Ground Truth Is a Process

A team reviews one hundred conversations. The classifier's proposed label appears beside each one. Reviewers correct obvious mistakes and accept the rest. The resulting spreadsheet becomes the gold standard used to evaluate that same classifier.

The evaluation looks reassuring. It partly measures how often people accepted a suggestion they had already seen.

This does not make assisted labeling useless. It makes the provenance of a reference label part of the evidence. “Human reviewed” can describe very different processes, with different kinds of dependence on the system being judged.

Write a procedure people can follow#

Reference labels need a defined unit of analysis, the evidence readers should inspect, category definitions, tie-breaking rules, and a way to preserve uncertainty. Without these, two careful people may answer different questions.

For the ticket-email example, a reviewer needs the message and reliable evidence about its sender and destination. The filter's claim that the email is suspicious is evidence about what the filter said. It is not independent confirmation that the ticket is fraudulent.

Have at least a subset of cases labeled independently before showing model suggestions. Compare the judgments and inspect disagreement. Adjudication should record why an interpretation won or why the case remains ambiguous. This makes the resolution available for improving the rubric, rather than leaving it in a meeting nobody can reconstruct.

Human agreement is a diagnostic of the task and protocol. It is not a mathematical ceiling that no model can exceed. Humans can make correlated mistakes, and a reference can be wrong despite unanimous agreement. Low agreement still deserves investigation before a small model improvement is treated as meaningful.

One sample cannot answer every question#

A useful evaluation needs coverage of ordinary traffic and enough difficult cases to expose important weaknesses. Those goals pull in different directions.

A natural sample estimates behavior under the real mix of cases. It may contain almost no examples of a rare failure. A targeted or stratified sample supplies those cases, but changes the mix. An independently double-labeled subset examines reference reliability and can overlap either group.

Keep those purposes visible in the report. Precision measured on a sample enriched with failures is not automatically production precision. A challenge set shows whether the system handles the selected challenges; it does not establish how frequently those challenges occur.

Sample sizes should follow the decision's required precision. Three examples of a class may expose an obvious defect. They cannot establish a reliable recall estimate: one additional miss changes the observed recall by a third. Repeated turns from one conversation may also be dependent, so counting each as an independent observation overstates the evidence.

Protect the exam from the lessons#

Use development cases to refine definitions and prompts. Reserve a separate test set for evaluating the chosen system. Split related conversations or episodes together when sharing them across sets would leak nearly identical evidence.

Repeatedly checking the test set and revising until it passes turns that set into development material. New labels generated by a model can help triage work, but agreement with those labels should be described as agreement with that model or ensemble until stronger validation exists.

A reference record should identify the input snapshot, taxonomy version, rater or source, whether suggestions were visible, adjudication, sampling cohort, and evaluation split. These fields explain what the eventual score means.

This is the same concern explored in Evidence requires Provenance. A reference is useful when another reader can inspect how it became authoritative. The word “gold” cannot supply that history.