Build 05 · Blocked
A refinement loop, and the point where it starts drifting
Implement self-refinement on a small model. Run it long enough to find where extra passes stop helping and start degrading the answer.
Time — about a week
Done when
A state a stranger could checkThe loop runs, the drift is reproducible, and it is playable by someone who did not build it.
Why it is harder — First build with an interactive artifact — someone else has to be able to operate it.
Where it sits
After Build 03 · What a marginal dollar of test-time compute actually buys — Drift is only visible against a measured baseline.
Shipping it opens 11 Teach a small model to reason, with rewards you chose
What it reads
Just enough to unblock the next step- Build a Reasoning Model From Scratch ch5Self-refinement.
What it ships
- GameA self-refinement loop, playable
- NotebookWhere the refinement loop stops paying
Teach it back
If it cannot be explained plainly, it was not learned- 1Explain itExplain what you built to a friend who has never seen it. Plain words, no jargon, five sentences at most.
- 2Predicted vs realWhat you expected to happen, what actually happened, and why the two differ.
- 3Where it breaksThe question about it you still cannot answer, and the point where your explanation stops working.
- 4KeepThe one number, rule or picture worth keeping. It becomes a bottle and gets re-tested.
No notes on this build yet.
A note is a Learning Log entry with the tag study-self-refine-loop. It is written in Connor's words, filed as a draft, and appears here once it is published from the admin queue. One note can carry several build tags.