A
Admin
**AI Safety & Alignment — executable template.** Official staff.
Green safety evals + user harm still happens (deployment gap: https://www.asihub.org/t/60). Below is a one-page review template we use after an agent incident. Critique and redlines welcome.
## Template
1. **Harm statement** (who, what irreversible effect, blast radius).
2. **Eval status at ship** (suite + version + harness). Mark contamination risk (https://www.asihub.org/t/73).
3. **Which gate was missing:** allowlist / state assertion / human approval / kill switch.
4. **Repro under poorer harness** (yes/no + link to log).
5. **Why the green eval did not cover this** (OOD task, tool not in suite, judge-only grading…).
6. **Fix + verification** (what re-run must pass before re-enable).
7. **Owner + clock** (who holds the revoke token).
## Discussion
What row would you add from the last incident you are allowed to describe (anonymized OK)?
— ASIHub Editorial (official staff account)
Green safety evals + user harm still happens (deployment gap: https://www.asihub.org/t/60). Below is a one-page review template we use after an agent incident. Critique and redlines welcome.
## Template
1. **Harm statement** (who, what irreversible effect, blast radius).
2. **Eval status at ship** (suite + version + harness). Mark contamination risk (https://www.asihub.org/t/73).
3. **Which gate was missing:** allowlist / state assertion / human approval / kill switch.
4. **Repro under poorer harness** (yes/no + link to log).
5. **Why the green eval did not cover this** (OOD task, tool not in suite, judge-only grading…).
6. **Fix + verification** (what re-run must pass before re-enable).
7. **Owner + clock** (who holds the revoke token).
## Discussion
What row would you add from the last incident you are allowed to describe (anonymized OK)?
— ASIHub Editorial (official staff account)