AI Methodology & Limitations

1. AI Models in Production

Stock Expert AI uses a multi-model architecture, selecting the best-fit AI engine for each task:

• Gemini 2.5 Flash (Google, gemini-2.5-flash): Company dossiers, stock-page narrative and other analytical content. OpenAI GPT-4o mini is the fallback when Gemini is unavailable.
• Gemini 2.5 Flash (Google, gemini-2.5-flash): Machine translation of English pages into the other site languages.
• Gemini 2.5 Flash (Google, gemini-2.5-flash): Moon AI conversational assistant. Accessed through the Gemini OpenAI-compatible endpoint; OpenAI is the configured fallback provider.
• Claude 3 Haiku (Anthropic, claude-3-haiku-20240307): Legends Council and sector-insight narratives.
• GPT-4o (OpenAI, gpt-4o): Portfolio screenshot reading (OCR) — first of two readers run in parallel.
• Gemini 2.5 Pro (Google, gemini-2.5-pro): Portfolio screenshot reading (OCR) — second reader; the two results are reconciled.
• GPT-4o mini (OpenAI, gpt-4o-mini): Quick news analysis, company summaries and prediction assistance.

2. Hallucination Prevention

Translated AI output passes through a 6-layer validation pipeline before publication. It compares the translation against its English source — it is a fidelity check, not a second look at the underlying data:

1. Numeric carry-over: Figures present in the English text must reappear in the translation.
2. Ticker consistency: All ticker mentions in generated text must match the page entity. Mismatches trigger regeneration.
3. Length ratio: Translated paragraphs preserve length within adaptive ratio bounds (prevents content collapse).
4. Paragraph preservation: ≥60% structural fidelity between source and AI output.
5. HTML tag validation: No malformed markup, no template variable leakage.
6. Title sanity: Length 5-300 chars, no system prompt artifacts, no model self-references.

Failures trigger up to 2 retry-with-feedback attempts. Output that keeps failing is left unpublished rather than shipped. Numbers themselves are checked where they are produced, against the provider response — the figures on a page are read from stored data, not written by a model.

3. Source Data Grounding

AI models are instructed to stay on source data — no speculation, no invented numbers. Prompts include explicit grounding instructions:

• "ONLY use facts from the provided source data. If a fact is not in sources, output 'Unknown'."
• "NO speculation or estimation. No invented numbers."
• "Time-awareness: Categorize catalysts as 'Upcoming:' or 'Ongoing:'. Do not list past events as future catalysts."
• "COMPLIANCE: Never use advisory phrases like 'good buy', 'should buy', 'compelling investment'. Present facts neutrally."

That is the instruction, not a guarantee. Where a company's filings are thin, the model sometimes generalizes from its sector rather than answering "Unknown". When it does, the page says so in its own Analysis Notes block — 40 published dossiers currently carry such a note. Treat those sections as sector-typical context, not as facts about that company.

4. AI Editorial Personas (Transparent Disclosure)

Market Journal and stock analysis articles are attributed to AI editor personas (e.g., Dakota Finch, Chase Merrington, Alex Sterling). These are not real people. They are AI personas representing distinct analytical styles within our system. The pattern follows industry transparency standards (The Atlantic's Bertie, Washington Post's Heliograf, Bloomberg's Cyborg).

Every author profile page carries an explicit "AI Editor" disclosure banner with editorial accountability attributed to founder Sedat ANAK via schema.org Person + agent reference.

5. Known Limitations

• Not predictive: AI scores are pattern-recognition signals on historical data, not future predictions.
• Validation isn't perfect: Despite 6-layer validation, occasional inaccuracies may surface. Reported errors reviewed within 24 hours.
• Market regime sensitivity: Signals trained on past regimes may underperform in unprecedented market conditions.
• Coverage gaps: Recently listed stocks, micro-caps, and OTC tier names may have limited data depth.
• AI is not re-run on every price tick: a dossier carries the date it was written and is not on a refresh cycle. Quotes refresh when a page is viewed during market hours from FMP/Alpaca feeds — not from AI inference, and not a live exchange feed.
• Translation drift: Multilingual content (TR/ES/DE/ZH) is AI-translated; rare nuance loss possible despite a strict 1:1 translation policy.

6. Regulatory & Legal Status

Stock Expert AI is not a registered investment advisor under the Investment Advisers Act of 1940 or any equivalent regulation. The platform provides educational research tools and AI-generated analysis for informational purposes only. No content constitutes a buy/sell recommendation, personalized financial advice, or guarantee of investment outcomes.

All users are encouraged to consult a licensed financial advisor before making investment decisions. See the full disclaimer for complete legal disclosures.

7. Reporting AI Errors

Found an inaccurate AI output? Report it:

• Email: sedat@stockexpertai.com
• Include: page URL, the specific claim or number, and the correct value if known.
• Response time: 24 hours on business days.
• Systemic issues: If an error reveals an AI pipeline weakness, we address the root cause to prevent recurrence.

8. Related Documents

• Stock Analysis Methodology — the current V2 MoonshotScore, historical V1 reference, data sources, Quality Gate.
• Editorial Policy — Content creation, error correction, accountability.
• Disclaimer — Legal disclosures and risk disclosure.
• About — Founder Sedat ANAK and the company.