QL Intelligence can make mistakes.

QL Intelligence is built on large language models and, like every AI of this generation, will sometimes return information that is inaccurate, outdated, or fabricated. Here's why — and how to use it well.

Why AI gets things wrong

QL Intelligence is powered by transformer-based language models. They generate responses by predicting the most likely sequence of tokens given the conversation, the tool results, and their training data. That probabilistic generation works remarkably well most of the time — and fails in three recurring ways:

• Hallucinations. The model can produce a confident, well-written claim that has no basis in the underlying data. This is most likely when it's reaching beyond what was retrieved by the tools it called.
• Stale knowledge. The model's training corpus has a cutoff date. For anything that has changed since — earnings, mergers, regulatory news, leadership shifts — it relies on the live tool calls (Polygon, SEC, Finnhub, Alpha Vantage, web search). If those calls didn't fire or returned thin data, the model may fill the gap with what it remembers — which may no longer be true.
• Math + numeric drift. Even with correct source data, models occasionally arithmetic-error. A percentage may be off by 10%, a ratio inverted, a date misread by a quarter.

What QuantLogix does to mitigate

Most of what makes QL Intelligence different from a generic chatbot lives in this category.

• The signal engine is authoritative, not the AI. Every Long-Term, Swing, and Day signal you see is computed by the 5-factor quantitative engine (Technical / Momentum / Fundamental / Options / Microstructure) — not generated by the AI. The AI quotes signal scores; it does not invent them.
• Tool-grounded responses. Most chat answers fire one or more typed tool calls (live quotes, SEC filings, analyst ratings, 13F holdings, options chains, walk-forward backtests). The AI is instructed not to fabricate numbers — if a tool returns "no data," it should say so rather than guess.
• Multi-agent debate for high-stakes asks. For deep-dive due diligence, the 10-agent panel (5 analysts in parallel → Bull → Bear → Bull rebuttal → Bear closing → Trader synthesizer) intentionally lets opposing views attack each other so the surviving claims are the cleanest.
• Verified signal track record. Every signal the engine produces is immutably logged at issuance and resolved against live market data. Engine self-validation runs nightly — broken factors get auto-down-weighted or zeroed via operator override. Win rates published at /longterm-signals, /swingtradesignals, and /daytradesignals are real outcomes, not back-tests.
• Cross-session memory + thesis tracking. Chat conversations and your open theses are preserved in your account. When the AI quotes your prior call on NVDA, that's a real prior thesis with a real outcome — not a hallucinated memory.

What you should do

Use QL Intelligence the way you'd use any junior analyst: as a force multiplier on your own judgment, not a replacement for it. Specifically:

• Double-check critical numbers. If the AI quotes a P/E ratio, a 13F holding %, or an analyst price target — and you're about to make a decision on it — open the underlying source (signal page, SEC filing, analyst panel) and confirm.
• Treat any forward-looking claim as a hypothesis. "NVDA will hit $200 by Q3" is the AI reasoning under uncertainty, not a forecast. Weight it with the signal score and the multi-agent verdict, not just the prose.
• Never trade on AI commentary alone. The signal engine + your own thesis are the gating layer. AI commentary is for context, comparison, and stress-testing.
• Flag obvious errors. If a response cites a non-existent SEC filing, a wrong CEO, or a date that's off by a quarter — let us know via /contact. We track these systematically.