AI-generated content on Nkosuo

Two paths, same shape

Every AI panel on Nkosuo runs through one shared contract. We always compute a deterministic rule-based explanation from structured data — that's the version the page renders by default. When ANTHROPIC_API_KEY is configured, we can additionally request a Claude-written version of the same explanation. The Claude path is forced into the same JSON schema, then passed through a guardrail validator before reaching you. If Claude's output contains forbidden language (buy / sell / hold / target price / guaranteed / "price will rise"), we discard it and fall back to the rule-based version with the violation reasons attached as warnings.

What we use AI for

Where enabled, the per-stock Intelligence Report can be synthesised by a large language model (currently Claude from Anthropic). The AI receives the same structured inputs you see elsewhere on the page — session recap, liquidity signals, news sentiment, fundamentals — and is instructed to produce a short descriptive summary and signal chips.

Hard rules baked into the system prompt

• No buy / sell / hold recommendations.
• No price targets, fair-value numbers, or forecasts.
• No language implying you “should” do anything.
• No fabrication — if an input is missing, omit the point.
• No cross-stock comparisons framed as advice.
• No personalised investment advice.

Every AI-generated block on the site carries a visible “Descriptive only, not a recommendation” footer and links back to our methodology.

Limitations

LLMs can still be wrong, outdated, or miscalibrated. We believe the guardrails substantially reduce harm, but they do not eliminate it. Treat AI-written paragraphs as a convenience layer over the structured data — if the structured data is the truth, the prose is the summary. If they diverge, trust the structured data and the primary sources.

What to do if you spot a problem

Email corrections@nkosuo.com. Material AI-related corrections will appear on our public corrections ledger along with the resolution.

Data we send to the model

Only the public, structured market context needed to produce the summary. We do not send your portfolio, watchlist, or account information to any LLM.