Cross-check and translate for Luciana

Step 1: Complete the methodology memo

Open the methodology memo and fill in the remaining sections. Document:

• The model you selected and why
• The threshold you chose and the precision-recall numbers at that threshold
• The limitations -- what the model cannot predict, what data quality issues affect its reliability, what conditions would make the predictions untrustworthy

Be specific about limitations. If you noticed anything unusual about the 2022 vintage or the 2023 data, note it here. If the model performs differently on certain vineyard plots, note that too.

Step 2: Cross-model review

Open a second Claude session. Give the second AI only Luciana's voicemail transcript, the data dictionary, and your completed methodology memo. Do not share the original session history.

Ask it to review: "Does the metric choice match the client's priorities? Is the threshold justified? Are there features I should have excluded? Is the model honest about its limitations?"

Read the review. A fresh perspective may catch things your original session normalized -- a threshold that could be better tuned, a limitation you understated, or a feature relationship you overlooked.

Address legitimate findings. If the second AI flags something that is not actually a problem, note why. Both responses sharpen your thinking.

Step 3: Translate the confusion matrix into Luciana's terms

The confusion matrix has numbers. Luciana needs barrels.

Not "precision = 0.83" but: "Of every 10 barrels the model flags as Reserve candidates, about 8 actually deserve it and 2 are standard barrels you will taste unnecessarily."

Not "recall = 0.85" but: "The model catches about 85% of the barrels that truly are Reserve quality. About 1 in 7 Reserve barrels will slip through without being flagged."

Direct AI to rewrite the confusion matrix results in this language. Check that the translation is accurate -- the numbers should map exactly to the matrix cells.

Step 4: Translate the feature importances

Luciana asked which production factors drive quality. Direct AI to extract feature importances from the selected model and translate them into production language.

Not "altitude_m has coefficient 0.034" but: "Barrels from the high-altitude Alto plot are more likely to reach Reserve quality. Fermentation temperature matters too -- barrels fermented below 26 degrees Celsius score consistently higher."

Ground every finding in Luciana's vineyard. She knows her plots. She knows her fermentation process. The translation should connect the model's numbers to decisions she already makes.

Step 5: Draft the findings summary

Structure the findings around Luciana's five requirements from the voicemail, not around the methodology. She asked for:

1. Catch the Reserve barrels
2. Which production factors drive quality
3. When to trust the model and when not to
4. Something that works for next harvest
5. Something she can explain to export partners

For each requirement, write a clear answer in her language. Lead with what she asked for. The methodology supports the findings; it is not the findings.

Review the voicemail transcript (materials/voicemail-transcript.md) one more time to confirm you have addressed everything she asked for.