Commodity-Price Drift Monitor — Make the AI Design the Checklist, Then Bring the Index

• buyer-pack.txt — what you hand the AI: your five-contract portfolio + the Section 8.4 price-adjustment clause. No index data.
• verification-pack.txt — what you bring, every month: the two commodity indices the AI cannot fetch for itself.

The split is the whole lesson. Start with the buyer pack.

---

Thursday, 2:47pm. Two browser tabs open: a steel PPI chart sliding south since January, an aluminum PPI climbing the other way. Your VP just pinged: "Quarterly review is Tuesday. What's our commodity exposure across the metals portfolio? How many contracts are at risk?"

You pull the file. Five fixed-price supply agreements, two commodity indices, one clause — Section 8.4 — that lets someone trigger a price renegotiation if the index moves more than 10% from the baseline you signed at. You've read that clause before. You've never actually calculated whether it's been tripped.

The natural move: paste the portfolio into an AI tool and ask "Which of these contracts is at risk?" That question asks the AI to judge risk using a number it doesn't have — this month's commodity index, which updates monthly on data.bls.gov, past the AI's training cutoff. The AI has your contracts and your clause, but not this month's index value. Without it, a risk verdict is either a refusal or a fabrication.

In our four runs on 2026-07-02, three tools refused to answer when asked for the verdict cold. One — Gemini — invented index values and flagged all five contracts as at-risk; the real answer, once the actual indices were supplied, was two. The refusal is the safer failure, but neither gives you a working monitor. The fabrication is worse: a confident wrong table, repeated monthly, is a credibility problem on a timer.

The method that works has three steps: make the AI design the monitor and name the data it's missing, supply that data yourself from a source you trust, then draft a one-page note that cites the index values and pull date so next month's run is a five-minute refresh. The skill is not getting AI to build a report. The skill is knowing which part of the report the AI cannot source — and bringing it yourself, every month.

---

✓ Every AI output quoted below is from a real run on 2026-07-02 — screenshots and full exported transcripts in evidence/. All four platforms (ChatGPT, Claude Sonnet, Gemini, Copilot) ran the full flow. Per-platform details are in the Reference section at the end.

---