Production environment

Entry points and traffic path

Where users and systems enter, which components handle traffic, and how requests are encrypted, routed, limited, and blocked.

• Domain and DNS
• TLS certificates
• Load balancer, gateway, or reverse proxy
• Firewall, WAF, CDN, private network, or VPN

Compute and runtime

The hosts, containers, clusters, functions, or managed platforms that run the application, plus operating systems, runtimes, and foundational software versions.

• CPU, memory, and storage quotas
• Container images or executable artifacts
• Runtime versions
• Scaling and process management

Network and environment isolation

How production is separated from development and test, which entry points are allowed, which services may connect, and whether outbound access is controlled.

• Network segments and security groups
• Ingress and egress rules
• Internal service discovery
• Separate production and non-production accounts or projects

Data and storage

Where live business data is stored and how it is classified, encrypted, retained, archived, and deleted, including who operates each data service.

• Databases, caches, and queues
• File and object storage
• Encryption at rest and in transit
• Retention and deletion rules

Identity, access, and secrets

Permissions for users, administrators, and service accounts; approval and audit of sensitive operations; and storage and rotation of passwords, tokens, certificates, and encryption keys.

• Least privilege
• Multi-factor authentication
• Service accounts
• Secret management and access auditing

Configuration and external dependencies

Environment-specific settings and dependencies such as models, payment, messaging, maps, identity providers, or internal systems, including behavior when those dependencies fail.

• Environment configuration
• Feature flags
• Third-party APIs and quotas
• Timeouts, retries, circuit breaking, and degradation

Observability and alerting

How health is evaluated across user experience, business outcomes, and technical components, and how actionable failures reach the responsible people.

• Metrics, logs, and traces
• Health checks
• Errors, latency, capacity, and business anomalies
• Alert severity, ownership, and escalation

Backup, recovery, and operating records

How data and service are restored after deletion, failure, a bad release, or disaster, and where routine procedures, incident steps, and contacts are recorded.

• Backup scope and frequency
• RPO and RTO
• Recovery exercises
• Runbooks, incident playbooks, and contacts

Isolation from non-production

Production accounts, data, secrets, and privileged actions should not be casually shared with development or test. Isolation reduces mistakes and limits the impact of a compromised account or component.

Identifiable and reproducible state

The running application, configuration, infrastructure, and database state should be identifiable. Critical parts should be reproducible from version control, automation, or controlled documentation.

Controlled, reversible change

Code, configuration, database schemas, models, prompts, and dependency upgrades are all production changes. They need review, testing, records, a release strategy, and a stop or rollback plan.

Least privilege and traceable actions

People and programs receive only the access required for their tasks. High-risk operations have extra controls, and administrative actions leave searchable audit records.

Observable service health

A running process is not enough. Teams must observe whether critical business journeys succeed, dependencies remain available, latency and errors stay within limits, and capacity has headroom.

Failure handling and recovery

Alerts need owners and response instructions. Backups must be proven restorable, and critical failure scenarios need exercised rollback, failover, degradation, or restoration procedures.

Managed capacity, performance, and cost

Expected demand, peaks, service quotas, and scaling methods must be understood, with monitoring for resource consumption and ongoing cost.

Durable operating ownership

System owners, technical owners, business contacts, supplier boundaries, on-call arrangements, and escalation paths must remain current when teams or vendors change.

Model and invocation path

Record the model, version or alias, service region, calling account, quotas, timeouts, retries, and fallback. A vendor change behind a model alias can be a production change.

Prompts, policies, and tool permissions

Version system prompts, output rules, and tool permissions. Reading information and changing orders, sending messages, or approving payments carry very different risks.

Knowledge and retrieval state

Define production knowledge sources, authorization, freshness, chunking and index versions, citation behavior, and removal of deleted material from retrieval.

Data transfer and retention

Confirm which inputs, files, business data, and logs go to a model or external service, whether they are retained, used for training, or processed across regions, and configure the service accordingly.

Output quality and safety boundaries

Use representative cases to evaluate correctness, citations, refusal, insufficient evidence, sensitive content, and unauthorized actions. One average accuracy number is not a complete production standard.

Human handoff and stop conditions

For money, responsibility, inventory, scheduling, compliance, high-risk actions, or missing evidence, define when automation stops and how context reaches the responsible person.

Change monitoring and re-evaluation

Models, knowledge, prompts, data distributions, and external APIs change. Re-run evaluations after material changes and monitor refusal, handoff, error categories, and business outcomes.