Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/alphaleaks60-maker/docs2/llms.txt

Use this file to discover all available pages before exploring further.

Pump.fun is adversarial by design. A significant share of activity on the platform is not organic price discovery — it is coordinated manipulation, bot noise, and deliberate traps set for retail participants. Alpha Leak runs a dedicated layer of adversarial detectors that identify these patterns in real time, ensuring the trading system never acts on manufactured signals and can exit positions the moment a trap is confirmed.

​
Threat coverage

ThreatDetectorML featureResponse
Coordinated bundle buyBundleDetector (5s time windows)token_bundle_confidenceToken flagged; anti-signal trigger at >70%
Bot noiseBotDetector (3 behavioural patterns)is_botWallet alpha capped at 30
Serial ruggerCreatorRiskScorer (rug rate, insider pct)creator_risk_scoreAnti-signal trigger at score >80
Copy-trade followerCopyTradeDetector (directional consistency)is_copy_traderSignal downweighted; originator resolved
Wash tradingToken risk scorer (buy/sell pattern analysis)wash_trade_pctAnti-signal trigger at >30%
Exit liquidity trapAntiSignalEmitter (smart money selling)—Anti-signal emitted; open position force-exited
Signal crowdingSignalCrowdingDetector (tracked SOL ratio)crowding_ratioScore penalty up to 50%

​
Bot detection

The BotDetector runs every 30 minutes and classifies wallets across three distinct behavioural patterns. Each pattern is independently sufficient for bot classification.
Trigger: 20+ distinct tokens bought within a 2-hour window.This pattern is characteristic of broad sniping bots that buy every new token regardless of quality. At 20+ tokens per 2 hours, no human is making individual buy decisions — the wallet is running automated logic against the token creation stream. These wallets produce high apparent volumes but carry no predictive signal for token success.
Detected bots are written to the wallets table with their bot_type, added to the known_bots Redis set, and capped at a maximum alpha score of 30. The is_bot field is a direct feature in the ML model, allowing the model to learn independent patterns correlated with bot activity.

​
Bundle detection

Bundles are the most common form of coordinated manipulation on Pump.fun. A group of wallets — often controlled by the same operator — buys a token within the same few seconds, typically at or near launch, creating the appearance of organic momentum while actually front-running the curve for a later dump. The BundleDetector groups trades into 5-second time buckets and identifies clusters of 3 or more wallets buying the same token within the same window. Each cluster is scored on two independent dimensions:

Amount uniformity

Coefficient of variation of buy sizes within the cluster must be below 0.3. Random wallets rarely buy the exact same SOL amount — uniform sizing indicates a single operator distributing capital across wallets.

Rank continuity

The span of buy ranks within the cluster must be less than or equal to the wallet count. Consecutive buy ranks in a tight time window indicate the wallets transacted in rapid succession from the same source, rather than arriving independently.

​
Bundle confidence classification

ClassificationCriteriaConfidence range
same_slot_coordinatedBoth amount uniformity AND rank continuity met0.7–1.0
similar_amountsAmount uniformity only0.5–0.7
time_windowTime proximity only (3+ wallets, same 5s window)0.3–0.5
The confidence score feeds into the ML model as token_bundle_confidence and into the AntiSignalEmitter as an independent trigger. A bundle confidence above 0.7 combined with a bundle buyer percentage above 30% is sufficient to trigger an anti-signal on its own.

​
Creator risk scoring

Not all creators are equal. The CreatorRiskScorer builds a longitudinal view of every creator with 2 or more tokens on-chain, combining rug rate, velocity, and insider presence into a single risk score. The most predictive metric is the rug rate: the fraction of a creator’s tokens whose last trade occurred within 10 minutes of launch and which never graduated. A creator with a 90% rug rate has a consistent and deliberate pattern of launching tokens that die immediately — typically because the creator or insiders dump before any retail participation arrives. Additional signals computed per creator:
  • Serial velocity — tokens launched per day over the last 30 days. High velocity combined with a high rug rate is an extremely reliable indicator of malicious intent.
  • Average insider presence — the average count of insider wallets (wallets with unusual early access or coordination) across the creator’s tokens.
  • Average bot buyer percentage — creators whose tokens attract almost exclusively bots are likely farming interaction metrics rather than building organic communities.
The creator risk score (0–100) is a direct ML feature and triggers an anti-signal at a threshold of 80.

