Trace a worked example

You operate a legal research assistant. A teammate proposes a change that sounds beneficial, but you require a trace connecting configuration to evidence. Here is the observed run:

A policy assistant must answer from weekly regulations in a strict JSON schema. Prompt-only JSON validity is 86%; RAG improves factual freshness but not validity. A small format-tuning set raises validity to 99%, while RAG supplies dated clauses. Removing RAG makes current-fact accuracy collapse, proving each layer’s role.

First identify the input and scope. Which user, segment, document, image, query, hardware profile, or task was involved? Next record the exact configuration: model or checkpoint, prompt, index, context policy, sampler, thresholds, and tool versions that matter for fine-tuning versus RAG. Then preserve the intermediate artifact that explains the result. Finally attach the user-visible output and measured consequence.

Write the trace as a sequence rather than a conclusion:

request + configuration
  -> intermediate evidence
  -> model or policy decision
  -> validation / fusion / routing
  -> user-visible action
  -> measured outcome

This format prevents hindsight from collapsing several stages into “AI error.” It also exposes where a deterministic check could have stopped propagation.

Use absolute counts alongside percentages. If success falls from 78 of 100 to 62 of 100, that is a 16 percentage-point decrease, not merely “16% worse.” If cost rises from $0.006 to $0.018 for one million requests, variable spend rises from $6,000 to $18,000. If a sample contains only ten cases from a critical language, one miss moves its rate by ten points; collect more evidence before claiming stability.

Measure behavior separately from knowledge: format validity and instruction adherence for tuning; recall@k and citation correctness for retrieval; end-to-end grounded task success for both. Include freshness updates, access-control tests, latency, cost, and regression suites. Pick one primary metric and list gates separately. Do not average a privacy breach, severe unsafe action, or failed authorization with stylistic quality.

Generate at least three explanations: input mix changed; a component configuration changed; or measurement changed. Then propose a discriminating test for each. Replay the same cases on old and new configurations, compare intermediate artifacts, and rescore both with the same rubric. This controls more variables than debating outputs by eye.

Teams fine-tune on changing policies and get stale answers, expect RAG to teach a stable output schema, train on synthetic mistakes, retrieve unauthorized documents, or evaluate only fluent answers. A tuned model may memorize sensitive examples; RAG may cite an irrelevant but similar passage. The likely failure should match the earliest divergent artifact. If it does not, revise the hypothesis.

Choose among keep, roll back, canary, route, or collect more data. State the owner and deadline. A rollback restores safety but does not explain root cause; preserve the failed configuration for offline reproduction. A successful fix adds the case to a regression set and updates the runbook.

The expert habit is modest: claim only what the trace supports. One run can demonstrate a mechanism, not a universal advantage. A coherent sequence with inspectable evidence teaches more than a polished before-and-after screenshot.