# The Verification Protocol

Three checks before any AI output leaves your hands. AI drafts. You verify. You own.

**When to run it:** anything going to a client, student, funder, court, employer, or the public. Skip it only for output that stays with you — brainstorms, private notes, rough summaries.

## Check 1 — Fact check

Highlight every name, number, date, quote, citation, and factual claim in the output. Verify each against a source you trust. The AI is not a source for its own claims; "are you sure?" proves nothing.

| Claim in the output | Source I checked | Pass/Fail |
|---|---|---|
| [CLAIM] | [SOURCE] | [P/F] |
| [CLAIM] | [SOURCE] | [P/F] |
| [CLAIM] | [SOURCE] | [P/F] |

Shortcut: ask the model to list every factual claim in its own output. Use that list as your checklist — then verify each item yourself.

## Check 2 — Fit check

Facts can be right and the document still wrong. Confirm:

- [ ] Meets length and format requirements: [REQUIREMENT]
- [ ] Contains every required element: [LIST]
- [ ] Written for the right audience: [WHO]
- [ ] Sounds like me, not like a press release
- [ ] Nothing important from my source material got dropped

## Check 3 — Name check

Read the whole thing once, as the recipient.

- [ ] I would sign this with my name
- [ ] I can defend every sentence without saying "the AI wrote that" — which is not a defense to a bar, an auditor, a principal, or a program officer
- [ ] Nothing in it would embarrass me if quoted back in six months

## Scoring

- **All three pass:** ship it.
- **One check fails:** fix, then re-run that check only.
- **Fact check fails twice on the same task type:** demote that task on your Stage 1 leverage map to "draft only" or "not AI."

## The time test

For each task type, record once:

- Minutes to verify AI output: [MINUTES]
- Minutes to do the task from scratch: [MINUTES]

If verification regularly costs more than half the from-scratch time, the task is a poor AI fit. Update your map.

## Two-week failure log

Keep this log for two weeks. You will learn your tool's specific failure pattern and stop verifying blindly.

| Date | Task | Check that failed | What the AI got wrong |
|---|---|---|---|
| [DATE] | [TASK] | [1/2/3] | [WHAT] |
| [DATE] | [TASK] | [1/2/3] | [WHAT] |
| [DATE] | [TASK] | [1/2/3] | [WHAT] |

Bring this log into Stage 3. It tells you which workflows are ready to build and which need a human checkpoint baked in.

Part of the ADAPT framework — vetslearn.com/framework