How a brain is computed

Data & accounting

Wallet history comes from Solana mainnet (Helius when configured, otherwise standard RPC). Each transaction is normalised from the wallet's own balance changes into buys and sells; tokens are enriched with launch time, launchpad, supply and market data (Jupiter / DexScreener), and SOL is priced at the minute of the trade.

• Window: statistics cover the trailing 90 days.
• Unit of analysis: a position = one wallet × one token, from first buy until the balance returns to ~0.
• Transfers are excluded from trade stats (airdrops, internal moves, phishing airdrops).
• Minimum sample: no archetype, score or tier below 20 trades and 5 distinct tokens — the brain is shown as “Unrated — thin sample”.

Metrics

Archetypes

One primary archetype per brain: the first row that matches, top to bottom. The triggering numbers are shown next to the label.

Flags

Overlay flags are added for every rule that applies.

Copyability

A 0–100 estimate of how much of the wallet's edge survives being followed with realistic latency (≈3 seconds) and slippage (≈5%). A weighted quality score — entry market cap (20%), profit per position (20%, thin profits are eaten by slippage), hold time versus latency (18%), execution pace (14%, bot-like volume scores low), edge type (16%) and shrunk win rate versus the population (12%) — is multiplied by a fidelity gate of 0.25 + 0.75 × f, where f is the worse of the hold-time and pace scores (0.7 when neither is known). The gate comes from our copy-trading directory work: a wallet a follower physically cannot mirror cannot score well, however much it earns, while a wallet holding ≥10 minutes at a manual pace keeps its full quality score. Many top wallets hold for seconds: a low score is the honest result, not a criticism.

Track score & tiers

trackScore (0–100) = stop-loss discipline 34% (1 − share of tokens below −50%) + profit share 28% + ROI 16% (−5% → 0, +30% → 100) + win rate 10% + sample size 12% (20 → 320 tokens). Dev-linked wallets are multiplied by 0.45. Unrealized PnL is never used. Tiers set the default seat supply and price; an admin can override pricing per brain, and supply never drops below seats already minted.

The win rate in both scores is shrunk toward the population before it is used: (wins + p × 20) ÷ (closed positions + 20), where p is the pooled win rate of every closed position across tracked brains. This is the posterior mean of a Beta prior worth 20 positions, so a wallet needs about 20 closed positions before its own record outweighs the assumption that it is average: 3 wins out of 3 stays close to p, while 120 out of 200 stays close to 60%. The shrunk rate is scored as lift over the population (p − 20 points → 0, p → half, p + 20 points → full weight). The raw win rate is still what the brain page shows.

Supply caps exist for capacity and fairness — alert fan-out, front-running and coaching load — not to create scarcity value. Seats are term-limited memberships with no revenue share, yield or buyback for holders.

Live alerts

Every new trade produces an alert with deterministic facts only:

• side, size in SOL/USD, market cap, entry delay after launch, launchpad and venue;
• position context: first buy or add, share of the peak position sold, realized result on sells, time held;
• in-character comparison against this wallet's own medians (e.g. “4.1× usual size”, “entered much later than usual”);
• safety context from token data: pool liquidity, top-holder concentration, mint/freeze authority, and the RugCheck summary (risk 0–100 where higher is riskier — ≤30 low, ≤60 medium, above that high — share of LP locked, and named risks). RugCheck results are cached for 6 hours; where RugCheck's risk wording contradicts the token's on-chain mint/freeze authority, the on-chain authority wins. LP lock is not shown for tokens still on a launchpad bonding curve, which have no LP yet;
• swarm context when the buy is one of several tracked brains entering the same token (below).

A short AI explanation then describes what the trade means within the wallet's pattern. It is constrained to the facts above and never recommends buying or selling. Claimed brains may have an alert delay; the public trade log on each brain page shows trades only after one hour.

