Idler · Environments that train AI to code

Available nowhi@idler.ai

IdlerRL environments for coding

Issue N°03Coll. 2026Ref. RL-2026-R02

MandateTrain AI to code at an expert human level. Real engineering, graded against a working result.

The corpus / index

*The corpus, rotating — one generative spine

Method

a skill → a graded test

From real engineering work to a graded world. The same five steps every time.

01PerceiveFind where coding agents break on real engineering work.

02RepresentTurn the task into an environment with a checkable result.

03BuildStand up the repo, the tests, and the grader.

04ScaleMass-produce variants. Early environments become training data.

05MeasureScore where models fail, and aim the next environment there.

the engineering work

Reproduce, localize, and fix real bugs in a live repo.

Build features across an unfamiliar codebase.

Restructure code without breaking what works.

Write tests, read diffs, and catch regressions.

real, graded, frontier

Environments from real engineering work, never invented benchmarks. The skill transfers.

Every step checked against a working result. Dense reward, not just pass or fail.

Built for the best models, aimed at the engineering they still get wrong.