Agent-Analyzed Copy Trading

Agent-analyzed copy trading runs an AI-reviewed twin next to an ordinary copy-trader so you can compare them. Both watch the same wallet with the same sizing and limits. The twin reviews each copied buy and may close it early; the original keeps its settings and positions.

Local or always on

The Copy trading workspace runs copies on your desktop, or you can use the always-on option at the top. For always-on, choose HivemindOS wallet to create a hosted Base wallet from your HivemindOS account, or connect an existing Bankr wallet. Both use the same target detection, limits, verified execution ledger, price, and reporting.

Before you confirm, HivemindOS shows the one-time hosted-wallet setup price alongside the copy-trading price. The wallet starts paused until you fund its address with Base USDC; then resume the monitor. The first $1 covers a rolling 30-day monitoring period. There is no fee on trades: a Live sale that locks in new profit shares up to 10% of it (less with more copied volume or HIVE staked), and losing, failed, skipped, paper, and unverified trades pay nothing. Canceling stops the monitor and revokes its trading access.

For hosted monitor setup, fees, and status, see Bankr Copy Trading.

Start in Paper, then promote

New desktop copies start in Paper, with simulated money.

1. On a Basic or Agent copy, choose Promote to Live.
2. Review the wallet, trader, sizing, limits, and whether trader sells are followed.
3. Confirm the real-money risk, or choose Keep Paper.

Promotion starts a separate Live run and stops the Paper copy. Paper results stay as history but never become real holdings. Experimental comparisons stay Paper only. Keep the desktop app running for local copies to follow new trades. Live copies pay trading and network fees.

One budget across both copies

When the ordinary copy and its twin both run, they split one trade amount and one daily limit. By default each gets half, so a $10 copy becomes two $5 entries. If one side stops, the other gets the full amount again.

New buys must fall within the copy’s age limit and happen after its current Live run starts. An unconfirmed trade blocks the next one until it is resolved; restarting never turns an unknown result into a new order.

Paper and live copies see each new Base block at the same moment, so paper results don’t gain or lose time by being simulated. Paper still models network, venue, slippage, and liquidity costs.

What the twin does after a buy

1. It mirrors the buy, just like the original.
2. It gathers market, liquidity, token-security, and past target-wallet evidence while the fill completes.
3. A fast safety check can close the twin’s position immediately for objective red flags: a confirmed honeypot, inability to sell, a non-transferable token, a malicious creator, extreme sell tax, or critically thin liquidity.
4. Everything else goes to GPT-5.6 Sol, which also searches the web for exploits, scam reports, and news.

The result is Keep, Close (only when calibrated confidence meets that trade’s threshold), or Uncertain (the position stays open). Confidence is calibrated only on earlier completed batches of 50 trades. If the model, network, or research fails, the position stays open; a failed analysis never triggers a sale. Slow reviews don’t delay the next wallet check.

Read the comparison

Each pair shows the original return, the twin’s return, and the difference since the pair started. That snapshot alone isn’t a reason to switch. The card also shows:

• Learning evidence progress toward 200 matured trade outcomes;
• held-out 95% edge and absolute 95% return intervals, and profit factor;
• maximum drawdown for both sides;
• modeled costs, reviewed/kept/closed counts, and analysis failures;
• Outcome retrospectives explaining what happened.

Open Show data for the latest confidence, close threshold, decision path, and sources.

A Live pair splits the configured amount across two real buys, and the twin may make an extra closing sale, which adds fees. A twin created from a Live copy starts stopped in Paper; promote it separately to use real money.

When a policy qualifies

EVO records both paths at 5 minutes, 30 minutes, 4 hours, 24 hours, and the target’s own exit, including modeled costs. It evaluates the policy in frozen, chronological 50-trade batches. A policy is only eligible for paper promotion when it has:

• at least 200 cost-aware outcomes matured through 24 hours;
• a complete unseen 50-trade validation batch;
• a paired bootstrap 95% lower bound above zero, and an absolute-return 95% lower bound above zero;
• a held-out profit factor of at least 1.2;
• maximum drawdown no worse than the original and no greater than 25%;
• no more than 5% failed-open Sol reviews in the held-out batch.

Passing never promotes a policy automatically and never turns on live trading. Real-money use needs your separate decision and a new safety review. EVO can refine the analyst prompt, evidence, calibration, and close policy, but never the source copy-trader, wallet, caps, or live setting. No test proves future profits.

Replay history in the strategy lab

The local strategy lab replays completed paper outcomes instantly, using only what was knowable at each fill. It proposes candidates on earlier trades and tests one on the latest untouched holdout. A candidate can change exactly one allowed close rule.

node scripts/copy-trading-strategy-lab.mjs --holdout 8 --purge 1 --candidate-limit 4 --agent-proposals --record-harness

A handful of trades is only exploratory. A result is paper-review-eligible only with at least 50 untouched holdout trades, a positive paired 95% interval, positive absolute return, no worse drawdown, and complete replay parity. The lab never edits your running copy or enables live trading.

Current agent versus hybrid paper test

From an agent-analyzed copy, choose Create current-agent vs hybrid paper test to add a new paper-only comparison. Both sides start fresh with the same target, sizing, limits, and cash. The hybrid reuses the current agent’s exact review, so only the exit rule differs: when the normal close rule doesn’t fire, it may trim 50% of a newly copied fill if Sol recommended closing at or above a 0.659 confidence floor or the review failed open. It trims only when security evidence is complete and shows a risk score of at least 20 or a risk flag. It is an experiment, and promotion needs the same evidence gates as above.

Requirements and privacy

• The analyst uses your ChatGPT sign-in, or OPENAI_API_KEY from your shared HivemindOS environment in API-key mode. Credentials are never written into a copy-trading setup or review.
• Token security checks use public GoPlus data. Missing security data alone never forces a sale.
• Web pages are treated as untrusted evidence; their instructions are ignored, and only short summaries, source links, and outcomes are kept.