Evaluation

Model Output Drift Watch

Catch quiet AI behavior drift before users become the monitoring system. They charge more.

Use when

A production AI workflow can change behavior because the model, prompt, tool results, retrieval index, or upstream data shifted under it.

Intermediatedifficulty
Evaluationcategory
Mia daily expansionsource

Cadence

Weekly or after model, prompt, retrieval, or tool changes

Verification

Representative outputs are compared against a baseline, regressions are classified with evidence, and risky changes are rolled back or queued for approval.

Structured loop spec

FieldValue
NameModel Output Drift Watch
CategoryEvaluation
TriggerWeekly or after model, prompt, retrieval, or tool changes
ObjectiveCatch quiet AI behavior drift before users become the monitoring system. They charge more.
Allowed inputsRelevant files, source notes, logs, tests, screenshots, metrics, or task state for this loop
Allowed actionsSelect a stable sample of real or synthetic cases covering common tasks, edge cases, refusal boundaries, formatting needs, and high-risk customer or internal workflows.; Run the current workflow with model, prompt, tools, retrieval filters, temperature, and input data recorded alongside the previous baseline where available.; Score differences as acceptable variance, improvement, regression, privacy risk, policy risk, formatting break, tool-use failure, or needs human review.; Patch one cause at a time: prompt, tool schema, retrieval metadata, guardrail, fixture, or model setting, then rerun the failed cases plus nearby controls.; Do not ship prompt/model/tool changes, publish generated content, or update customer-facing claims until must-pass cases are green or an explicit approval accepts the risk.
VerificationRepresentative outputs are compared against a baseline, regressions are classified with evidence, and risky changes are rolled back or queued for approval.
Stop conditionStop when the verifier passes, the budget is exhausted, no progress is made, a blocker appears, or approval is required.
BudgetSet a time, turn, token, retry, file, or dollar cap before running the loop.
Approval boundaryHuman approval required before publishing, sending, deleting, spending, changing accounts, touching production, or making reputational/legal/financial commitments.
Safe outputDraft, report, checklist, table, or approval-gated recommendation
Works withClaude, ChatGPT, Gemini, any tool-using AI assistant

Steps

  1. Select a stable sample of real or synthetic cases covering common tasks, edge cases, refusal boundaries, formatting needs, and high-risk customer or internal workflows.
  2. Run the current workflow with model, prompt, tools, retrieval filters, temperature, and input data recorded alongside the previous baseline where available.
  3. Score differences as acceptable variance, improvement, regression, privacy risk, policy risk, formatting break, tool-use failure, or needs human review.
  4. Patch one cause at a time: prompt, tool schema, retrieval metadata, guardrail, fixture, or model setting, then rerun the failed cases plus nearby controls.
  5. Do not ship prompt/model/tool changes, publish generated content, or update customer-facing claims until must-pass cases are green or an explicit approval accepts the risk.

Prompt

Run the Model Output Drift Watch loop. Build or reuse a stable sample covering common tasks, edge cases, refusal boundaries, required formats, and high-risk workflows. Run the current AI workflow and compare against the previous baseline with model, prompt, tools, retrieval filters, temperature, and data version recorded. Classify differences as acceptable, improved, regressed, privacy risk, policy risk, formatting break, tool failure, or needs review. Patch one cause at a time and rerun failed cases plus controls. Do not ship risky prompt, model, tool, retrieval, or customer-facing changes until must-pass cases are green or explicitly approved.

Run in Claude Code

Paste this into Claude Code (or any tool-using agent) to run the loop bounded: one change per round, the same verification every round, durable state files, and explicit stop conditions.

Run the "Model Output Drift Watch" loop from AI Loop Library (https://ailooplibrary.com/loops/model-output-drift-watch/) as a bounded loop.
Goal: Catch quiet AI behavior drift before users become the monitoring system. They charge more.
Rules: one change per round; run the same verification every round (Representative outputs are compared against a baseline, regressions are classified with evidence, and risky changes are rolled back or queued for approval.); append each round to docs/loops/model-output-drift-watch/progress.md and update docs/loops/model-output-drift-watch/state.json; stop on verifier pass, 8 rounds, 3 consecutive failed verifications, no progress, a blocker, or anything needing human approval (money, production, outbound, deletion). Finish with a proof report: rounds used, changes made, verification output, remaining risk, and the next human decision.

Get the MCP server + agent pack

Tags

LLMevalsquality monitoringproduction AI

Related

Back to all examples