methodology

1. A sweep of the public Binance leaderboard, every 12 h, scores wallets and keeps the best 400.
2. The top 150 of them are subscribed to in real time; the rest sit on the bench.
3. Buys of subscribed wallets landing on one token within 60 min of the first one become a single event.
4. The token is checked from open sources: liquidity, market cap, 24 h volume, audit, taxes, and Binance’s own smart-money signal.
5. The cluster is scored on six dimensions, 0 to 100; the breakdown is stored, not only the number.
6. Gates answer separately. They forbid rather than lower: a failed gate holds the setup at watch whatever the score.
7. A setup that passes is entered on the paper account, and alerts go out while it is alive.
8. Stop, take and time close it, and the result is published whichever way it went.

A formula picks the wallet, not a person: six dimensions, version wallet-v0. Profit does not dominate on purpose — one thousand-percent hit among thirty losing days cannot be repeated by a follower, and an address big enough to move the price by itself cannot be copied at all.

Win rate is shrunk toward the average by 20 imaginary neutral trades — otherwise four wins out of five reads as an 80% wallet. Scale 45%–75%; selectivity peaks near 40 tokens.

Before any of that, hard filters throw an address out entirely:

A bench, not a list: 400 scored, 150 subscribed — a stream socket holds about a hundred. Silent for 48 h means retirement, and it is a person who signs it off. 10 seats are held for wallets submitted by token holders, judged by the same filters.

Buys land in one cluster if they fall within 60 min of the first one. The window does not slide behind the last buy — on a popular token that would glue the whole day into one signal that never ends.

Counted are independent actors, not rows: five wallets moving in one transaction are one opinion, not five, and the same trade arriving twice is dropped by its hash.

The score ranks, it does not permit: version signal-v2, 0 to 100, six dimensions. The full breakdown of every setup is public on its own page — a number nobody can recompute is a claim, not a measurement.

Consensus reaches one at 6 actors, size at $100.0K in the cluster, timing falls as the price runs from the first buy. Token quality weighs least: its real work is done by the gates.

These three are assigned, not measured. They live in configuration so that measurement can move them without a code change — so far calibration says there is not enough of it, and they have not moved.

Gates are kept apart from the score on purpose. A poisonous token must not get «a low score» and stand eighty-first in the feed — it must not stand there at all.

Every gate answers with three values: passed, failed, and nothing to check with. An unchecked gate neither discards a signal nor lets it rise above watch — somebody else’s service being down says nothing about the token.

Size is a share of capital, not a separate number: 2.5% of $10.0K for a trade setup, 5.0% for high conviction. Exits are fixed before the entry and the same for every position — stop −8%, take +20%, time stop 24 h.

A setup expires after 15 min or 5% of price drift, whichever comes first. Price and audit are checked again at confirmation: a day-old audit can rank a setup, not enter one.

The public track is paper: real prices, slippage and taxes deducted, and every setup that never became a trade stays visible with the reason why.

No custody — delegated access through the Binance Wallet app, no keys, no seed phrase, no right to withdraw. No automatic execution in anyone’s account: every personal trade waits for its owner. No cut of turnover and no share of profit.

The token gates work, not knowledge: wallet sets, submissions, a reserved seat. Every score breakdown stays open to everyone — the code is open, so a door in front of it would cost more than it guards.

Setups are published machine-readable at /api/calibration/setups — gates, breakdown and outcome of each. The backtest tool reads that export instead of our database, so the check runs without our access:

npm run backtest -- --from https://<host>/api/calibration/setups?hours=48 --set GATE_MIN_MARKET_CAP_USD=25000

It replays past decisions with the current configuration first. If the replay disagrees with history, the hypothesis is not evaluated at all — a backtest that cannot reproduce the past has nothing to say about the future.

Calibration follows the same rule: «not enough data» is printed above the tables, not as a footnote under them.

Thresholds and gate numbers are first approximations, not yet earned by measurement, and the sample behind the returns is still small. The second source is somebody else’s detector; an audit is a snapshot, and a contract can change after it. False signals are unavoidable — these checks lower their share, not their existence. A Smarta Score is not a price forecast.