# Editable AI product requirements document

OpenMax editorial worksheet. Replace bracketed fields using authorized evidence. The document remains DRAFT until a real owner records an approval for a specified version and purpose. This form is not a compliance certification, release approval or evidence of an implemented OpenMax integration.

## 1. Document status, owner and requested approval

- PRD ID / version / date: [ ]
- Document owner / accountable product owner: [ ]
- Required reviewers and decision authority: [ ]
- Status / next decision date: [ ]
- Requested decision and purpose; separately record permission to build and release: [ ]

## 2. Executive summary

- User, blocked task and proposed outcome: [ ]
- Why decide now; alternatives if the proposed implementation changes: [ ]
- Decision requested, not a presumed launch commitment: [ ]

## 3. Problem, evidence and counterevidence

- Current authorized observations and source IDs/versions/dates: [ ]
- Exact claims supported; inference and missing evidence: [ ]
- Baseline, excluded participants and alternative explanations: [ ]
- Unresolved research question / owner / resolution date: [ ]

## 4. Target users and contexts

- Operator role, affected people, task, environment and language: [ ]
- Initial inclusion/exclusion and permission conditions: [ ]
- Keyboard, assistive-technology and other access needs: [ ]

## 5. Goals and metric definitions

For each metric: [ID; event; numerator; denominator; cohort; window; data source; owner; proposed threshold; rationale; decision]. Separate operator acceptance from factual quality. Define percentiles and whether failures are excluded. Label unvalidated planning targets and keep unavailable baseline/cost data unknown.

## 6. Non-goals

For each exclusion: [capability; reason; enforcement needed; condition and authority required to reopen]. A label or prompt does not enforce a boundary.

## 7. Scope and requirements

For each requirement: [stable RQ ID; observable required behavior; rationale/source; priority; dependencies; acceptance-check IDs]. Split independently failing behaviors. Distinguish product outcomes from implementation preferences.

## 8. Experience and recovery

- Entry/loading/preview/empty/denied/timeout/correction/exit behavior: [ ]
- Existing user work preserved; manual continuation: [ ]
- Status messages, focus and required accessibility checks: [ ]

## 9. Data, integrations and AI configuration

- Authorized sources, versions, inputs, outputs, identifiers and freshness: [ ]
- Missing/conflicting/stale/inaccessible source behavior: [ ]
- Model candidate, prompt/retrieval/evaluation-set versions and generation settings: [ ]
- Training/fine-tuning only if applicable; configuration not selected stays unknown: [ ]

## 10. Security, privacy and specialist review

- Information flow, request-time/display-time permissions: [ ]
- Logs, sensitive content, retention, deletion and changed access: [ ]
- Qualified review owners, evidence, outstanding decisions and blocking effects: [ ]

## 11. Dependencies, constraints and fallback

For each dependency: [system/team/vendor; owner; evidence; missing condition; deadline; delay consequence; authorized fallback]. Identify the cost basis without invented prices.

## 12. Risks and triggers

For each risk: [failure; affected people; detection; response owner; stop/restart conditions; residual decision]. Name the version of any framework used. Creating a mitigation task does not prove mitigation.

## 13. Acceptance, release and rollback

- Each check: [ID; requirement; input/setup; expected result; actual result; evidence; configuration; PASS/FAIL/NOT_RUN; owner/date].
- Limited-pilot and expanded-release conditions / authorized decision: [ ]
- Open defects, specialist reviews and unresolved blockers: [ ]
- Disable/manual recovery/support; externally irreversible effects: [ ]

## 14. Questions and decision history

- Open question: [ID; owner; evidence needed; due date; blocking effect].
- Decision: [ID; old choice; new choice; reason; approver; date; version/purpose; affected RQ/checks/docs].
- Preserve superseded records. Do not silently alter a target to make a result pass.