How this site uses AI
Last updated: 31 August 2026
Short version: every SCORE and NUMBER on this site - Intrinsic Value, Quality, Psychology, Discovery, Moat, the Value/Long Score, Trade Setup - is computed, not AI-written. A handful of specific features additionally use Claude (Anthropic’s AI model) to turn already-computed numbers into a written paragraph. Every one of those paragraphs is labelled where it appears, every time, with the words “AI-written summary of the site’s data - not advice.”
What’s computed (never AI)
The DCF, the quality calculation, the psychology and discovery readings, the Moat Score, the Trade Setup gates and targets - all of it is deterministic arithmetic run against live market data, exactly as How the scores work describes. No AI model ever sees this arithmetic or influences what it outputs. This is true whether or not any AI feature below is even configured on a given deployment.
What’s AI-written, and where
- Ask boxes (Deep Dive, and My Portfolio for signed-in users) - a grounded answer to a typed question, built only from that page’s own already-computed numbers.
- Natural-language screening (Home, Scanner) - a plain-English query translated into the Scanner’s own real filters (country, universe, sector); the resolved filters are always shown back to you before the results run.
- Portfolio AI watchdog - a nightly “what changed” brief for holdings you’ve opted a portfolio into, sent only when something material actually happened.
- Personalised weekly brief - the weekly watchlist email’s written paragraph, covering your own watchlist moves, portfolio health changes, and relevant new research from that week.
- “Explain this number” - the (i) next to a Deep Dive gauge, explaining what that specific figure means for that specific stock.
- Admin research-note drafting and comment spam/abuse triage - admin-only tools that draft a first-pass blog post for review, and flag a pending comment for the admin’s attention. Neither one publishes or moderates anything by itself - a human reviews and clicks Save/Publish, or Approve/Reject, every time.
None of it gives investment advice. Every AI feature is instructed to describe what the site’s own numbers show and to cite which section a claim rests on, never to recommend buying, holding or selling anything - the same rule the computed scores themselves follow.
Models, and why two
Claude Haiku 4.5 is the default model for every AI feature above. Research-note drafting and the weekly brief use Claude Sonnet 5 instead, because both call sites draft longer first-pass text meant for review (an admin drafting one note, or one brief per subscriber per week) rather than a short grounded answer - low-frequency enough that the higher per-call cost stays immaterial. Nothing about which model runs changes what data it’s allowed to see or what it’s allowed to say.
Limits, and why
Using an AI feature requires signing in - free accounts get 20 questions a day; a Plus subscription (where offered) raises that to 300 a month. A site-wide monthly spend cap stops all AI features except the owner’s own once reached, and every call is logged (which feature, which model, how many tokens, estimated cost) so usage stays visible and bounded. “Explain this number” additionally caches each explanation per stock and metric, so the same figure is only ever explained once, no matter how many visitors read it.
Your data
Nothing from My Portfolio, your watchlist, or your email address is ever sent anywhere except into the grounded prompt for a question you yourself asked (e.g. your own Portfolio Ask, or your own watchdog brief) - never exposed through the public snapshot pages, the JSON API, or the MCP server, and never used to train any model. See the privacy policy for the site’s full data-handling practices; this page only covers what the AI features specifically do and don’t do with what they see.