Post your loop here and I'll help you design the error signal

[fitz2882]

The one thing LoopGain needs from you is a single non-negative number per iteration that says how wrong the current output is. Lower is better, zero means done. Everything else is derived from that number's trajectory.

That's the whole integration — and it's also the part people get stuck on. So: describe your loop in a reply and I'll suggest a signal. No format required. "I have a RAG pipeline that re-retrieves when the answer looks incomplete" is enough to work with.

Starting points, if yours resembles one of these:

Loop	Error signal
Agentic coding (write → run tests)	number of failing tests
JSON / structured extraction	number of schema violations
RAG with self-correction	number of required facts still missing
Self-refinement with an LLM judge	gap to target, e.g. 10 − quality_score
Lint / format loop	lint error count
Tool-use retry chain	consecutive failure count

observe() takes a number or any sequence — hand it the raw list of failing tests and it uses the length. So often there's no conversion work at all.

Two rules, and one escape hatch:

1. Non-negative.
2. Gets smaller as the output gets better.
3. If your quality is fuzzy and has no natural zero, run target_error=None. LoopGain then stops when the number stops improving, wherever that plateau lands, instead of waiting for a target that never arrives.

The failure mode to avoid is a signal that's noisy but doesn't actually track quality — a judge score that wobbles ±2 on identical output will read as oscillation. If you're unsure whether yours is stable enough, say so in your reply and we'll look at it.