行业知识 · v2.3.0 · 资料核对 2026-10-03

AI Evaluation and Launch Plan

在交互知识库中阅读下载 Markdown 原文

1. Decision

• System / version / candidate:
• Decision: feasibility / selection / acceptance / release / expansion / regression:
• Baseline:
• Decision owner and reviewers:

2. Task and risk taxonomy

Include core, high-value, long-tail, refusal, adversarial, permission, and high-impact cases.

3. Dataset governance

• Source, date, authorization, and sensitive-data treatment:
• Development / validation / test split:
• Leakage controls:
• Version and change log:
• Refresh and retirement:

4. Rubric and evaluators

• Human qualification and adjudication:
• Model-judge calibration and agreement:
• Deterministic checks:

5. Metrics and gates

6. Error analysis

7. Security, privacy, and red-team checks

• Data authority and deletion:
• Tenant/access isolation:
• Prompt injection and untrusted content:
• Tool authorization and irreversible action:
• Abuse, unsafe output, fairness, transparency:
• Professional/legal reviews required:

8. Operational readiness

• Trace/version coverage:
• Dashboards and alerts:
• Capacity, rate limits, supplier/fallback:
• Human sampling and coverage:
• Runbook, support, incident roles:
• Rollback mechanism tested:

9. Rollout stages

10. Decision

• Go / conditional Go / hold / rollback / stop:
• Evidence:
• Violated gates:
• Conditions, owners, due dates, and expiration:
• Rollback triggers:
• Next evaluation trigger: