Launch Lumiere

Trust and verification

How to evaluate AI trading-tool claims before connecting money or a broker

Treat guaranteed returns, fixed win rates, risk-free automation, unexplained “AI,” missing source provenance, hidden costs, pressured deposits, and unclear account permissions as reasons to stop and verify—not as reasons to move faster.

Reviewed by Lumiere Trading Mentor

Direct answer

What claims should you distrust when evaluating an AI trading tool?

Distrust any claim that removes uncertainty: guaranteed profits, a fixed or near-perfect win rate, “risk-free” returns, effortless income, or a model that supposedly predicts sudden market changes. Pause when the seller cannot explain what the tool does, identify current sources, show a test methodology with costs, state every account permission, distinguish read-only analysis from order submission, or provide a clear way to revoke access. Verify the company, people, claimed registrations, evidence, total costs, and broker authority independently before connecting an account or funding anything.

Who this checklist is for—and what it cannot decide

This checklist is for self-directed traders comparing research, chart-analysis, pre-trade review, signal, or auto-trading products, especially when a tool may connect to a brokerage account. It helps turn a marketing claim into specific evidence requests.

It cannot certify a provider, determine whether a product is lawful in every jurisdiction, prove that a registered person is trustworthy, or make an investment suitable. Educational software, research tools, investment advisers, brokers, exchanges, and auto-trading services can have different legal roles. Identify the role actually being performed and verify the applicable regulator rather than assuming every trading tool must hold the same registration.

Eight red flags and the evidence to request

Claim or conditionWhy it mattersWhat to request
Guaranteed or risk-free returnsAll trading involves risk; certainty language hides downside and can be a classic fraud warning.Do not proceed on the promise. Ask for the legal entity, offering documents, and independently verifiable record.
Fixed or extraordinary win rateA rate without period, sample, benchmark, losses, costs, and out-of-sample method is not decision-usable evidence.Dated methodology, complete sample, drawdowns, fees, spreads, slippage, exclusions, and reproducible results.
“AI-powered” without a defined jobThe label can exaggerate or misstate what the system actually does—often called AI washing.The exact task, inputs, outputs, human review, failure modes, evaluation period, and material limitations.
No source or timestampModel output can be inaccurate, incomplete, outdated, misleading, or invented.Named underlying sources, freshness, market session, fallback behavior, and a way to inspect the cited evidence.
Unclear broker permissionsReading market data is materially different from viewing balances, changing settings, or submitting orders.A permission inventory: data read, account read, order create/modify/cancel, withdrawal ability, paper/live scope, and revocation steps.
Auto-trading sold as effortlessAuto-trading can let a third party send instructions directly to a brokerage account for immediate execution.Entity registration where applicable, order authority, strategy limits, kill switch, incident process, and independent broker confirmation.
Returns shown before total costsSubscriptions, commissions, spreads, slippage, financing, taxes, and market impact can change the result.Net results under disclosed cost assumptions and a complete recurring and transaction-cost schedule.
Urgency, secrecy, or deposit pressurePressure reduces time for independent checks and can be used to push money toward an unverified entity.Stop. Verify identity, domain history, contact details, custody, withdrawal rules, and regulator records independently.

The eight-step due-diligence sequence

  1. Identify the legal entity and role.Record the company name, people, domain, location, custody arrangement, and whether the product claims to provide education, research, advice, brokerage, exchange, signal, or auto-trading services. Verify claimed registrations on the regulator's own site.
  2. Separate analysis from authority.Write down whether the tool only displays information, reads broker context, drafts an order, or can create, modify, cancel, or transmit an order. Confirm Paper versus Live scope and who performs the final action.
  3. Interrogate every performance claim.Reject guarantees. For any historical result, require the period, full sample, benchmark, drawdown, losing observations, selection rules, out-of-sample method, and costs. A screenshot or a few winners is not a performance record.
  4. Test the AI claim.Ask what AI changes, which inputs it uses, how the output was evaluated, what a failure looks like, and when a human must intervene. Product language should match the capability actually implemented.
  5. Inspect provenance and freshness.Require named sources, timestamps, session and time-frame context, missing-data behavior, and fallbacks. Check consequential information against multiple independent sources.
  6. Inventory permissions and data.List every credential, account field, order capability, storage location, retention period, subprocessor, and revocation path. Grant the minimum necessary permission and avoid a live account when a paper or sandbox test is available.
  7. Recalculate after all costs.Include subscriptions, spreads, commissions, slippage, financing, taxes, withdrawal conditions, and the cost of errors. Compare like-for-like periods and do not turn a backtest into a forward promise.
  8. Run a reversible test.Use a paper, sandbox, or read-only environment first; test access removal and support response; document a stop condition; and independently confirm that revocation worked before considering money or live permissions.

Broker-connection questions for an IBKR user

Before enabling any connection, ask for a one-page permission map. It should answer: What account is visible? Is the session Paper or Live? Which market-data and account fields can be read? Can any order be created, modified, cancelled, or transmitted? Where do credentials remain? What network path is used? How is access revoked? What happens when data is stale, missing, or contradictory?

A read-only connection is a narrower boundary, not proof of accuracy, security, performance, or suitability. Paper trading is also a simulator and does not establish live execution quality. Use the IBKR Paper TWS setup guide to inspect Lumiere's read-only connection boundary and documented simulator limitations.

How Lumiere applies the same checklist

This is a public product-boundary check, not a third-party certification or a claim that Lumiere is risk-free.

QuestionCurrent public boundaryWhat it does not prove
Returns or win ratesLumiere does not promise profits, a win rate, loss prevention, or suitability.That a review will improve a trading outcome.
What the AI doesThe hosted product organizes source context, chart evidence, invalidation, risk arithmetic, and a bounded paper plan for user review.That generated analysis is correct, complete, current, or a recommendation.
Source provenanceThe review surfaces provider, request status, freshness, fallback, and evidence labels where available.That a source is error-free or independently sufficient.
Broker accessThe optional Local Connector is limited to read-only IBKR Paper TWS market context in V1.That local configuration, entitlements, network security, or data availability will be identical for every user.
Order authorityLumiere does not transmit customer paper or live orders in V1; the trader independently controls any broker action.That a user's later order will fill at a planned price or stay within planned loss.
Data and privacyThe public Privacy Policy separates anonymous operational analytics from optional product and research telemetry.A universal privacy or security certification.

Official sources behind this checklist

The following U.S. regulator sources support the red-flag and verification principles above. Their warnings do not establish that a particular product is fraudulent, and they do not replace jurisdiction-specific professional advice.

For one dated illustration, FINRA's July 29, 2025 auto-trading warning described promotions claiming consistent monthly returns of more than 10 percent. That is an example of the promotional language FINRA flagged, not a threshold below which a return claim becomes safe, credible, or suitable.

CFTC source reviewed August 4, 2026: Customer Advisory: AI Won't Turn Trading Bots into Money Machines.

FINRA, SEC, and NASAA source reviewed August 4, 2026: Artificial Intelligence and Investment Fraud.

FINRA source reviewed August 4, 2026: Risks of Auto-Trading Services Offered by Unregistered Entities.

SEC enforcement source reviewed August 4, 2026: SEC Charges Two Investment Advisers with Making False and Misleading Statements About Their Use of Artificial Intelligence.

Investor.gov source reviewed August 4, 2026: Protect Your Money: How to Avoid Investment Scams.

Inspect the workflow

See how a source, limitation, invalidation, and risk boundary appear in one illustrative review.

Open the sample reviewUse the private worksheet