# External Test Oracles Make AI-Generated Code Trustworthy ## Claim AI-generated code only becomes verifiably correct, not just plausible, once it's checked against a test oracle that neither the AI nor the human reviewer authored or controls. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds strongest when a mature, pre-existing, adversarial test corpus already exists (a language's own regression suite, an established compliance test set) — the oracle's independence is what does the work. - Weakens in greenfield domains with no pre-existing ground truth, where the "oracle" would itself have to be built by the same human/AI pair it's meant to check, and so isn't fully independent. - Doesn't by itself guard against the oracle's own plumbing lying (see the linked CRLF-normalization failure mode) — the oracle needs periodic verification too, not blind trust. ## Evidence Ekin Ertaç's from-scratch PHP interpreter ("Phargo"), written in Rust by an AI agent and directed by an author who states he "cannot audit the code," is validated entirely against PHP's own ~22,000-file upstream `.phpt` test suite rather than against tests either the author or the AI wrote. The project reached a real WordPress front page and `/wp-admin/` render specifically because that external, adversarial scoreboard — "cannot be flattered, negotiated with, or prompted into a better mood" — caught bugs a non-expert human reviewer would have missed, including several "Potemkin" features that ran without erroring while doing nothing at all. ## Links - [[AI & the Future of Knowledge Work MOC]] — the independent check that converts plausible output into verified output - [[3. Resources/Permanent Notes/A Test That Cannot Return No Cannot Validate Yes]] — parallels: a check only demonstrates what it can also refuse; an oracle you didn't write is what supplies that discriminating power - [[3. Resources/Permanent Notes/Eval Quality Is the Rate-Limiter on Safe Agent Autonomy]] — extends: names the mechanism (eval quality) this claim shows a concrete instance of - [[3. Resources/MOCs/AI Agent Safety MOC]] — applies-in: the maker/checker discipline this claim generalizes into a governance principle ## Tests / falsifiers - Find a project that gates AI-agent code behind an AI-authored (not third-party) test suite and compare its post-merge regression rate against Phargo-style external-oracle projects; if rates are comparable, oracle independence isn't doing the claimed work. - Check whether Phargo's pass-rate curve keeps climbing once fuzzing moves past the fixed 22,000-test corpus — a static oracle that stops catching new bugs once "solved" would undercut the general claim. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Literature Notes/I Don't Know Rust. My AI Is Rewriting PHP in It Anyway]] — reports the Phargo project's scoreboard mechanic, the CRLF harness bug, and the WordPress-boot milestone.