行业知识 · v2.3.0 · 资料核对 2026-10-03

Safety, Launch, and Operations Reference

在交互知识库中阅读下载 Markdown 原文

Risk-based depth

Increase testing, approvals, human control, audit, and rollout caution as impact, scale, autonomy, sensitivity, and irreversibility increase. High-risk domain conclusions require qualified professional and legal review.

Pre-release gates

Verify:

• Product scope and intended users.
• Data authority, minimization, access, retention, deletion, and logging.
• Tenant isolation, secret handling, Prompt injection, tool authorization, and misuse.
• Quality and severe-tail thresholds by segment.
• Human confirmation, refusal, escalation, correction, and undo.
• Capacity, latency, cost, rate limits, and supplier failure.
• Monitoring, alerts, sampling, runbook, rollback, and incident roles.
• User disclosure, support, training, and operations ownership.

Hard gates must be pass/fail. Do not compensate a critical privacy failure with higher average answer quality.

Rollout ladder

1. Offline evaluation.
2. Internal/sandbox use.
3. Shadow mode without affecting users.
4. Trusted cohort with strong review.
5. Small random canary.
6. Gradual expansion by segment and capability.
7. General availability with continuous sampling.

For each stage, define traffic/users, duration, owner, quality/safety/cost thresholds, human coverage, expansion decision, and rollback trigger.

Monitoring

Monitor service, model/system quality, safety, behavior, cost, and business outcome:

• Availability, errors, P50/P95 latency, queue and tool failure.
• Task success, refusal, groundedness, regression and feedback.
• Unauthorized access/action, injection, sensitive output and abuse.
• Input/task/user distribution drift and model/version changes.
• Tokens, tool calls, retries, loops, human review and cost per success.
• Adoption, completion, rework, retention, complaint and downstream outcome.

Combine automated alerts with scheduled human sample review.

Rollback triggers

Predefine triggers for critical security/privacy events, harmful or unauthorized actions, quality below floor, cost runaway, uncontrollable latency/error, corrupted telemetry, untraceable version, or failed human-control mechanisms.

Use a degradation ladder: disable write actions, require confirmation, route to safer model, use retrieval-only/search, reduce scope, switch to human/manual, or turn off the feature.

Incident first response

1. Assign incident commander and recorder.
2. Confirm scope, severity, continuing impact, and affected versions.
3. Contain with disable, rollback, limit, revoke, or isolate.
4. Preserve relevant evidence and logs within privacy rules.
5. Notify required technical, product, security, legal, operations, and support owners.
6. Publish confirmed facts, unknowns, next checkpoint, and owner.

Restore safely before conducting blame analysis. After stabilization, document timeline, root/system causes, affected users, remediation, regression tests, and prevention owners.

Conditional Go

Use conditional Go only when unmet items are not hard gates. Give each condition an owner, deadline, validation, expiration, operating restriction, and automatic response if it remains open.