A post (X or a tracked Telegram channel) containing a token's full contract address. Cashtags alone don't count — they're too ambiguous to grade honestly. Each caller/token pair counts once, at the earliest post we saw.
The token's forward price return from the time we ingested the call — not a cherry-picked entry, not the day's low. We report the median return across all of a caller's measured calls (medians can't be rescued by one lucky moonshot), plus hit rate and the share of calls that dumped. Calls younger than 2 hours are too fresh to grade; calls older than 10 days age out of the window.
Every receipt displays n, the number of measured calls behind it. Below n = 8, a caller's tier is 🟡 provisional — never a full 🟢 alpha or 🔴 exit-liq badge. Small samples get small claims.
Every call is snapshotted into an append-only ledger the first time we see it. If a call later disappears from the source, it is flagged deleted — and stays in the caller's record, still measured against the chain. The receipt shows how many of a caller's calls were deleted after the fact.
Nothing reaches the Signals channel on vibes. A candidate token — surfaced because graded wallets are already buying it — must pass a ribbon of twelve named gates, all at once. One fail, no alert. The gates:
vetothe token verifies live at alert time — not rugged, stale, or already dumping
deployerthe deployer wallet's on-chain track record — serial ruggers are an automatic no
confluenceenough independent, graded wallets are actually in, weighted by measured grade
price_ceilingrefuses tokens that already ran — we don't alert you into a top
dispersionearly buying is spread across holders, not concentrated in one bag
early_bundlethe first buys weren't one bundler wearing many wallets
velocitybuy flow is accelerating — thin conviction with no momentum doesn't clear
falling_knifeno buying a bleeding chart without fresh flow behind it
lead_qualitythe wallets that led the move are measured leaders, not just bodies
cohort_poisona poison-majority cohort is never a buy — no matter how many winners ride along
authority_integritymint/freeze authority revoked, no Token-2022 honeypot extensions (below)
lp_safetyLP lock and honeypot status via independent scanners (below)
Why you can trust it: every evaluation is written to an append-only gate log with the names of the gates that failed — and blocked near-misses are posted publicly as WATCH digests, so you see what we rejected and why, in the same channel as the passes. What we don't publish: the numeric cutoff behind each gate. The checks are public; the tuning is the edge — and a published cutoff is a recipe for building tokens that skim just under it.
Before anything is alertable we read the mint account itself. An unrevoked mint authority means the deployer can print supply after you buy. An unrevoked freeze authority means they can freeze your token account so you can never sell. Certain Token-2022 extensions — non-transferable, permanent delegate, abnormal transfer fees — are the Solana equivalent of an EVM honeypot contract. Any of these is an automatic block.
Verifiable: run getAccountInfo on the mint against any Solana RPC — the same fields we read are public to everyone. Honest failure mode: if the RPC is unreachable the check reports unknown, which never fakes a pass and never blocks a real entry on an infra hiccup.
Liquidity-lock percentage and honeypot status come from third parties we don't control — RugCheck for Solana, GoPlus for EVM chains. Unlocked liquidity or a honeypot flag blocks the signal.
Verifiable: both endpoints are free and public — paste the same contract address and you get the same report we gated on. Honest failure mode: best-effort external data; when it's unavailable the gate doesn't pretend to know.
Confluence is counted in independent entities, not wallets. Before a wallet cluster reads as conviction, we collapse it using two signals that are expensive to fake at once: shared funding lineage (wallets funded from the same source are one operator) and co-buy lockstep (wallets that keep buying the same tokens within tight windows of each other are one operator). One person running twenty wallets collapses to one entity — and the signal prints the deflated count and the inflation factor.
Why you can trust it: to look like N independent wallets you'd need N funding histories and N uncorrelated buy tapes — at which point you are, functionally, N traders. What we don't publish: the window lengths and overlap ratios that define lockstep.
Every tracked wallet is re-graded on one question: does copying this wallet, at our detection latency, make the copier money? Not the wallet's own PnL — leaderboard wallets routinely profit from entries no follower can get. Forward marks for a wallet's buys come only from other wallets' fills (a wallet cannot wash-trade its own grade), taken as medians, never a single thin fill. Wallets with measured-negative copy value are labeled poison and their signal weight drops to zero — a demoted wallet can never produce an alert.
Why you can trust it: positive boosts are gated behind a shuffled-label permutation test — no wallet gets extra weight until the top of the roster provably beats what random selection over the same trades would produce, and that claim ships with confidence intervals. Protective vetoes are never gated: refusing to boost requires proof of skill; refusing to follow a measured loser does not. What we don't publish: the grading horizon, the poison threshold, or the weight multipliers.
The wallets that were accumulating a token are watched continuously. When that graded cohort flips from accumulation to distribution, the exit flag fires — off wallet flow, which moves before price does. And since 2026-07-27, insider dumps are surfaced, not suppressed: wallets carry the poison label because of prior behavior, so a poison cohort exiting a token is often the earliest real distribution signal — it now escalates instead of being filtered as noise. When a poison cohort piles into the buy side, that posts as an explicit FADE — intel, never a buy signal.
Why you can trust it: every exit-flag fire is logged with the cohort's buy and sell flow at the moment it fired — recomputable from public chain data. What we don't publish: the window and flow thresholds that define "distribution onset."
A separate layer sits above every signal and only permits or blocks — it never trades and holds no keys. It enforces: a macro regime gate (risk-off or distribution regimes halt new entries entirely; defensive regimes cut size — including a BTC-structure throttle that halves size but never binary-flattens), a hard per-position ceiling, a max concurrent positions cap, single-bag and top-3 concentration caps, and a daily-loss circuit breaker that halts all new entries after a defined drawdown on the day.
Why it matters: entry quality is worthless without exit and sizing discipline — this layer is the precondition for any automation. What we don't publish: the cap values and the breaker threshold.
Brand-new launches have no wallet history to grade, so the Genesis Feed covers a token's first hours on pure on-chain heuristics: authority integrity, LP lock + honeypot, deployer track record, and early-flow dispersion, composed into one of three verdicts — ACTIONABLE / WATCH / AVOID. "No strong signal yet — too early to call" is an allowed answer, and the most common one.
Why you can trust it: every feed ships a coverage note stating what was NOT checked — discovery currently depends on public launch feeds (not the millisecond of mint), and authority/deployer checks are Solana-only today. We'd rather publish the gap than imply coverage we don't have.
Every open position is scored into one of three lanes: HOLD / TRIM / EXIT. The checks run in strict order: ① a rug kill-switch first — liquidity draining or a one-hour price collapse fires loudest, for every hold type, regardless of PnL; ② a trailing stop off your peak gain, scaled to the position's hold-type (scalp / swing / hold); ③ a hard cut-loss from your entry; ④ smart-money distribution — the graded wallets that bought with you are leaving. Everything is entry-aware: a routine pullback on a bag still green against your entry is not an exit, and the model says so instead of panicking you out.
Why you can trust it: positions are reconstructed from on-chain swap history — entries, sizes, and peaks are all recomputable from the chain, not self-reported. What we don't publish: the trail percentages, the stop level, and the rug-switch cutoffs.
Every number is recomputable from public chain data plus the timestamped call ledger. If a caller believes a receipt is wrong, the dispute is arithmetic, not argument — the inputs are listed above, and our grading artifacts are snapshotted every 6 hours into a hash-chained ledger published to a public GitHub repo we don't control the timestamps on, so past numbers can't be quietly rewritten.