Skip to main content

Reliability: Independent Retries

3. Independent Retries: Sample, Then Select​

Use this lever when the target is clear, the work is cheap to repeat, and a reliable signal can select the best result. Two plans may satisfy the same spec while one is less invasive; one implementation may miss an edge case that another catches.

The operator move has two parts:

  1. Generate meaningfully independent candidates. Separate contexts are stronger than one thread repeatedly revising its own framing.
  2. Apply independent selection pressure. Prefer executable tests; use human review for scope and intent; use an LLM judge only when the artifact is bounded and the criteria are explicit.

If a single attempt succeeds with probability R and k attempts are independent, the chance that at least one succeeds is:

R_eff = 1 − (1 − R)^k

At R = 0.95, two independent attempts produce an effective step reliability of 0.9975. The number is illustrative: shared framing and weak judgment destroy the independence that the formula assumes.

The diagram therefore includes both generation and selection. More samples without a trustworthy judge produce more output, not more reliability.

Reliability levers

Independent retries need selection pressure

Sampling raises effective reliabilitynoisy stepimplementationcandidate Afresh contextcandidate Bfresh contextcandidate Cfresh contextjudgetests / reviewR_eff ↑benefit depends on independent attempts and a trustworthy judge
Retries improve a bounded noisy step only when attempts are meaningfully independent and a separate signal can select the best candidate.

For rate limiting, retry the bounded plan or implementation step—not the entire contaminated workflow. This lever does not repair weak grounding, bad task shape, or a missing review boundary.

Next: Reliability: HITL Checkpoints