Say vs do (claimed brains only)

Tweet-to-trade correlation runs only for brains claimed by the wallet owner who also linked their X account, via the official X API. For each bullish post mentioning a token, an event is recorded when all hold: position ≥$500 at the post, first buy ≤72h before it, market cap ≥1.5× the average entry, and ≥50% of the position sold within 24h. It is marked as a fast exit if ≥80% was sold within 1h or the first sell came within 15 minutes. The “Promote-then-sell pattern” flag needs ≥3 events in 90 days, an event rate ≥40% of pre-positioned bullish mentions and a lift ≥2.0 versus the wallet's own exits after comparable price moves without a post. Each event links the post and transaction signatures.

Anti-scalping: if a claimed owner sells ≥50% of an alerted position within 60 minutes after the alert published, their revenue share is held pending review and a notice is shown on the brain page.

Swarms

A swarm is three or more tracked brains making their first buy of the same token within 10 minutes of the first of them. It is “forming” at 3–4 brains, a “swarm” at 5–9 and “large” at 10 or more. Each brain counts once; a later, separate wave into the same token forms its own swarm. Entries that land in the same Solana slot are kept but marked, because near-simultaneous entries may be one operator or a bundle rather than independent decisions; the count of distinct slots is shown next to the member count. Swarms are detected the moment a tracked trade is ingested and rebuilt from full trade history every 10 minutes, so newly tracked wallets add to past swarms too.

For context we show the base rates measured by the Narrative Radar study (Aug–Oct 2026) on pump.fun KOL clusters: ≥5 KOLs in 10 minutes preceded a runner (≥$100K market cap, farm launches excluded) in about 15% of cases, ≥10 KOLs in about 50%. These are external figures, not ours.

Our own outcome for each swarm uses only prices we observe: the highest market cap in any tracked brain's trade of the token in the 24 hours after the swarm started, against the market cap at its first entry, and separately against the market cap when the swarm became detectable (the third entry) counting only trades after it. Because we only see a price when a tracked wallet trades, the peak is a lower bound. Aggregate rates use swarms whose 24-hour window has closed. Neither our figures nor the study's are a prediction or advice.

The token in a swarm is hidden from the public for one hour after the last member joined; holders of an active seat on any member brain see it in real time. Member brains and timing are always shown.

Connectome

The connectome maps how tracked wallets tend to enter the same tokens, borrowing the vocabulary of brain wiring diagrams (FlyWire, neuPrint). It describes historical co-movement among tracked wallets only; it does not show who knows, follows or trades on information from whom. Recomputed every 10 minutes over the last 30 days.

• Neuron: one brain (wallet). An entry is the wallet's first-ever buy of a token, counted when it falls in the 30-day window.
• Synapse A → B (“B tends to enter after A”): for every token both entered, with B's entry 0–60 minutes after A's, add idf(token) × e^{−Δt/15 min}. Entries in the same second or within 2 slots (~1 s) are treated as near-simultaneous (bots, bundles, a shared trigger) and carry no direction. idf(token) = ln((N + 1) / (n + 1)), with N the wallets active in the window and n the wallets that entered that token, so a token everyone bought adds nothing and a rarely shared token adds most.
• Significance: the same weight is recomputed under 100 shuffles that permute entry times within each token among its entrants (keeping token popularity and each wallet's activity, breaking the order). A synapse is kept only when A→B shares ≥3 tokens and its weight is ≥3 standard deviations above the shuffled mean. Each synapse shows its co-entry count, median lag and example tokens older than one hour.
• Region: Louvain communities on the undirected graph (A→B and B→A weights summed), with a fixed random seed so regions are stable between runs. A region is named after its hub (the member with the highest influence) plus its dominant trait: “fast-entry cortex” when members' median entry delay after launch is under 60 s, otherwise “lead nucleus” when their median lead share is ≥0.6, otherwise “cluster”. Wallets without significant synapses form their own one-neuron region, shown as “Unwired · name”.
• Influence (PageRank): weighted PageRank (damping 0.85) on the reversed graph, so rank flows from wallets that enter later to the wallets they entered after; scaled so the top neuron is 1. Lead share = outgoing ÷ (incoming + outgoing) synapse weight.
• Input share of A→B = its weight ÷ all synapse weight into B. The signal cascade from a wallet follows the FlyWire information-flow model: starting from that wallet, any wallet whose summed input share from already-reached wallets is ≥0.3 is reached at the next step; the arrival time adds up median lags. The strongest path between two wallets minimises Σ −ln(input share), i.e. the chain where each step explains the largest part of the next wallet's entries.
• Firing: live trades as they are ingested, shown per wallet as buy/sell and a size bucket (<1, 1–10, >10 SOL), without the token. Token replays and synapse examples use trades older than one hour.
• Token wiring on a brain page: its top 120 tokens by bought USD in 30 days, and capital rotations — a sell of token A followed by the wallet's next buy of a different token within 3 minutes.

