A
Admin
**Show and Tell.** English-primary. Seeking critique, not endorsement.
We drafted a one-page “harness card” that launch posts could paste beside any agent demo. Goal: make poorer-harness re-runs and grader type visible without a 20-page appendix.
## Draft fields
- Model + tools + sampling
- Step budget / retries / wall-clock
- Grader type: environment assertion | human | LLM-judge (triage-only?)
- Task table (n, domains)
- Median / p95 tool calls and token/$ per success
- One poorer-harness re-run result
- Kill / override path
Inspired by threads on agentic evidence https://www.asihub.org/t/62 and LLM-as-judge limits https://www.asihub.org/t/54.
## Ask
What field is missing that would have caught a false positive in your last eval cycle? What field would you delete as noise?
— ASIHub Editorial (official staff account)
We drafted a one-page “harness card” that launch posts could paste beside any agent demo. Goal: make poorer-harness re-runs and grader type visible without a 20-page appendix.
## Draft fields
- Model + tools + sampling
- Step budget / retries / wall-clock
- Grader type: environment assertion | human | LLM-judge (triage-only?)
- Task table (n, domains)
- Median / p95 tool calls and token/$ per success
- One poorer-harness re-run result
- Kill / override path
Inspired by threads on agentic evidence https://www.asihub.org/t/62 and LLM-as-judge limits https://www.asihub.org/t/54.
## Ask
What field is missing that would have caught a false positive in your last eval cycle? What field would you delete as noise?
— ASIHub Editorial (official staff account)