[JacobStephens2]

Summary

 GLOSSARY.md
+  Eval, Eval suite, Eval task, Arm, Trial, Grader, Grade, Replay
 docs/adr/
+  0019-session-logs-become-session-traces.md

The #698 grilling settled two things, and this PR records them.

• Eval terms (Q10). Anthropic's "Demystifying evals for AI agents" is the nearest thing to a standard, so the terms follow it where it has a word: Eval suite, Trial, Grader. Arm, with the Baseline and Candidate arms, and Replay fill its gaps. Its Task becomes Eval task here, because the Ticket entry already tells agents to avoid "task".
• ADR 0019. Session logs become Session traces: one OTLP JSON Lines file per session, holding an OpenTelemetry log record for each Harness event as it arrives, plus spans. It answers Spec: Session traces, OTLP JSON Lines in place of Session logs #700. The glossary renames Session log to Session trace only once the switch is built, so Runs working on today's logs aren't told to call them Session traces.

Part of #698.

Evidence

• The terms and the decision come from the answered rounds of .scratch/grill-with-docs/698-evals.md on the evals branch: Q3, Q10 and Q23–Q28.
• No Avoid line clashes. The glossary avoids "baseline" for the Usual range and "task" for Tickets. The Baseline arm and the Eval task name different things.
• The ADR's claim that the GenAI semantic conventions and the OTLP file format are both still Development was rechecked on 2026-10-20: the conventions are still at 547934a, with no releases.

Merge Danger

Door: two-way

Documentation only: the code still writes Session logs.

Blast Radius: docs

Runs look terms up in GLOSSARY.md. The new entries name things no code has yet, and no existing entry changes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01BnFUcJ51dAFoj97E8FSXQF