Methodology

How the roast is generated

Each portfolio you submit is processed in three deterministic stages. The LLM runs exactly once per portfolio — whether you pay or not. Payment unlocks the locked sections of a roast that has already been written and cached; it does not trigger a second LLM call.

• Screenshot OCR (if you upload an image): we call gemini-3-flash-preview to extract a JSON list of holdings. No image is stored after extraction.
• Currency normalization: deterministic code (no LLM) converts non-USD positions and computes percentage weights against your total. This step runs in a Cloudflare Worker — see /worker/src/currencyNormalizer.js.
• Roast generation: the normalized portfolio plus our published prompt are sent to gemini-2.5-pro. The prompt is the “Five Blades of Execution” structure — Opening Salvo, five themed critiques (concentration, correlation, cost, behavior, recovery), and a Redemption Arc. Output targets 600–800 words.
• Asset analytics (background, anonymized): we call gemini-2.5-flash-lite to track which asset names appear across roasts so we can prioritize what to improve. Names only — no portfolio composition, no user identity.

Data sources

• The portfolio itself: whatever you paste or upload. We do not connect to your broker. We do not look up live prices. The roast reasons over the data you give us, period.
• CFTC Commitments of Traders (used by our weekly newsletter, not the per-portfolio roast): published Friday 3:30 PM ET by the U.S. Commodity Futures Trading Commission. Source: cftc.gov.
• Research we cite in blog posts: Vanguard, Morningstar, S&P Dow Jones SPIVA, DALBAR. Specific studies are linked inline in the post that cites them — citations live in the article, not in this hub.

What we explicitly don't do

Known limitations of the LLM

Large language models hallucinate. Ours can too. Specifically:

• Ticker / ISIN ambiguity: if you submit only a short ticker, the model may guess the wrong listing. Include the full security name or ISIN when you want precision.
• Fee estimates: when a fund's expense ratio isn't in the model's training data, the roast estimates a typical fee for that asset class. Estimates are flagged as such in the cost section; verify against the fund's factsheet before acting.
• Asset class assignment: edge-case holdings (e.g., closed-end funds, structured notes, niche ETFs) can be miscategorized. The concentration critique is only as good as the categorization.
• No live data: prices, holdings, and underlying-index composition are frozen at the model's training cutoff. Recent reconstitutions or splits may not be reflected.

Who's behind it