Docs / Reference / Glossary

Glossary

The terms below are used precisely and consistently across the product, the docs and our conversations with you.

Traces and evidence

Trace: One complete run of your agent: the request that started it, every event it produced, and the final output. The unit that gets judged.
Event also called a span: One step inside a trace: a model call, a tool call, a retrieval, a chain wrapper. Events form parent and child chains, and a piece of evidence belongs to a specific one.
Field: A named part of an event, such as its input, its output, or a tool result.
Evidence: The passage of the trace a code or a verdict rests on, kept tied to its place in the trace rather than copied out of it. This is what makes a number auditable: you can open any instance and read the words behind it.

Open code: One human observation about one trace: a short label, the evidence it rests on, and an optional note. The raw material of the method. See Open coding.
Axial code also called a failure mode: A named category consolidating many open codes, with an explicit scope. The unit that gets monitored, charted and reported. See Axial coding.
Calibration: Two or more people coding the same traces independently and reconciling the differences by tightening code definitions. See Collaboration.

Judge: The model that decides whether a trace exhibits a given axial code. Returns a structured verdict, never free text.
Verdict: One judge decision for one trace against one code: yes, no, or insufficient, with a confidence and the supporting evidence.
Insufficient: A verdict meaning the trace genuinely does not contain enough to decide the question. Not a negative, and excluded from the denominator of the rate.

Monitored code: An approved axial code promoted to daily measurement. How many can be active at once is capped by your plan. See Monitoring.
Trace day: The day the traffic actually happened. Rates are attributed to trace day, so a spike lands where the traffic was, not where the scoring was.
Evaluation day: A completed trace day that has been scored. Partial days are not evaluated, so a rate is never computed from half a day of traffic.
Error rate: For one code, flagged traces over evaluated traces on a given trace day. On the lead tile, the share of evaluated traces flagged by at least one monitored code.
Sampling: The share of a day's traces that gets judged. Ingestion is not sampled; only paid judging is. The rate is set by your plan. See Monitoring.