IRLLABIndependent Reinforcement Learning Lab
IRL LAB / FIELD NOTE 001EST. 2026

Small worlds.
Hard questions.

A personal laboratory for rebuilding reinforcement learning from first principles—one controlled environment, exact baseline, and honest uncertainty statement at a time.

Study 01 is runningExact analysis complete · empirical conclusions withheld

01 / CURRENT STUDY

A noisy corridor is fast. A southern route is longer but protected. Where does the optimal route switch—and does a learning algorithm target the same boundary when exploration continues?

Q-learningSARSAExpected SARSA
Read the live research scaffold
exposed corridor protected route hazard cells
Known dynamicsTwo consequence lawsPaired seed panel

THE LAB CONTRACT

Mechanisms before leaderboards.

  1. 01Solve what can be solved.

    Known finite environments get exact dynamic-programming baselines before training begins.

  2. 02Match the comparison.

    Methods share interaction budgets, seed panels, and held-out evaluation conditions.

  3. 03Separate design from evidence.

    Calibration shapes the protocol. Full-run artifacts support the published result.

  4. 04Show what would change the conclusion.

    Unresolved actions, instability, and inconvenient outcomes remain visible.

REPRODUCIBILITY LEDGER

Evidence has a paper trail.

StudyShortcut or Shelter?
Exact modelComplete
Full artifactsPending validation
Public resultWithheld