pages from outside the production service sandbox. The test therefore measures faithful evidence preservation and publication—not whether a model independently finds the right passage.

Five cases cannot estimate a general success rate, and third-party defenses can change. The NASA PDF browser_error, W3C cloudflare_challenge, and SEC access_denied outcomes are retained verbatim.

Dataset v0.1 · Published by VeriGlow · Contains no customer accounts, private documents, or credentials.