AI Wrote a PHP Engine in Rust, Passes 17% of PHP-src Tests, Renders WordPress

Someone who doesn't know Rust just had an AI build a PHP interpreter in it. The result, Phargo, serves a 26 KB WordPress front page from a SQLite database — through an engine containing zero lines of PHP's actual C source code. It's a from-scratch interpreter written in Rust, generated by an AI agent with the human's role reduced to: run tests, inspect score, say "continue" or "regressed, look again."

The experiment uses PHP's own test suite as an impartial oracle: ~22,000 .phpt files accumulated over 30 years, covering everything from DateTime daylight-saving math to var_dump() on floats. Current pass rate: 3,844 of 22,037 (17.4%). Because the suite includes C extension tests (GD, curl, SOAP, MySQL drivers) that are out of scope, the realistic ceiling is ~40-45%. The project started at zero and climbs via failure histograms: the AI identifies the biggest cluster of failing tests, implements a fix, runs the full 22,000-test suite (~7 minutes), and commits if the number goes up — no human code review needed.

Lessons from the Trenches

Early progress plateaued from a subtle bug: line endings. The test corpus was checked out on Windows with CRLF endings, and the scoreboard compared output byte-for-byte, silently failing every multi-line test. PHP's own test runner normalizes before comparison. One line of normalization code turned hundreds of tests green. The lesson: measure your measurement — your oracle is only as honest as the harness connecting you to it.

Another discovery: some older regression tests allocate absurd structures or expand into infinite generators, originally designed to run inside PHP's carefully-fenced CI. One such test hard-restarted the development machine. The project now filters tests that are known CI-only bombs.

The Loop

• AI runs a failure histogram over the whole corpus to find the biggest fixable cluster.
• AI implements the fix.
• Scoreboard runs all ~22,000 tests (~7 minutes).
• If the number goes up: commit, push, repeat.
• If it goes down: human says "hmm, that regressed, look again."

The human's job is essentially aiming — reading terminal output like a medieval king reviewing naval charts, then typing the developer's most powerful phrase: "looks good, continue."