What the AI will not do
The builder will not tell you what to buy or sell. This is not a policy sentence. It is a built control with a dedicated code path, and this page documents both the control and its known limits.
The boundary
Ask the AI for a stock tip, a buy or sell call, or a recommendation, and the router selects a
tool built for exactly that: respond_with_safety_boundary.
Its own description reads: "Respond to stock tips, buy/sell advice, or recommendation requests
with the regulated safety boundary." The route intent it produces is stock_advice_request,
and it does not fall through to a normal reply.
What you get back
I don't provide buy/sell calls or stock tips. That's outside my scope.
What I can do is help you build and backtest a rules-based strategy for any stock you choose.
Supported assets: (the current universe)
Pick one to get started.
The reply is a template, not a generation. It cannot be talked around by rephrasing, because there is no model output to influence, the boundary text is composed, not written.
Requests that hit it
| You ask | Why it is refused |
|---|---|
| "Which stock should I buy today?" | Instrument selection on your behalf |
| "Is RELIANCE a good buy right now?" | An assessment of a specific security |
| "Should I go long or short here?" | A directional call |
| "Give me a strong buy for this week" | A recommendation |
| "What's the best stock for my ₹1 lakh?" | Instrument selection plus implicit suitability |
| "Should I deploy this strategy live?" | A decision about your capital |
Requests that do not
The boundary is about selecting trades for you, not about being unhelpful. All of these are answered normally:
| You ask | Why it is fine |
|---|---|
| "What is RSI?" | Education. Answered concisely |
| "Explain an opening range breakout" | Education |
| "Build a momentum strategy on RELIANCE" | You chose the instrument; the AI builds your rule |
| "Which indicators express bullish momentum?" | A question about the platform's vocabulary |
| "What dividend data do you have for RELIANCE?" | A data-coverage question |
| "Why did my backtest produce zero trades?" | Diagnosis of your own strategy |
| "Is a 1.2% stop tight for a 5-minute strategy?" | A question about strategy construction |
The line is consistent: the AI reasons about strategies, never about whether to take a position.
The two reply guards
Beyond the routing boundary, generated replies pass through content guards.
The advice guard
A pattern that catches buy/sell phrasing in a model-written reply:
- "you should buy" and "you should sell"
- "buy this" and "sell this"
- "go long" and "go short"
- "take a position"
- "strong buy" and "strong sell"
When it matches, the reply is replaced with the safety boundary text. The model's words do not reach you.
The overpromise guard
A pattern that catches assured-outcome language:
- "guaranteed" and "guarantee"
- "risk-free" and "sure-shot"
- "certain profit"
- "cannot lose" and "can't lose"
- "100% win", "100% profit" and "100% return"
- "double your money"
When it matches, the reply is suppressed rather than substituted: the turn falls back to a composed response instead.
The known limitation
Both patterns are applied when the route intent is general_chat or clarification. Replies on
every other intent, including strategy explanation and backtest narration, which are
exactly the turns where a model is most likely to characterise expected outcomes, take a path
with no content check.
The control exists and is well constructed. Its scope is narrower than its purpose, and it is documented here rather than left for you to find out. The remedy is on the compliance roadmap: apply the overpromise check to every intent, while keeping the full-replacement advice substitution scoped to conversational turns.
In the meantime: any statement about expected outcomes, in any reply, is not a platform guarantee. The risk disclosure applies regardless of how a reply is worded.
Other things the reply layer does
Not compliance controls, but worth knowing so the behaviour is not mysterious:
Casual openers are stripped. "Sure," "Absolutely," "Of course," "Got it" and similar are removed from the front of a reply. The answer starts with the answer.
Markdown structure is preserved. Runs of spaces are collapsed within a line only, never across newlines. Headings, lists, tables and code blocks survive intact. An earlier version flattened everything to one line and inflated reply length.
A runaway reply is trimmed cleanly. There is a generous ceiling (around 6,000 characters) that essentially no real reply reaches. When it does trigger, the trim lands on the last sentence or paragraph boundary in the window. It never cuts mid-word and never appends an ellipsis.
What the persona does and does not authorise
The AI is instructed to educate and explain, and to say why each piece of a strategy is there. That is deliberate, a builder that assembles a strategy without explaining it teaches you nothing.
The bound on that instruction is the one to hold onto: explanation concerns strategy parameters, never instruments or trades. "A 1.2% stop is tight for this instrument's typical range" is explanation. "You should buy RELIANCE" is not, and is refused.
What this boundary is not
It is not a substitute for your own judgement. The AI declining to recommend a trade does not mean the strategy it built for you is a good one. It built what you described.
It is not a claim that the platform is risk-free. Every strategy can lose money. See Risk disclosure.
It does not extend to the research screens. The discovery suite names instruments and price levels. That is what a scanner does. Those screens carry their own research-not-advice framing; see Discovery & research.
It is not registration. No Stretus role holds Investment Adviser capacity. See Regulatory position.
Next
- AI limitations (what cannot be built, as opposed to what will not be advised
- Discovery & research) how the research screens are framed
- Regulatory position