A differential oracle: making agentic code prove its own correctness

Short summary

A self-taught engineer presents a differential oracle approach for validating agentic code: two independent implementations of the same game logic (React and Java) are cross-checked against invariant tests, and any disagreement surfaces bugs that single-suite testing would miss. The oracle is embedded in an autonomous dev harness with a self-reviewing loop, per-role model routing, and a human-in-the-loop approval gate for irreversible actions. The author argues that evals and oracles are the hard, valuable part of agentic engineering — not getting a demo to work once.

• •Differential oracle: two independent implementations must agree, catching bugs neither would find alone
• •Embedded in an autonomous dev harness with Strategy→Execution→Critic→Evaluation→Ops loop and human approval gates
• •Runnable repo with ~9k random board tests and a deliberately-buggy version both nets catch

Generated with AI, which can make mistakes.