Concepts
Predictive TDD
Test-driven development in which every check is preceded by a stated, falsifiable expectation. Before a test runs, the agent writes down what it expects to happen: whether the code compiles, whether the test fails, and with which message or value. Afterwards it compares the prediction with the actual result. A wrong prediction is a signal to stop and understand the code before changing it, not something to explain away.
Work proceeds in Red–Green–Refactor cycles, one test at a time: Red makes exactly one new behavior fail for the predicted reason, Green adds the smallest change that passes it, and Refactor improves the design while every test stays green.
The approach is inspired by Ted M. Young's Predictive TDD and TDD Game and Jon Jagger's cyber-dojo.
EXACT Coding
The lab's own family of TDD workflows for coding agents, built on Predictive TDD. Every EXACT Coding workflow consists of the same four building blocks, each a separate command, skill or agent: a test list, Red, Green and Refactor.
The test list comes first. The agent writes the behaviors it is going to build as inactive tests before any production code exists, then activates them one at a time. A test that already passes when it is activated is accepted as evidence that an earlier change generalized; no failure is manufactured for it. Refactoring follows the Four Rules of Simple Design: passes the tests, reveals intention, no duplication, fewest elements.
The workflows differ in where each building block runs — all in one shared context, each in its own isolated subagent, or a mix of the two — in how complete the test list is and whether it is cross-checked against the specification, and in the rules the refactoring applies. The Opus line was built on Claude Code with Claude models, the SOL line on pi with GPT models; either line also runs on the other harnesses.