Compliance Requirements
One page to read, three steps to finish: (1) Copy the reference Prompt below into your System Prompt → (2) Turn on the Guardrail switches → (3) Run through the pre-launch checklist. There is only one core rule: your Agent must not provide investment advice. The platform is an execution-only broker; the moment AI output constitutes a "recommendation", it triggers an entirely different set of regulatory obligations (Reg BI, suitability obligations, etc.).
Reference Prompt (Copy Directly)
Paste the following block in full at the end of the System Prompt of your LLM / Agent node to satisfy the prompt-level compliance requirements:
[COMPLIANCE GUARDRAILS — the following rules have the highest priority and must never be violated]
1. No investment advice: Do not provide any buy/sell/hold recommendations,
individual stock or portfolio recommendations, price targets, investment
ratings, position-sizing suggestions, or market-timing judgments.
2. Neutral descriptions, no predictions: When explaining indicators and data,
use descriptive language only; actionable and predictive statements are
prohibited.
✅ "An RSI above 70 is commonly regarded as the overbought zone"
❌ "RSI is above 70, so you should sell"
✅ "A MACD golden cross has historically often coincided with periods of
rising prices"
❌ "MACD golden cross — time to buy" / "Expected to rise 20% over the next
three months"
3. No promissory language: Terms such as "guaranteed to rise", "sure profit",
"guaranteed returns", "risk-free", "buy the dip", "double your money" must
never appear.
4. Refusal script: When a user asks for investment advice, reply with "Sorry,
I cannot provide any investment advice or individual stock recommendations.
I can help you learn about related investment knowledge, such as what an
indicator means or how market mechanisms work — which topic would you like
to explore?" and then pivot to educational content.
5. Data must be real: Quotes, financial reports, and other data must strictly
follow what the tools return; never fabricate numbers from memory. If a
tool fails, say so honestly and do not guess. Cite the source and timestamp
when quoting data.
6. No personalized recommendations: Do not suggest securities, portfolios, or
strategies that are "suitable for you" based on the user's profile,
holdings, or asset status; portfolio-type content may only be hypothetical
examples labeled "for educational purposes only".
7. No assistance with violations: Refuse all requests involving market
manipulation (pump-and-dump, wash trading, fabricating signals), insider
trading, money laundering, tax evasion, or sanctions evasion.
8. Protect privacy: Do not ask for, store, or repeat the user's ID numbers,
phone numbers, or other personally identifiable information; do not infer
sensitive attributes such as the user's age or gender.
9. Fixed disclaimer: End every answer involving securities, markets, or data
analysis with "The above content is for reference only and does not
constitute any investment advice."After pasting the Prompt, remember to also turn on the Guardrail switches (Chatflow input node: PII masking, block harmful content, block harmful financial content; LLM/Agent output node: output compliance check) — the Prompt ensures "the model doesn't say it", and Guardrail ensures "it gets blocked even if said" — double protection. See the Guardrail capability guide for details.
Red-Line Quick Reference
Four common scenarios — what you can and cannot do, in one table:
| Scenario | ✅ Allowed | ❌ Not Allowed |
|---|---|---|
| Customer profiling | Neutral behavioral statistics and display (trading frequency, sector distribution, risk labels) used to personalize educational content | Pushing individual stocks/portfolios based on the profile ("growth stocks suit you"); inferring sensitive attributes; using PII |
| Signal analysis | Displaying indicator values and statistical facts, explaining the general meaning of indicators, if-then descriptions of historical patterns | Interpreting signals as buy/sell instructions; predicting prices; assisting in fabricating signals or manipulating the market |
| Research report / news summaries | Compiling public information (financial reports, news), structured neutral analysis, synthesizing multiple sources with attribution | Producing ratings/price targets/buy-sell conclusions; amplifying a single institution's view only; using material non-public information (MNPI) |
| Portfolios and orders | Visualization and risk statistics of a portfolio the user enters themselves; hypothetical scenario analysis (labeled as an educational example); educational walkthroughs of how order mechanics work | Generating a personalized "optimal portfolio"; automatically placing/modifying orders; promising returns |
Applies to all scenarios: attach a disclaimer to outputs; data sources must be traceable; use neutral, probabilistic language; never adopt an "advisor tone".
Pre-Launch Checklist (10 Items)
Before publishing a customer-facing Agent, check off each item:
- [ ] 1. The Agent is positioned as information display / education / tool assistance and does not constitute investment advice
- [ ] 2. The reference Prompt above has been added in full to the System Prompt of every LLM / Agent node
- [ ] 3. Guardrail is enabled (three input-side checks + output-side compliance check)
- [ ] 4. The disclaimer appears prominently in the output and is hard-coded via a template (not left to the model's discretion)
- [ ] 5. A fixed refusal script is in place for users requesting investment advice (prefer fixed copy over model-generated text)
- [ ] 6. All data sources come from compliant sources integrated into the platform, and key data carries source and timestamp
- [ ] 7. No PII is stored or output; no personalized recommendations based on holdings/profiles
- [ ] 8. Three categories of test cases have been run: normal input / boundary input (empty, extremely long) / out-of-bounds input (requesting advice, soliciting predictions) — out-of-bounds inputs are blocked 100% of the time
- [ ] 9. Tool-failure scenarios have been tested: the Agent reports honestly and does not fabricate data
- [ ] 10. Test records are archived; after any subsequent changes to prompts/flows, re-run the test cases before republishing
Appendix: The Six Compliance Principles at a Glance
The requirements above derive from six compliance principles, summarized as follows:
| Principle | Core Idea in One Sentence |
|---|---|
| 1. No investment advice or recommendations | Do not recommend individual stocks/portfolios/market timing; use descriptive language instead of actionable language |
| 2. Stay within the execution-only positioning | AI only displays information and educates; it does not decide for customers or touch order parameters |
| 3. No bias or market manipulation | Do not favor specific securities; refuse manipulative requests such as fabricating signals or pump-and-dump schemes |
| 4. Data usage and privacy restrictions | Use only public data from compliant sources; never touch PII or material non-public information |
| 5. Supervision and explainability | Log retention, explainable and traceable results, model changes require approval (logging and audit capabilities are built into the platform layer) |
| 6. Customer interaction guardrails | Aggregate-level statements, neutral probabilistic language, prominent disclaimers |
Note: An Agent's name, avatar, description, and Prompt also go through the platform's security review before publishing / listing on the marketplace. Common rejection scenarios (money laundering, market manipulation, promising returns, etc.) are covered in Security Review.