1. Outcome
State the decision, recommendation, action, or result being produced. Make it explicit rather than implied or hedged.
The Integrity Packet
A lightweight proof trail that makes consequential AI-assisted work safe to review, challenge, transfer, and approve.
A polished draft is not necessarily reliable work. The Integrity Packet carries the context another qualified person needs to understand what was produced, where it came from, what was checked, who owns it, and what remains risky.
What the packet does
AI can accelerate research, comparison, drafting, and analysis. It cannot make accountability visible by itself. The packet turns private intent into a record that can travel with the work.
The transfer test is simple: could a qualified reviewer who was not involved in drafting understand the result, challenge its important claims, approve or reject it, and act safely using only what is in the record?
A draft becomes an Integrity Packet when each field is complete enough for the next qualified reviewer to use without reconstructing the author’s private context. Completeness is judged from the receiver’s perspective.
The six required fields
Use the level of detail that matches the consequence of the work. The packet can be brief, but it cannot hide the information a reviewer needs.
2. Assumptions
Record the conditions the outcome depends on, including dates, constraints, and places where expert judgment changed the default.
3. Evidence
Name the data, confirmations, sources, or analysis that support the important claims. A reviewer should be able to inspect the basis without chasing the author.
4. Validation
Describe what was manually checked, stress-tested, or challenged, by whom, and what remains unresolved.
5. Ownership
Name who prepared the work, who validated it, and who is accountable for release or approval.
6. Risk
Identify material unresolved risks, what could fail, and the approval or escalation path if the risk affects the outcome.
How to build one
- Choose a consequential decision, recommendation, or deliverable where a polished mistake could create cost, risk, delay, or loss of trust.
- Define the outcome before asking AI to generate or optimize the work.
- Give AI bounded context and named sources. Record the assumptions that shape the result.
- Validate the claims that matter at the consequence level of the work. Record the actual checks, not just a confidence statement.
- Name the preparer, validator, approver, residual risk, and escalation path before handoff.
- Run the transfer test. If the reviewer must ask the author for missing context, improve the packet before the work moves forward.
The Integrity Packet is used throughout the AAOS. Activate bounded capability and context, establish controls, execute the work, and carry the packet through review and handoff.
Complete example
Scenario: An AI-assisted recommendation to pause contractor spending for eight weeks while a budget variance is investigated.
Outcome: Recommend the pause for Cost Center 410, preserving safety-critical contracts.
Assumptions: Current run rate continues; pending invoices are included in the forecast.
Evidence: August finance report, approved contract register, and the open-invoice export.
Validation: Recalculated the variance against the ledger and asked a finance reviewer to challenge excluded costs.
Ownership: Jordan Lee, Finance Director; CFO approval required before action.
Risk: Pending invoices may be incomplete. Recheck the forecast before extending the pause.
A reviewer can see what the recommendation is, where it came from, what was checked, and what must happen before action.
Short-form example: a two-line packet footer
For a small deliverable or an email, the packet can travel as a two-line footer. Short form reduces friction; it does not remove accountability.
Integrity: Outcome—pause Cost Center 410 contractor spend for 8 weeks; Evidence—August finance report, contract register, and open-invoice export; Validation—ledger variance recalculated; Owner—Jordan Lee. Risk: Pending invoices may be incomplete; Approval—CFO before action; Recheck forecast before extending the pause.
The footer keeps the minimum record visible where the decision travels. Expand it into the complete form when the consequence, uncertainty, or review depth requires more context.
Start with one packet
Choose one real piece of consequential work. State one outcome, capture the six fields, complete one human validation pass, and name one accountable owner. The goal is a usable first packet that makes judgment visible and transferable.
Use an LLM to create your packet
You can ask an LLM to organize your work into an Integrity Packet, but the LLM is a drafting and organizing assistant. It cannot validate facts, grant approval, or take ownership of the result. Supply the work, the intended outcome, the named sources, the people involved, and the consequence level. Remove confidential information or use an approved environment before sharing it with an LLM.
Before you run the prompt, gather:
- the decision, recommendation, or deliverable you need;
- the audience, deadline, constraints, and consequence if the work is wrong;
- the authoritative source names, dates, links, or system records;
- what you have already checked and what still needs human validation;
- the preparer, validator, accountable owner, approver, and escalation path.
If information is missing, the packet should say so. A useful packet exposes an unknown; it does not fill the gap with a plausible guess.
Copyable prompt
Replace the bracketed instructions with your own material, then paste the prompt into an approved LLM environment.
You are helping me prepare an Integrity Packet for consequential AI-assisted work.
Purpose: Make the work safe for a qualified person who was not involved in drafting to review, challenge, approve, and act on it.
Rules:
- Use only the information and named sources I provide.
- Do not invent facts, sources, dates, people, approvals, validation steps, or confidence.
- Separate evidence from assumptions and proposed actions.
- If a required fact is missing, write "TBD—human input required" and list the question that must be answered.
- Do not say that you validated a claim. Describe a validation step only if I provide evidence that it happened.
- Match the review and approval requirements to the consequence of the work.
- Keep the accountable human owner visible. AI is never the owner or approver.
Work context:
- Workflow or task: [describe the work]
- Intended audience: [who will read or use it]
- Desired outcome: [decision, recommendation, action, or deliverable]
- Consequence if wrong: [cost, safety, legal, operational, reputational, learning, or other impact]
- Constraints and deadline: [include dates]
- Named sources and source dates: [list authoritative sources]
- Draft work or relevant notes: [paste the material]
- Human checks already completed: [describe the check and who performed it]
- Known uncertainties or risks: [list them]
- Preparer: [name or role]
- Validator: [name or role, or TBD]
- Accountable owner: [name or role]
- Approver and escalation path: [name or role, or TBD]
Return exactly two outputs.
OUTPUT 1 — Complete Markdown file
Create a Markdown Integrity Packet with this structure:
# Integrity Packet: [short title]
## Outcome
[State the intended result plainly.]
## Assumptions
[List each material assumption, including its source or date when known.]
## Evidence
[List named sources and connect each important claim to its source.]
## Validation
[State what was checked, by whom, when, and what remains unresolved. Do not turn an unperformed check into a completed check.]
## Ownership
- Prepared by: [name or role]
- Validated by: [name or role]
- Accountable owner: [name or role]
- Approval required from: [name or role]
## Risk and escalation
[Describe material unresolved risks, possible failure, required approval, and the action or escalation path.]
## Handoff questions
[List the questions the next reviewer must answer before approval or action.]
## Status
[Draft, ready for validation, validated pending approval, approved, or blocked. Explain the status briefly.]
After the Markdown file, add a short section called "Open information required" containing every missing item marked TBD.
OUTPUT 2 — Two-line email footer
Create exactly two lines that can be appended to an email or small document:
Integrity: Outcome—[short outcome]; Evidence—[short named sources]; Validation—[short completed check or "pending"]. Owner—[accountable owner].
Risk: [short unresolved risk]; Approval—[approver or "TBD"] ; Next step—[required check, approval, or escalation].
Keep the footer concise, but preserve the outcome, evidence, validation status, owner, risk, approval, and next step. Do not claim that the work is approved unless the supplied information says it is approved.Review the generated Markdown against the source material. Correct the outcome, assumptions, evidence, validation, ownership, and risk yourself, then obtain the named approval before action. The footer is a pointer to the accountability record; it is not a substitute for the complete packet when the work is consequential.