The AI Output Verification Pack

You are my verification desk, acting as four people on every pass: a forensic fact checker who has caught fabricated sources before, a research librarian who confirms a citation by finding the actual source, an audit senior who ties every number back to a document, and a paranoid managing partner whose name goes on this and who has been burned once.

Before we start, ask me and wait for my answers:

1. What am I verifying?
   (A) A report or memo with citations
   (B) A financial model or a deck full of numbers
   (C) A legal or contract document
   (D) A research piece or literature review
   (E) A board or investment committee pack

2. Where is the material? This decides how the next prompts run.
   (A) I upload the draft AND the source documents it cites into this Project.
   (B) I paste the draft text, and I will paste sources when you ask.
   (C) I use the Claude add-in for Microsoft Word or Excel on the live file.
   (D) I pull sources from a governed connector (our VDR, a filings database, a market-data terminal).
   (E) A mix.

3. Context: who is the audience for this deliverable, who signs it, and what is the worst outcome if a wrong line gets through (refund, regulator, lawsuit, lost deal, reputational)?

Once I answer, confirm the setup in three lines, then wait for the material. On every later pass: judge a claim only against a source you can actually see. If you cannot see the source, say UNVERIFIED, never assume it is fine. End every answer with: "Verification pass only. A named human signs before this ships."

Using the material from prompt 01, check that every citation in the deliverable actually exists.

For each reference, footnote, link, and source named:
1. State the claim it is attached to.
2. Verdict: EXISTS (you can confirm it from the uploaded sources), UNVERIFIED (you cannot see the source to confirm), or LIKELY FABRICATED (the source has the shape of an AI invention: plausible title, real-sounding author or journal, but no trace in the provided material).
3. For anything not EXISTS, say exactly what you would need to confirm it.

Output a numbered citation table: claim, cited source, verdict, what is missing. Rank the LIKELY FABRICATED items to the top. Do not pass a citation as fine just because it looks well formatted. A clean format is how fabricated sources hide.

End with the count: X exist, Y unverified, Z likely fabricated. Then "Verification pass only. A named human signs before this ships."

Using the material from prompt 01, trace every quotation and every attributed statement in the deliverable.

This is the Deloitte failure mode: quotes pinned on court judgments and papers that never said them.

For each quote or "X said / the ruling found / the standard requires" line:
1. Find the exact source and location (page, section, paragraph) in the provided material.
2. Verdict: TRACED (the source says this, word for word or in clear substance), PARAPHRASE DRIFT (the source exists but does not support the wording or the meaning), MISATTRIBUTED (a real source, wrong author or wrong document), or UNTRACEABLE (no source in the material supports it).
3. Quote the actual source line next to the deliverable's version so I can compare.

Flag every PARAPHRASE DRIFT, MISATTRIBUTED, and UNTRACEABLE item as a blocker. A quote with no traceable source does not ship. End with "Verification pass only. A named human signs before this ships."

Using the material from prompt 01, tie every number in the deliverable back to a source.

For each figure, percentage, total, and date that carries weight:
1. Trace it to the source document, the input I provided, or a calculation you can reproduce from those.
2. Verdict: TIED (sourced or reproducible), DERIVED (you can recompute it from tied inputs, show the arithmetic), or UNSOURCED (no basis in the provided material).
3. For DERIVED numbers, recompute and confirm the deliverable's figure matches. Flag any mismatch.

Output a tie-out table: number, where it appears, source, verdict, recomputed value if applicable. List every UNSOURCED number and every recompute mismatch at the top as blockers. Never accept a number because it sounds reasonable. End with "Verification pass only. A named human signs before this ships."

Using the material from prompt 01, test each substantive factual claim against what the cited source actually says.

This is the EY failure mode: a real citation attached to a claim the source does not support.

For each claim that cites or implies a source:
1. Read what the source actually says in the provided material.
2. Verdict: SUPPORTED (the source backs the claim), OVERSTATED (the source says something weaker or narrower), CONTRADICTED (the source says the opposite or materially different), or SOURCE NOT PRESENT (cannot test without the source, mark UNVERIFIED).
3. For OVERSTATED and CONTRADICTED, quote the source line and show the gap.

Rank CONTRADICTED first, then OVERSTATED. These are the lines that survive a citation check but fail on substance. End with "Verification pass only. A named human signs before this ships."

Using the material from prompt 01, hunt for assertions stated with more confidence or precision than the sources support.

Find:
1. False precision: a specific figure or date where the source is vague or a range.
2. Hardened hedges: a source that says "may" or "estimated" turned into "will" or "is".
3. Stale as current: a past data point presented as today's number, with no date.
4. Single source stated as consensus: "studies show" or "it is well established" resting on one or zero sources.
5. Plausible filler: smooth sentences that assert a fact with nothing behind them.

For each, quote the line, state why it overreaches, and give the corrected, defensible version. End with "Verification pass only. A named human signs before this ships."

Using the material from prompt 01, check the deliverable against itself.

1. Numbers that contradict across sections (a figure on page 2 that disagrees with the same figure on page 9).
2. Totals that do not foot, subtotals that do not sum, percentages that do not reconcile to their base.
3. Dates, names, and defined terms used inconsistently.
4. Conclusions that the body does not support.

Recompute every total and percentage. Show the arithmetic. Output a list of every contradiction and every failed recompute with the two conflicting locations. End with "Verification pass only. A named human signs before this ships."

Now verify your own verification. Do not skip this.

1. Re-read prompts 02 through 07 and the deliverable. List anything you marked TIED, EXISTS, TRACED, or SUPPORTED that you cannot actually defend from a source you can see. Move it to UNVERIFIED.
2. Produce a sign-off sheet with four lines, each PASS or BLOCK:
   - Citations: every cited source EXISTS or is removed.
   - Quotes: every quote TRACED to a real source.
   - Numbers: every figure TIED or DERIVED and recomputed.
   - Claims: every claim SUPPORTED by its source, no CONTRADICTED line left.
3. Count the unresolved items. State the rule plainly: if the count is above zero, this deliverable is NOT cleared to ship.

If anything is BLOCK, list exactly what must be fixed and stop. Do not soften the verdict to be helpful. A clean sign-off you cannot defend is the exact failure that cost Deloitte the refund. End with "Verification pass only. A named human signs before this ships."

Using prompts 02 through 08, produce the verification report a reviewer or auditor can keep.

1. One-line verdict: CLEARED or NOT CLEARED to ship.
2. Blocker list: every fabricated citation, untraceable quote, unsourced number, and contradicted claim, each with its location and the fix.
3. The sign-off sheet from prompt 08.
4. A verification log: what was checked, against what sources, and the date, so there is a record of the pass that survives later edits.
5. The single highest-risk line in the document and why.

Keep it to two pages. This is the artifact that proves the work was checked. End with "Verification pass only. A named human signs before this ships."