Two wallets reacting to the same post, call or listing look like a synapse when one is merely faster; reciprocal pairs (both directions of similar weight) usually mean a shared external trigger.

Agents

Seat holders can run an agent: an automated trader that trades memecoins in the style of a brain. It never copies the brain's wallet or its trades; it hunts its own entries with rules derived from the brain's profile, deterministically, with every rule linked to the statistic it came from:

• Entries: launchpad (pump.fun when ≥60% of the brain's tokens were pump.fun launches), phase (bonding curve when ≥60% of its pump.fun entries were on the curve, pools when ≤20%), market-cap band (25th–75th percentile of its entry market caps), age band (25th–75th percentile of its entry delay after launch), a momentum gate for momentum styles, and a demand gate: net SOL into the curve (on-chain reserves) or 5-minute buys/sells for pool tokens. A number we can't measure yet is unknown, never zero, and blocks the entry.
• Safety: mint and freeze authority revoked (read on-chain), top-10 holders excluding the curve or pool, RugCheck score. A brain flagged rug-exposed gets a stricter RugCheck limit.
• Sizing: a share of free SOL per entry from the brain's sizing style (fixed, variable, conviction), capped at 1 SOL per position and a number of open positions.
• Exits: time stop at the brain's 75th-percentile hold, a stop-loss tightened or widened by how often the brain cut losses, a take-profit (laddered when the brain usually sold in steps), initials-out and a trailing stop where its exits show them. Exits never act on a price older than 30 seconds; a token without a price for 15 minutes is valued at zero and sold if any route exists.
• Judge (optional): when the rules pass, a fast language model reads the brain's playbook, the rules and the candidate's numbers and answers buy or skip in at most 30 words. It has an hourly budget; without it the rules decide alone.

Paper mode (default) fills at a real Jupiter quote, less a latency haircut (1.5% on the curve, 0.5% in pools), the network and priority fee and token-account rent, so paper results are not flattered. Live mode is opt-in: the server holds the agent's key (encrypted at rest), signs Jupiter swaps and records the actual amounts from the confirmed transaction. Every order is keyed so it can never be sent twice. Risk limits: trades per hour, open positions, a daily loss limit that halts the agent, a per-agent deposit limit (excess deposits halt the agent until withdrawn) and platform kill switches. The owner can withdraw to their own wallet at any time.

Every candidate an agent evaluates is recorded on its public tape with the reason: all buys and sells, every near miss (one rule short) and a 1-in-30 sample of far misses. Decisions informed by a fresh swarm are shown to the public one hour later, like the swarm itself. Leaderboards rank paper and live agents separately, by return on starting capital. Agents are an experiment: memecoins are extremely risky, most lose money, and an agent's results say nothing about the brain's own wallet.

Brain Picks

A pick is a token that matches a brain's trading style right now, found by the same rules the brain's agents use. Picks never place orders: each one is tracked on paper against live prices and closed by the style's exit plan, and that paper outcome is the brain's public pick track record. A pick says a token matches this wallet's style; it is not a recommendation and not financial advice.

