Principles
Principles
Primitives: properties that make LLM-generated code converge on correctness. Method: how the lab runs the measurement loop. Evidence: the literature and stack analysis behind the primitives (formerly the lever project, consumed here).
- bundled-enforcementYou cannot get the value without the verification; the check is inseparable from the thing it checks.derived-obligationsThe system computes what correct looks like from the spec; each codegen step is a layer where wrong output cannot be expressed.prescriptive-failureWhen the system rejects, it names what to do next, not just what went wrong.vacuity-detectionCatch code that technically passes but verifies nothing, the test with no assertions, the gate that always returns true.
- effect-sizes-not-leaderboardsReport effect sizes and trial counts, not ranks. A position on a board is not a finding.find-the-frictionLeverage hides where the work drags. Start there, not where it is convenient to measure.keep-what-survivesShip what survives measurement; discard the rest. The record is what held up.measurement-firstDesign the measurement before the tool. Drag that cannot be scored cannot be removed.simulate-to-iterateRun the work as a cheap, repeatable experiment so the configuration space can be searched, not guessed.