Demo

Two reviewers. One found nothing.

This is a real Veridu review of a real pull request, shown exactly as it came back. Nothing on this page is running — it is the stored result of one run against fastify/fastify#6965, a two-file change of +42/−2 lines that fixes a 404-handler bug.

The panel
openai
gpt-5.5
0
findings
anthropic
claude-sonnet-5
1
finding

Same diff, same instructions, same moment. One reviewer read this change and had nothing to say about it. Had you run only that one — which is what every single-model reviewer is — this page would be blank.

mediumcorrectness·lib/four-oh-four.js:78 found by claude-sonnet-5

Object.create-based context breaks own-property enumeration semantics

Switching _404Context from a shallow copy (Object.assign({}, …)) to a prototype-linked object (Object.create(…)) means hooks inherited from instance[kFourOhFourContext] are no longer own enumerable properties of _404Context. Any code elsewhere that enumerates the context via Object.keys(), JSON.stringify(), object spread, or Object.assign(target, _404Context) will silently omit the inherited hooks (preHandler, preValidation, etc.), even though direct property access (context.preHandler) still works via the prototype chain.

What it checked
  • line 78: `const _404Context = Object.create(instance[kFourOhFourContext])` replaces the prior `Object.assign({}, instance[kFourOhFourContext])`
  • Object.create sets its argument as the prototype, not as own properties, of the new object
  • Only `onSend` is explicitly assigned as an own property on `_404Context`
What it inferred — confidence: possible

If any consumer of the 404 context object relies on own-property enumeration (Object.keys, spread, JSON serialization) rather than direct property access, the fix will silently drop inherited lifecycle hooks for such code paths, reintroducing a variant of the bug being fixed.

Veridu keeps those two apart on every finding. A reviewer that cannot show what it verified is asking to be believed, and the fastest way to lose a team is a confident wrong answer with no way to check it.

One example is an anecdote

So we adjudicated every finding across 10 pull requests of a production codebase by hand. Of 44 findings judged real, only 4 were found by both models — 9.1% overlap. 40 were real and unique to one reviewer, against 2 false positives.

That is a small sample, self-adjudicated, on one codebase — so treat it as ours, not as a benchmark. An independent run of four reviewers across 146 merged pull requests found 93.4% of issues caught by exactly one reviewer and none by all four, and concluded that a second reviewer is additive rather than duplicative. Different tools, different codebase, same shape.

Run this on your own pull request

Public repos need nothing but the URL. Private ones stay private — connect a repo, or run the CLI locally against your own API keys.

npm install -g veridu  ·  veridu review <pr-url>