ABOUT
Dracula, king of all vamps
Every new Solana narrative sets off a wave of launches on the same theme. Several migrate, then they PvP each other — vamping liquidity, holders and attention — until one coin takes the narrative and the rest bleed out.
Dracula is a research agent that reconstructs those battles: who launched first and where, what each coin was paired to, how holders and liquidity moved between rivals, which KOLs and wallets backed the winner, which callouts carried it, and what tokenomics it ran. Every battle it studies becomes training data, so it gets better at telling, while a battle is still live, which coin is most likely to take it.
HOW IT WORKS
- 01
Data
A live scanner reads every new pump.fun launch, groups copycats by ticker and name, and starts tracking a battle the moment its second coin migrates. Deterministic Python then pulls each coin's trades, holders, launchpad, migration, tokenomics, KOL wallets and callouts, and compresses every coin into feature snapshots at fixed points: T+1m, T+5m, T+15m, at migration, migration+30m, T+24h — each using only data from before that moment.
- 02
Agents
Specialised analysts — timing, holder flow, KOL wallets, callouts, tokenomics — each read one view of those summaries (never raw transactions) and make claims that must cite the numbers they rest on. A comparison analyst lines the coins up.
- 03
Reviewer
Every citation is checked against the collected data before an evidence reviewer grades what's left. It can only downgrade a claim, never rescue one. What survives becomes the post-mortem: why the winner won, ranked by evidence strength, with its data limitations.
- 04
Model
Each finished battle is a labelled example. The win model trains walk-forward by date (never a random split), is calibrated, and scores live battles. The agents explain its output; they never replace it. After a battle resolves, the prediction is checked against what happened and the miss is kept.
HOW TO READ THE NUMBERS
- Win probability — the model's estimate that a coin takes the battle. It is a probability, not a call: a 60% coin loses 4 times in 10.
- Shaded band — the confidence range around that estimate. Two coins whose bands overlap are not meaningfully apart.
- n and confidence — every probability shows how many past battles the model trained on and its confidence. Early on that number is small and confidence is low.
- Winner — the first coin to hold the highest market cap for 10 straight minutes once two or more coins have migrated, counting only coins that pass minimum fee and liquidity filters. If no coin does, the battle has no clear winner and isn't used for training.
- Accuracy — hidden until the model has trained on about 50 battles. The track record lists every resolved prediction, misses included.
SCOPE AND LIMITS
- Solana only. Launchpads: pump.fun first, then LetsBonk and Bags; others as “other”.
- Read-only. Dracula holds no keys, signs nothing and executes no trades.
- KOL wallets and callouts come from curated lists and indexed social sources, so coverage is partial.
- Historical market data is rebuilt from on-chain transactions; when a battle's history is incomplete, the affected numbers are marked unavailable rather than guessed — see each post-mortem's data limitations.
- Data: Solana RPC (Helius), launchpad programs, DexScreener. The read-only API is public at api.vampagent.fun.