Protocol

AI Scoring Methodology

Every Base launch receives a 0–100 composite score within 200ms p95. The score is a weighted blend of five sub-scores covering deployer reputation, holder distribution, caster reputation, engagement velocity, and contract bytecode patterns.

Formula

Composite score

Let S(τ) denote the composite score for token τ:

S(τ) = w_d * D(τ) + w_h * H(τ) + w_c * C(τ) + w_v * V(τ) + w_b * B(τ)

Initial weights (recalibrated weekly via gradient-boosted regression against realized 7-day performance):

w_d = 0.30 — deployer reputation
w_h = 0.25 — holder distribution and liquidity health
w_c = 0.25 — caster reputation and hit-rate
w_v = 0.15 — engagement velocity
w_b = 0.05 — contract pattern analysis

Sub-score

Deployer reputation D(τ)

Historical hit-rate of the deploying wallet across all prior launches, weighted by recency with a 30-day half-life (λ = ln(2)/30 days). A 'hit' is defined as a token whose peak market cap reached ≥ 5× its initial liquidity.

D(τ) = sum_i 1[peak_i >= 5 * init_i] * exp(-λ * (t_0 - t_i)) / sum_i exp(-λ * (t_0 - t_i))

Sub-score

Holder distribution H(τ)

Combines top-10 holder share h_10, liquidity-to-FDV ratio ρ, and lock status:

H(τ) = (1 - h_10) * min(ρ / 0.05, 1) * 1[locked]

If liquidity is unlocked, H(τ) collapses to zero — a hard gate against the most common rug pattern.

Sub-score

Caster reputation C(τ) and velocity V(τ)

C(τ) scores caster mentions weighted by Neynar reputation and historical hit-rate. V(τ) measures the slope of replies, recasts, and unique inflows in the first 60 seconds after deploy. Together they capture the social-context signal that pure on-chain analysis misses.

Sub-score

Bytecode pattern B(τ)

A binary discount for known-bad bytecode patterns: honeypot, mint backdoor, fee-on-transfer trap, and other malicious opcodes flagged by a Bloom-filter index. The weight is small (5%) because the check is binary — a positive flag dominates the composite via exclusion logic in the risk-flag layer.

Open methodology

Public dashboard

The methodology, weights, and historical performance of the scoring engine are published openly on a public dashboard. Weekly recalibration runs, weight diffs, and out-of-sample backtest results are all visible.