Agent-Analyzed Copy Trading

Agent-analyzed copy trading lets you compare an ordinary copy-trader with an isolated, research-assisted twin. The original configuration keeps its settings and positions. The twin watches the same wallet, uses the same sizing and safety limits, and records its own cash, positions, decisions, and results.

What happens after a copied buy

The twin still mirrors the buy first, just as the original does. As soon as the token is known, it begins preparing current market, liquidity, contract or mint security, and prior target-wallet performance evidence while the fill completes.

Immediately after the fill, a fast safety gate checks objective evidence. A confirmed honeypot, inability to sell, non-transferable token, malicious creator, extreme sell tax, or critically thin liquidity can close the twin’s position without waiting for a model. Upgrade, freeze, mint, metadata, holder-concentration, new-pool, and other caution signals are not treated as proof on their own.

All other trades go to GPT-5.6 Sol as the expert adjudicator. Sol receives the prepared evidence packet, current market data, prior target-wallet outcomes, and the fast-gate findings, then uses web search for current exploits, scam reports, project information, material news, and credible warnings.

The result is one of three decisions:

• Keep leaves the position open.
• Close exits only the twin’s copied position when the empirically calibrated confidence meets that trade’s threshold.
• Uncertain leaves the position open because the evidence was not strong enough.

Sol’s stated confidence is calibrated only against completed earlier evaluation batches for the same twin. The current 50-trade batch cannot teach or grade itself. With little history or missing security evidence, the close threshold becomes more conservative. If the model, network, or research step fails, the position stays open and the failure is recorded. An analysis failure never triggers a sale.

Reading the comparison

Each twin still shows the original return, agent-analyzed return, and current difference in percentage points from the moment the pair started. That live snapshot is useful, but it is not evidence that the policy should be promoted.

The card separately shows:

• Learning evidence progress toward 200 matured trade outcomes.
• The held-out 95% edge interval when enough data exists.
• Agent-analyzed and original maximum drawdown.
• Modeled execution costs, reviewed/kept/closed counts, and analysis failures.

Open Show data to read the latest calibrated confidence, raw model confidence, close threshold, decision path, review summary, and source links.

Start in dry-run. A live twin performs a second real buy before analysis and may perform a real closing sale afterward, so it uses additional wallet funds, gas, and model/web-search usage. Creating a twin from a live source leaves the twin stopped until you explicitly start it.

How EVO uses the results

Every reviewed fill records both paths at fixed 5-minute, 30-minute, 4-hour, and 24-hour horizons: what the original hold would have returned and what the evolved decision returned. Both paths include modeled network, venue, slippage, and liquidity-impact costs. If the daemon is unavailable long enough to miss a horizon’s bounded observation window, that horizon is marked missed instead of substituting a later price.

The policy is frozen under an explicit version and evaluated in chronological 50-trade batches. Calibration for a batch can use only matured earlier batches. EVO does not mark a policy eligible until it has all of the following:

• At least 200 cost-aware outcomes matured through 24 hours.
• A complete unseen 50-trade validation batch.
• A paired bootstrap whose 95% lower confidence bound is above zero.
• Maximum drawdown no worse than the original path.
• No more than 5% failed-open Sol reviews in the held-out batch.

Failing a gate means not eligible yet; it never rewrites the original configuration. Passing all gates means eligible for promotion, not automatically promoted. A later product flow can deliberately create a new frozen evolved version from an eligible candidate.

EVO can improve the analyst prompt, evidence policy, calibration, and close policy in later experiments, while the source copy-trader, wallet signing rail, trade caps, and dry-run/live setting remain protected boundaries. This is walk-forward, evaluation-driven adaptation—not model-weight retraining and not a guarantee of better returns.

Requirements and privacy

The analyst prefers an existing ChatGPT OAuth connection and falls back to OPENAI_API_KEY from the shared HivemindOS environment when API-key mode is selected. OAuth tokens and API keys stay in runtime credential storage and are never written into a copy-trading configuration or review. OpenAI’s GPT-5.6 Sol model supports Responses API web search and structured outputs.

Token security reads use public GoPlus endpoints and fail open when evidence is unavailable; unavailable data alone never forces a sale. Web pages are treated as untrusted evidence. Their instructions are ignored, and only bounded review summaries, public source links, compact evidence flags, and evaluation outcomes are retained in copy-trading state.