21 lines
1.1 KiB
JSON
21 lines
1.1 KiB
JSON
{
|
|
"baseline": "/tmp/nemotron-replay-prod-20260602095438-openclaw-incumbent.jsonl",
|
|
"candidate_id": "langgraph_incident_kernel",
|
|
"contract_report": "/tmp/nemotron-replay-prod-20260602095438-langgraph-contract-report.json",
|
|
"contract_valid": true,
|
|
"fixtures": "/tmp/nemotron-replay-prod-20260602095438-sanitized-fixtures.jsonl",
|
|
"graded_output": "/tmp/nemotron-replay-prod-20260602095438-langgraph-graded.jsonl",
|
|
"graded_records": 50,
|
|
"grading_report": "/tmp/nemotron-replay-prod-20260602095438-langgraph-grading-report.json",
|
|
"input_records": 50,
|
|
"inputs": "/tmp/nemotron-replay-prod-20260602095438-sanitized-candidate-inputs.jsonl",
|
|
"label_grading_applied": true,
|
|
"normalized_output": "/tmp/nemotron-replay-prod-20260602095438-langgraph-normalized.jsonl",
|
|
"normalized_records": 50,
|
|
"result_records": 50,
|
|
"results": "/tmp/nemotron-replay-prod-20260602095438-langgraph-candidate-raw.jsonl",
|
|
"schema_version": "agent_replay_pipeline_report_v1",
|
|
"scorecard": "docs/evaluations/agent_langgraph_replay_scorecard_2026-06-02.json",
|
|
"scorecard_written": true
|
|
}
|