/evals/evaluation_plan.md
Evaluation plan
Defines target behaviors, known risks, input constraints, and the decision the evaluation should support.
Artifact sample
This page uses sample data and representative structure only. It is not a client case study, testimonial, metric claim, or evidence of a delivered result.
Representative structure
The value is not the numbers in this sample. The value is the evidence structure: what was evaluated, why it matters, what failed, and how the release decision should change.
/evals/evaluation_plan.md
Defines target behaviors, known risks, input constraints, and the decision the evaluation should support.
/analysis/failure_taxonomy.md
Groups reproduced failures by trigger, suspected cause, severity, and owner.
/gates/release_gate_checklist.md
Lists blockers, warnings, thresholds, and review behavior before rollout.
/decision/ship_revise_stop_memo.md
Summarizes the evidence and recommends whether the path should ship, be revised, or stop.
Sample table
These rows show the shape of an evaluation table. They are synthetic examples and should not be read as engagement evidence.
The table is intentionally synthetic. A real client table would need approved redactions, measurement context, and permission before publication.
| task id | baseline | candidate | failure mode | decision |
|---|---|---|---|---|
| sample_0142 | blocked | passes | wrong tool call | include in release gate |
| sample_0188 | passes | blocked | stale retrieval | revise before rollout |
| sample_0211 | blocked | passes | missing context | keep in replay suite |
| sample_0273 | passes | blocked | unsafe action | block release |
Sample JSON
Real client artifacts should include evidence metadata. This sample shows the intended labeling pattern.
{
"artifact_type": "representative_eval_summary",
"data_status": "sample_not_client_data",
"workflow": "support_triage_agent",
"decision": "revise_before_ship",
"gates": {
"unsafe_action_rate": "blocker",
"stale_retrieval": "warning",
"human_escalation_path": "required"
}
}