• Universe: picks are generated for brains with at least one active seat, at least one agent, or a place in the Hive's style universe, and only when the profile is rich enough to derive a style.
• Market rules: the same deterministic entry rules as the brain's agents (see Agents): launchpad, phase (curve or pools), entry market cap within the brain's 25th–75th percentile, entry delay after launch, liquidity, momentum, the demand / flow gate and swarm confirmation where the style asks for it.
• Safety rules: mint and freeze authority revoked, top-10 holders limit, RugCheck score — again identical to the agents' checks.
• Council: cross-brain context from every tracked brain — an active swarm on the token, tracked-wallet buys and sells in the last 30 minutes, and connectome leaders among the buyers, shown as chips on the pick. A token is vetoed when two or more tracked brains sold it and none bought.
• One-liner (optional): a language model may add one line on why the token matches the style, labelled with the model. It runs on an hourly budget and never decides whether a pick is made.
• Limits: one pick per brain and token per 24 hours, at most 6 picks per brain per hour.
• Style size: the style's position size as a share of a bankroll, from its sizing rules (base size, up to the maximum by conviction).

Paper outcome: from the price at the pick, the style's exit plan is simulated on live prices — stop-loss, partial sells on the take-profit ladder, initials-out, a trailing stop after the first take-profit and the time stop. A token with no price for 15 minutes is closed as dead with its remaining value at zero. Fees and slippage are not modelled, so real fills would do worse. The outcome is the result as a share of the paper stake.

Seat holders (and the brain's owner) see every pick the moment it qualifies. Everyone else sees picks one hour later; until then the token is hidden, and the public track record counts only picks older than one hour. The Brain Picks leaderboard ranks brains by average paper outcome per closed pick, with at least five closed picks. Past paper outcomes don't predict future results; memecoins are extremely risky.

The Hive

The Hive is KOL Brain's own public showcase agent. It runs on the same engine as seat holders' agents, but instead of one brain's style it draws on all of them, and it is funded by the pump.fun creator fees of the KOL Brain token. Its wallet, every decision, order, style switch and fee claim are public. Its profits stay in its own wallet: nothing is distributed, bought back or paid to token holders, and holding the token gives no claim on the Hive or its results.

• Style league: the Hive paper-trades the derived styles of up to eight top-rated brains in parallel (strongest track records, at most a third per archetype), each with the Hive's own limits, and scores every style over the last 72 hours: return including open positions, minus half the deepest drawdown, plus a small win-rate term. Thin samples are pulled toward the brain's track record and a style needs three closed trades before the Hive will follow it; a small exploration bonus gives styles with few trades a chance to prove themselves.
• Switching: the Hive trades the best-scoring style and changes only when another style leads by at least two percentage points and the current one has been in charge for six hours, so it does not chase noise. Every switch is published with its reason and the full score table.
• Council: before an entry, the Hive checks what every tracked brain is doing on that token: an active swarm, tracked brains buying or selling in the last 30 minutes, and whether connectome leaders are among the buyers. This adjusts position size, can veto a token tracked brains are leaving, and is given to the judge with the active style's playbook. Decisions that use this live, seat-gated information are shown to the public one hour later.
• Risk: a per-position cap, a maximum number of open positions, a daily loss limit that halts entries until the next UTC day, a gas reserve that is never traded, and kill switches. It trades on paper until the platform switches it to live.
• Funding: the token's creator fees are claimed into the Hive wallet on a schedule; each claim is recorded with its transaction. A public reconciliation compares the wallet's on-chain balance with the recorded deposits, fee claims and trades, and lists any transaction the records do not explain.

Limitations

• Only wallets that are tracked are analysed; positions held in other wallets are invisible.
• Sells into DCA or limit orders, airdropped tokens and wallet-to-wallet moves can distort individual positions.
• Market caps and prices for very new tokens can be missing or approximate at the moment of trade.
• Past behaviour does not predict future results. Survivorship bias applies to any leaderboard.

Spotted an error? Use the disputes form.