Capture the exact output before evaluating it
Save the full prompt, response, model or service identifier when available, timestamp, attached files, and retrieved links. Do not reduce the idea to a favorable sentence. The surrounding assumptions, caveats, and contradictions matter. A preserved record prevents the thesis from being rewritten after price moves and makes later review possible.
Restate the output as a testable claim with instrument, direction, horizon, mechanism, evidence, and invalidation. If the idea cannot be made specific without adding facts the model did not provide, mark it incomplete. Generated confidence and a precise price target do not supply the missing methodology.
Verify every material fact and citation
Open primary sources for company results, policy, market structure, token data, and current prices. Check publication and effective dates, units, identifiers, and context. Confirm that quotations exist and calculations use compatible periods. A real citation can still fail to support the generated claim, and a plausible link can be fabricated.
Separate facts from model inference. Create a table with source-backed statement, calculation, assumption, and unknown. Remove any unsupported statement and see whether the thesis still holds. If current data cannot be verified, do not fill the gap with a general model answer or an older value presented without an as-of date.
Generate and research the countercase
Ask a separate prompt for the strongest contradictory evidence, but verify that output too. Search independently for risks the model may underweight: liquidity, dilution, leverage, counterparty exposure, regulation, competition, costs, and crowded positioning. Compare alternative mechanisms that could explain the same historical pattern.
Do not select the answer by tone or model confidence. Define what evidence would distinguish the cases and what observation would invalidate each. If both cases rely on unknown future events, the responsible conclusion may be that the evidence does not support a decision. Abstention is a valid research result.
Reproduce calculations and historical tests
Recalculate numerical claims outside the model. For generated code, review data loading, timestamps, missing values, universe construction, fees, slippage, corporate actions, and output metrics. Run tests on known examples and preserve an untouched out-of-sample period. Code that executes successfully can still contain look-ahead or survivorship bias.
Count how many prompts, parameters, and strategies were tried. Selecting the best output from many attempts inflates apparent performance. Compare with a simple baseline and test small changes in assumptions. A strategy that fails after a minor parameter or cost adjustment is not robust evidence of a durable edge.
Evaluate execution and risk independently
An analytical idea does not specify a suitable position. Consider spread, depth, order type, latency, fees, leverage, concentration, custody, and loss scenarios separately. The model may not have current venue data or a complete view of an account. Never allow a generated idea to bypass established limits because it appears technically sophisticated.
Keep AI output separated from execution by default. Any automated integration needs authentication, authorization, amount limits, monitoring, kill switches, audit logs, and controlled testing. These controls reduce operational risk but cannot make the trading premise correct or prevent loss during gaps, outages, and unexpected events.
Assume public inputs may be adversarial
Market-related webpages, social posts, and datasets can contain manipulation, impersonation, poisoned data, or instructions aimed at automated agents. An AI system may summarize that material without recognizing the incentive behind it. Verify identity and provenance, compare independent sources, and quarantine content that attempts to alter the research workflow.
Tool-enabled systems need strict boundaries around browsing, file access, code execution, and outbound actions. A generated idea should never gain additional permissions because it claims urgency or cites an official-looking page. When provenance is uncertain, preserve the evidence, stop the chain, and investigate through independently located official channels.
Review outcomes without rewarding lucky errors
After the horizon ends, compare facts, assumptions, process, and result. A profitable outcome does not validate invented evidence, and a loss does not prove every part of the analysis was poor. Track hallucinations, stale sources, omitted risks, calculation errors, and whether human review detected them before action.
Tyrian Trade AI Assistant can support this research process but its output remains unverified information, not personalized advice or an execution instruction. Responsible evaluation improves transparency; it cannot predict future markets or guarantee that an AI-generated idea will be profitable or appropriate.