Trace a worked example

You operate a general-purpose chat model. A teammate proposes a change that sounds beneficial, but you require a trace connecting configuration to evidence. Here is the observed run:

For “Explain how phishing works for an awareness class,” base model A gives operational credential-theft steps. After preference tuning, model B refuses everything, including defensive advice. Reviewers prefer model C: it gives high-level mechanics, warning signs, and safe simulation guidance without deployable payloads. The pair becomes useful training data only after the rubric records why.

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 alignment and RLHF. 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.

Separate helpfulness, factuality, instruction following, calibrated uncertainty, harmlessness, and refusal quality. Use pairwise blinded human ratings, challenge sets, disaggregated language results, and win rates with uncertainty. Measure both harmful-compliance rate and benign-refusal rate because optimizing only one is misleading. 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.

Reward hacking produces responses that look polished to the scorer without being correct. Sycophancy mirrors a user’s false premise; mode collapse makes every answer sound alike; over-refusal blocks benign requests; hidden evaluator leakage inflates results. Alignment can also be brittle under new languages or jailbreak framing. 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.