Use this checklist when a human reviewer needs to inspect an AI paper trading agent before the next prompt change, risk change, watchlist change, or sample extension. The checklist keeps the review focused on evidence instead of confidence.
Trading Boy does not execute live trades, hold funds, or provide financial advice. This checklist reviews simulated paper-trading evidence only. A human reviewer can decide the next paper-mode action, but this checklist does not approve live trading, signals, copy trading, or real-capital deployment.
Use the checklist after the sample is closed. A pass means the paper evidence is clearer, not that the agent is safe, profitable, or ready for live execution.
| Review gate | Human question | Pass evidence | Fail action |
|---|---|---|---|
| Sample boundary | Was the sample closed by a planned date or decision count? | Start date, end date, version, and inclusion rule are visible. | Close the sample and label exclusions before review. |
| Prompt version | Can every decision be tied to one prompt and rule set? | Rows include the same version label. | Split mixed rows into separate samples. |
| Entry evidence | Do entries name setup, thesis, invalidation, and paper risk? | Each entry can be reviewed before outcome bias. | Tighten the output format before another sample. |
| Skip evidence | Do skipped trades name the blocked condition? | Skips cite unclear data, risk, invalidation, exposure, or setup mismatch. | Add a required skip-reason field. |
| Risk behavior | Did the agent respect paper size, exposure, drawdown, and pause limits? | Risk limits pass on entries and skips. | Run the risk control workflow. |
| Outcome review | Are winners and losers judged by rule fit before PnL? | The review separates process quality from simulated result. | Rewrite the review note before changing rules. |
| Next action | Is the next action narrow, paper-first, and testable? | One change, more collection, lower size, or retirement. | Reject broad prompt rewrites from thin evidence. |
Sample: A simulated AI paper agent produced 38 rows in a four-week review window: 15 entries, 17 skips, 4 exits, and 2 excluded records. The version label is consistent, and every row uses the same output format.
Human finding: Entries are clear enough to audit. Skips are incomplete because six rows say market unclear without naming the blocked condition. Risk behavior passes because simulated size and exposure stayed inside the written limits.
Checklist decision: The sample is not rejected, but it cannot justify a broad strategy change. The next paper version changes one field: every skip must name the blocked condition from the written rules. The setup, watchlist, size cap, and review cadence stay fixed.
Archive note: The reviewer saves the checklist beside the sample and links it to prompt versioning so the next review can compare skip quality before and after the change.
The AI trading agent decision review workflow explains how to inspect entries, skips, exits, and risk notes row by row. This checklist turns that inspection into a human signoff before any prompt or rule changes.
If the rows are inconsistent, use the AI trading agent output format template first. A human review should not have to invent missing fields after the sample closes.
The AI paper trading agent evaluation checklist scores the sample as a whole. This human review checklist focuses on whether a person can defend the next paper-mode action from the available evidence.
Pair both with paper-trading limitations so the review never implies that clean simulated records guarantee live performance.
Allow only a small set of outputs. The reviewer can keep collecting paper evidence, tighten one output field, reduce simulated size, add one skip condition, split the sample by regime, retire the workflow, or request a second human review. These outputs keep the system testable because each one maps to a specific paper-mode action.
Reject outputs that are too broad or too confident. Approved, ready, safe, autonomous, and profitable are not useful paper-review labels. A clean human review means the sample is easier to inspect. It does not prove live execution quality, liquidity, slippage, uptime, or future returns.
Archive the date range, version, reviewer, sample size, number of entries, number of skips, excluded rows, failed checklist gates, and exact next action. Also archive the reason for doing nothing when the checklist passes but the sample is still small. No change is often the most disciplined output because it preserves comparability for the next review.
The archive should link back to the AI paper trading agent workflow, the risk review, and the prompt-versioning record. That gives future reviewers a trail from setup to decision to human signoff without relying on memory or chat history.
It is a checklist a human reviewer uses to inspect simulated AI trading agent decisions, skipped trades, risk notes, sample quality, and next actions before changing a prompt or rule.
Review after a fixed paper sample, after repeated incomplete rows, after a risk-control breach, or before any prompt, output, watchlist, or size change.
No. Human review here applies to simulated paper-trading evidence only. Trading Boy does not execute live trades, hold funds, or provide financial advice.