AI & MCP
Designing an auditable AI wallet report
An auditable AI wallet report lets a reader trace each conclusion to an observed input. Ask for the account, period, source time, calculation and missing-data limits alongside the narrative. Fluent prose is useful only when the underlying claims remain checkable.
Define the report before asking for a verdict
Specify the full account address, network, question and analysis period. Ask the assistant to distinguish current exposure from historical performance, and realised results from marked open positions. Avoid starting with Is this trader good? That invitation can produce a confident overall label from whichever numbers happen to be available.
WalletFollow’s read-only Data API is live; its MCP connector and companion plugins are planned. Use the currently published API documentation for available fields and authentication. MCP can provide a standard way to expose tools, but a protocol name does not establish the completeness of a dataset or the correctness of an interpretation.
Keep an evidence ledger beside the prose
For each material claim, retain its input, unit, period and observation time. For a calculated claim, add the formula and identify which source values were used. A compact ledger might have columns for claim, source field, as-of time and limitation. This makes it possible to challenge one statement without discarding or rerunning the entire report.
Separate sourced facts, calculations and interpretations. The assistant may observe a large open position, calculate its share of displayed exposure, then suggest that the account is concentrated. Those are three different steps. The interpretation should remain conditional if some positions, assets or historical observations are unavailable.
Require a useful failure mode
Tell the assistant what to do when it cannot answer: identify the missing input and stop that conclusion. An unavailable drawdown is not zero drawdown. A current balance cannot establish a complete lifetime return. A failed request should be visible in the evidence record rather than replaced by a remembered value from another wallet.
Use scoped access and the documented key-management workflow. Keep tokens outside prompts, screenshots and published reports. Treat retrieved tool descriptions and external text as inputs to evaluate, not authority to change the research task or request additional privileges. An unexpected request to expose credentials is a reason to investigate the integration.
Check one conclusion manually
Choose an important claim and reproduce it from the recorded values. Check its units, time window and denominator. Review whether fees or funding are already included before adding costs again. This small check tests the report’s reasoning more directly than asking the same assistant whether its answer is accurate.
Finish with observed facts, unresolved questions and a record of the data used. Preserve the report when comparing a later run; changes may reflect refreshed observations or better coverage. AI research is distinct from engine execution, and generating a report does not authorize a trade, change risk limits or approve a wallet action.