​
Copy-trade detection

The copy-trade detection system maps the directional relationship between wallet pairs. When wallet A consistently buys a token a predictable number of seconds before wallet B, across dozens of tokens, it is likely that B is following A through an alert system or automation. The CopyTradeDetector runs on a 15-minute cycle and classifies each detected follower relationship into one of three types:
TypeTypical latencyMeaning
bot_copy< 5 secondsAutomated mirroring via on-chain listener; signal is already traded out
alert_copy5–60 secondsAlert service (e.g. Telegram bot) relaying the originator’s trade
manual_copy60–600 secondsHuman watching alerts and manually executing; signal still has value
The system resolves which wallet is the originator and which is the follower. Signals from followers are treated with lower confidence than signals from originators — a follower’s buy means someone else already paid a lower price, and the information content of that buy is diminished relative to the originator’s.
Copy-trade classification is directional: wallet A following wallet B does not imply wallet B follows wallet A. The graph captures asymmetric influence relationships across the tracked wallet set.

​
Anti-signal logic

The AntiSignalEmitter is the system’s last line of defence before capital is deployed. It scans every token with an active buy signal every 30 seconds, checking six independently computed risk signals. An anti-signal fires when two or more of the following conditions are simultaneously true:
Anti-signal triggers
1. Creator risk score > 80
   Example: "Creator [X] has 85% rug rate across 47 tokens"

2. Insider buyer pct > 40%
   Example: "62% of buyers are insiders/bundled wallets (8/13)"

3. Exit liquidity pattern detected
   Example: "4 smart wallets selling (12.3 SOL) while retail buying (3.1 SOL)"

4. Wash trade pct > 30%
   Example: "41% of volume appears to be wash trading"

5. Bot buyer pct > 60%
   Example: "78% of buyers are bots"

6. Bundle confidence > 70% AND bundle buyer pct > 30%
   Example: "Coordinated bundle (82% confidence, 45% of buyers)"
The multi-trigger design is intentional. Any single metric can be noisy — a 70% bot buyer rate on a brand-new token might simply mean no humans have heard about it yet. When two or more triggers fire simultaneously, the evidence is overwhelming.
Anti-signals published to trade:signals are consumed immediately by the live trader, which force-exits any open position in the flagged token. There is no grace period — the exit is executed at the next available opportunity regardless of current PnL.
Each anti-signal payload includes the full evidence object — not just that it fired, but why, with the specific numbers. This is logged and persisted to PostgreSQL for auditing and post-hoc analysis.

​
Exit liquidity detection

The exit liquidity trigger is one of the most reliable indicators of a dump in progress. The pattern requires all three conditions to be simultaneously true:
1

Prior smart-wallet buy signals

The token has had tracked-wallet buy signals in the last 15 minutes. This establishes that smart money was recently active and may have established a position at a lower price.
2

Smart wallets now selling

Those same tracked wallets are selling in the last 10 minutes. The pivot from buying to selling is the core signal — smart money is exiting.
3

Retail wallets now buying

Untracked (retail) wallets are simultaneously buying. The smart money exit is being absorbed by retail inflows, creating the classic exit liquidity dynamic.
The detector computes the tracked sell SOL versus retail buy SOL ratio as part of the evidence payload, quantifying how significant the exit is relative to new inflows. A ratio above 2:1 (smart money selling twice as fast as retail is buying) indicates an aggressive dump in progress.

​
Impact on ML features and live positions

Adversarial detection outputs flow into the ML pipeline in two ways: Direct ML features — is_bot, creator_risk_score, token_bundle_confidence, wash_trade_pct, is_copy_trader, and crowding_ratio are all included in the 68-feature standard vector. The ML model learns to weight these signals in combination with wallet quality and market features. Live position management — anti-signals bypass the ML inference loop entirely and act as direct force-exit instructions. This ensures that even a high-scoring position is closed immediately if adversarial evidence emerges after entry.

Wallet intelligence

How bot classification affects alpha scores and the confidence weighting of wallet signals.

Market regime

How signal crowding detection interacts with regime classification to apply score penalties.

ML features

The full 68-feature vector and how adversarial features are positioned within it.

Circuit breakers

How the live trader responds to anti-signals and enforces position limits under adverse conditions.

Build docs developers (and LLMs) love

Get started for freeTalk to us