AI

The Data Stack: What an AI Agent Actually Needs to Score a Token End to End

An AI agent is only as good as the data you feed it. Here's the full stack a memecoin scorer needs, layer by layer, and where each layer lies to you.

· 7 min read · Blackhat Empire

The Premise: An Agent Is a Pipeline, Not a Brain

Everyone wants to talk about the model. The model is the least interesting part. An AI agent scoring a memecoin is a data pipeline with a judgment layer bolted on top. If the pipeline is thin, the judgment is theater. If the pipeline is poisoned, the judgment is dangerous.

So let's walk the stack end to end — what an agent actually needs to look at a token and produce a defensible score. Memecoins are extremely high risk and most go to zero. No score, human or machine, changes that. What a good score does is tell you why something is likely to fail, not promise it won't.

Layer 1: Contract and Chain State

Before anything else, the agent needs raw chain truth. That means contract address, deployer address, mint authority status, freeze authority status, ownership renouncement, and the full transaction history of the deployer wallet.

On Solana this is the token mint account and its authorities. On EVM chains it's the contract bytecode, ownership functions, and any proxy patterns. The agent needs to flag:

  • Mint authority still live — supply can be inflated at will
  • Freeze authority live — your tokens can be locked
  • Deployer linked to prior rugs — same wallet, same pattern
  • Proxy or upgradeable contracts — logic can change after you buy

This layer is deterministic. There's no AI judgment here, just parsing. Get it wrong and every layer above is noise.

Layer 2: Liquidity and Pool Structure

A token without liquidity is a number on a screen. The agent needs pool addresses, LP token ownership, whether liquidity is locked and for how long, and how concentrated the pool is.

Key signals: LP burned vs. held by deployer, pool age, liquidity depth relative to market cap, and whether the pool has been migrated or recreated. A token that has had three pools in two days is telling you something.

This is where you pull the actual chart and pool data — on GMGN at https://gmgn.uk (mirror: https://gmgn.fr). The agent should be reading the same pool state a human trader reads, not a cached summary.

Layer 3: Holder Distribution

The agent needs the holder list, not just the holder count. Specifically:

  • Top 10 and top 50 concentration
  • Cluster analysis — wallets funded from the same source
  • Fresh wallet ratio — how many holders were created recently
  • Insider overlap — holders who also bought the deployer's last five tokens

A token with 2,000 holders where 40% of supply sits in wallets funded from one address is a single entity wearing a crowd costume. The agent must be able to see through that.

Layer 4: Trade Flow and Volume Quality

Volume is the most gamed metric in this market. The agent needs trade-level data: buy/sell ratio over time windows, average trade size, wash-trade detection (repeated round-trip wallets), and volume concentration by wallet.

A useful frame: is volume coming from many independent wallets at varied sizes, or from a tight set of addresses cycling the same tokens? The second one is a signal, and it's not a good one.

Layer 5: Social and Narrative Inputs

This is the layer where agents most often embarrass themselves. Social data is real signal, but it's also the easiest to fake. The agent needs:

  • Mention velocity — is attention accelerating or flat?
  • Account age and history of the accounts amplifying it
  • Copy-paste detection — identical posts across many accounts
  • Source diversity — one Telegram group vs. genuinely independent voices

If every mention traces back to the same handful of accounts created last week, the "narrative" is a manufacturing run. Treat social as a weighted input, never a primary signal.

Layer 6: The Scoring Layer Itself

Now the model. The job here is not to predict price — it's to output a structured risk profile with confidence levels and the specific reasons behind each component. A good agent scorecard looks like:

  • Contract risk: pass / warn / fail
  • Liquidity risk: pass / warn / fail
  • Distribution risk: pass / warn / fail
  • Flow quality: pass / warn / fail
  • Social authenticity: pass / warn / fail

Then a summary that names the two or three biggest red flags. Not a single number out of 100 that hides the reasoning. See /v2/dyor/reference.html#metrics for how we frame these metrics.

Where the Data Comes From

An agent needs multiple independent sources per layer, because any single source can be wrong, stale, or manipulated. Cross-check contract state against two RPC providers. Cross-check holders against more than one indexer. Cross-check social against raw message data.

If you're building or tuning an agent, the community channels are where you'll see real-time signal and real-time noise side by side — the main alert channels and the chain-specific chat groups are listed at https://blackhat.finance/channels.html. Watching how fast those alerts fire and how the chat reacts is itself a data source.

The Honest Bottom Line

An agent scoring tokens end to end is a research accelerator, not an oracle. It compresses hours of manual checks into seconds and forces you to look at the layers humans skip when they're excited. It will still be wrong sometimes. Most memecoins still go to zero.

The point is to know why you're wrong when you are, and to have the receipts. Build the pipeline properly and the score means something. Skip the boring layers and you've built a very confident machine for repeating your own biases back at you.

Read the rules at /v2/dyor/reference.html#rules before you trust any score — including ours.

Community

Stay connected across the chains:

Charts and on-chain research: https://gmgn.uk.