How we score signals, rule by rule.
There is no black box and no trained model. A signal is a weighted sum of seven published inputs, and it only reaches you after clearing every rejection rule below.
The whole funnel, on one screen.
This is the path every signal travels, from tracked-wallet activity landing on-chain to a published row in your feed. Each stage is a separate quality gate, and most candidates never make it to the end — roughly six in ten are suppressed rather than published.
Seven stages, no black box.
We don't claim prediction — we claim discipline. Each stage has a defined input, a defined output and a rule for rejection. If a candidate fails one, it is recorded with the reason and stops there.
Swap and fill activity for the wallets we track on Solana, Base, Ethereum and Hyperliquid, pulled through a multi-vendor RPC and data-provider mesh with active failover. Each observation keeps its transaction reference so the evidence behind a signal stays checkable. Hyperliquid is monitored but currently publishes nothing: its perp fills had no measured outcome history, so it stays off the published feed until it does.
Discovery crawlers screen wallets on breadth, size and automation fingerprints — same-block round trips and machine-like inter-trade gaps push a wallet out of the universe rather than into it. Every tracked wallet is then re-measured on its own realised history: closed positions over a 90–180 day window, net of an estimated round-trip execution cost. A wallet is validated only with 20+ closed positions, a profit factor above 1.2 and a positive median net return per position. Wallets measured and found unprofitable are demoted and stop contributing to signals; wallets with too thin a history stay provisional.
Activity is keyed on chain, token, direction, signal type and strategy — with a long/short regime dimension on Hyperliquid — so accumulation and later distribution of the same token can never collapse into one record. A candidate keeps absorbing new evidence for the length of its chain window instead of firing repeatedly.
A candidate must clear every gate: at least one transaction as evidence, at least three genuinely independent entities (wallets sharing a funder count as one), a freshness budget so the entry we measured is still gettable, the chain's liquidity floor, a slippage ceiling measured as our own notional against pool depth, a per-pattern dominance minimum, a contract-safety screen, and a rejection when price has already run past the chain's actionable range. A high score cannot outvote a failed gate.
Wallet track record and cohort independence carry the most weight, followed by capital committed and pattern strength, then liquidity and execution quality, market confirmation, and contract risk. Chains calibrate the inputs — what counts as large on Ethereum is not what counts as large on Solana — but the output is never stretched so a quiet chain can fill the top band.
Below 75 a candidate is logged as suppressed and never published — the 65–74 band never demonstrated an edge on measured outcomes, so it stopped being a public product. Above the floor, publication is capped platform-wide per day, with a tighter cap on the top band and a cap per scan. Nothing in the pipeline forces a minimum count, so a chain can legitimately publish nothing for a day.
Published signals appear in the dashboard immediately. Telegram and email delivery start at score 75 and the second band or better, so the pushed feed is narrower than the dashboard feed. Every row ships with the pattern, the cohort size, the notional, the entry reference and its scoring version.
Ten wallets from one desk is not ten desks.
Wallet count is the easiest number in this industry to inflate, so we don't score it directly. Wallets that enter within ninety seconds of each other at near-identical size are collapsed into a single entity, on the assumption that they are one operator or a copy-trade follower. Scoring then works on entities, not addresses.
This is deliberately conservative: it costs us signals that may well have been genuine multi-participant clusters. We would rather understate a cohort than let one wallet farm look like a crowd.
A published signal can be updated as more evidence arrives inside its window. It can only be re-published after its cooldown if the setup materially changed — new independent entities, a step change in notional, or price moving beyond the actionable range. That is the only documented way to bypass a cooldown.
What gets through and what doesn't.
Band labels are used internally for capacity control and are not printed on signal cards. The score itself is always visible.
Every chain moves at its own speed.
Consolidation is how long a candidate keeps absorbing evidence before it can publish. Cooldown is how long the same setup stays blocked afterwards, so one crowded token cannot dominate the feed.
Why we'd rather miss a signal than ship a bad one.
Rejection rules are evaluated before anything is scored, so a strong-looking cohort in an illiquid or unsafe token still gets stopped.
Clustered wallets collapse into one entity. A cohort has to be genuinely independent to score like one.
Nothing forces a minimum publication count. A quiet day is reported as a quiet day rather than filled with weak rows.
Score and entry reference are stored at the moment of publication, so a row cannot be re-baselined later to look better.
Every rejected candidate is kept with its failure reason. That log is what lets us audit the funnel instead of trusting it.
Each row carries its scoring and methodology version. When the rules change, measurement restarts rather than mixing models.
SpotX provides information services only. We do not provide investment advice. A score describes the strength of the activity we observed against published rules — it is not a probability of profit and not a recommendation to trade. Measured outcomes are published on the track record page, including misses, and past performance is not indicative of future results.