
What happened
The author ran 433 Claude Code-written PHP questions through a custom Artisan command, questions:verify-php, on PHP 8.2.0; 403 matched the registered answers and none mismatched.
Why it matters
The zero-mismatch result covers only code output against registered answers, not whether the explanation or the other answer choices are sound, so the verification is narrower than a clean pass suggests.
Summaries like this, every morning.
The setup matters as much as the result. Before generating anything, the author fixed rules: questions about code output must be executed in PHP and checked against the answer, and no question goes live until a person has reviewed it. Claude Code handled the execution checks during creation, and a separate Artisan command was built to re-run all code-containing questions at once, described in its own help text as an annual review tool. The command sorts each question first, skipping multiple-choice and web-server-dependent items, then running fill-in-the-blank choices and plain output questions through a separate PHP process with errors included in the comparison. That sorting logic is where the gap appeared: a text-prefix condition meant questions phrased as file or terminal executions never reached the comparison step, even though they asked for output. The author later widened the condition and added explicit reasons to exclude two questions that read POST body input and one that is a terminal command rather than a PHP file. Separately, re-checking the local database against the JSON turned up five missing questions and several stale answer choices, which shifted the first tally from 397 matches and 434 code-containing questions to the corrected 403 and 433. The author also notes what the command cannot see: answer choices that are conspicuously longer than the rest, and hints left in the question text, were found by reading, not by execution, and Claude Code made the fixes.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
BOOKERS now supports MCP, so users can connect their account to Claude or ChatGPT (web version, confirmed work…

IMPACTLAKE began offering a ChatGPT plugin for impactlake(TM) and added Model Context Protocol (MCP) support…

Google DeepMind introduced Gemini 4 Hydrogen and Gemini 4 Carbon, with Hydrogen being integrated into Google S…

On a Dragonwing RB3 Gen 2, 6 of 9 InternVL3.5-2B graphs profiled via the manual AI Hub chain

The author installed the open-source agent platform OpenClaw on an EPSON Endeavor NY2100S (Celeron P4600, 4GB…

On a 1-minute-3-second Japanese clip, LDX Voice detected all 9 misrecognitions, matched 11/11 glossary terms…
