Read the page top to bottom in the order of the six numbered cards: that is the order the bot works in. If a card's number is not moving, the panel for that stage explains why.
Top bar
- Phase: which stage of the plan the bot is in. It flips from "collecting data" to "scoring channels" on its own once 50 calls have been scored.
- Telegram: green means the user session is logged in and reading channels; the number is how many channels it reads.
- Chain: where launch and trade data comes from.
rpc_logs = decoded straight from a Solana node. "behind" = delay between the blockchain and the bot, usually under a second.
- Triage: "LLM on" = an AI model assesses candidate channels before joining; "rules only" = no key set.
- up / db / updated: process uptime, database size (old rows are pruned hourly), last page refresh (every 30 s and after important events).
Card 1 · Launches seen
New pump.fun tokens created in the last hour. Beneath: today's total, how many pumped 3x+ this hour, and how many are odd, meaning they use a non-standard price curve. Odd tokens are tracked but never scored or traded, because our price model is only exact for the standard curve. Hundreds of launches an hour is normal; most stay near 1.0x forever.
Card 2 · Channels watched
Telegram channels the bot reads. Seed = from the guide's list. Trial = discovered by the bot and being evaluated: its calls are recorded and scored but it would never be traded. Active = it has enough scored calls to hold a tier.
Card 3 · Calls seen → matched
A call is a source saying "buy this token now". There are two kinds and the card shows them separately:
- Channels: a Telegram message containing a token address. Matched means the token is a pump.fun launch the bot saw created, so it knows exactly how many seconds after launch the post came. Unmatched calls are usually tokens from other launchpads or tokens older than our window. The line beneath shows how many channel calls were first to post the token among the channels we watch, how many were copied (near-identical text to an earlier post elsewhere = a reposter), and how many were deleted afterwards (a channel hiding losers).
- Wallets: a trader wallet's first buy of a token within 10 minutes of launch (0.1 SOL or more, not an insider). These are always matched because we saw the buy on-chain. The line shows how many in the last hour and how many different tokens.
Why split them? Wallet calls are far more numerous, so a single combined number would hide what the channels are doing.
Card 4 · Calls scored
Ten minutes after a matched call, the bot replays it: "if I had bought 2 seconds after this call with my size, and followed my exit rules, what would I have made after every fee and price impact?" That number is the call's result, and it is what every source is judged on. Beneath: the share of winners and the average result. A call can be unscorable if the token is on an odd curve or had no trades.
Card 5 · Sources rated
Channels and wallets earn a tier from their scored calls. Golden needs 15 scored calls (20 for wallets), a statistically positive result, a win rate above 45%, and no deleted posts; it is the only tier the bot would trade. Silver is promising, Bronze and unranked are watching only. Three losing calls in a row drop a source immediately. Everything is unranked in the first days; that is expected.
Card 6 · Finding channels
Candidates found in messages, forwarded posts, Telegram recommendations and name search. Each is read (no join needed) and scored 0–100 by the AI for "is this a real Solana meme-coin call channel?". Below 60 is rejected without a join. The best are joined, at most 10 per day, 5 minutes apart, to protect the Telegram account. The panel Finding channels lists them; hover a rejected row for the reason.
Card 7 · Paper trading
The bot trades on paper: real signals, real prices, simulated fills. When a source at or above the minimum tier posts a call, the engine runs the safety checks, applies the risk limits (max open positions, one per source, one per creator, daily and weekly loss caps, a cool-down after 3 losses in a row), sizes the trade (Golden full size, Silver half, Bronze a quarter, never more than 2% of the bankroll or 2% price impact), waits the modelled send delay, and fills through the bonding curve at that moment's reserves, charging every fee. A fill whose price moved more than the slippage limit during the delay is recorded as a failed transaction: no position, fees lost, exactly as on-chain. Open positions are then managed on every trade of that token by the same exit rules the scorer uses: principal out at +100%, trailing stop 25% below the peak once above +50%, exit on net selling, on an insider sale, before graduation, at the 10-minute limit, or at a 30% loss of executable value.
The card shows total paper P&L; the panel shows the risk state, today's P&L against the daily cap, all-time and 24-hour statistics, results by source tier, why signals were rejected, open positions live, and closed trades. Click a closed trade for its fills. Controls: pause (no new positions), resume, set size, KILL (close everything and stop), un-kill. Creating the file data/KILL on the server has the same effect as KILL. Only Golden-tier results count for the decision to go live; Bronze trades exist to validate the pipeline while sources earn their tiers.
Activity
A live diary of everything the bot does, newest first, in sentences. The dropdown filters to calls, channels & discovery, pumps, or errors. Hover a line for the raw data. Typical lines: a call being seen and matched with its timing; an AI verdict on a candidate; a wallet buy recorded as a call; a scored result; a safety-check verdict; a token reaching 3x or graduating.
Recent calls
The table behind cards 3 and 4. After launch = seconds between the token's creation and the post. First? = this source posted before any other channel we watch. Peak = highest price in the 10 minutes after the post, as a multiple of the launch price. Our result = the replayed P&L. Exit = why the simulated trade ended: principal_out / trail = took profit, stop = loss limit, momentum = selling pressure, insider_sell = an early wallet dumped, time_stop = 10 minutes passed, graduation = the curve completed. Click a source or token for details.
Channels & scores
Win % = share of scored calls that made money. Typical P&L = average after dropping the best and worst 10%, so one lucky call cannot flatter a channel. Low estimate = a conservative bound that must be above zero for Golden. Profit factor = total wins ÷ total losses. Speed = median seconds after launch the channel posts. Deleted % = share of posts later removed; deleters cannot be Golden.
Trader wallets
Same scoring as channels, plus: Exit quality = how close to the token's peak the wallet sells (1.0 = the top). Early % = share of its buys inside 15 s of launch; high = a sniper bot whose speed we cannot copy. Creators = how many different token creators it bought from; low = an affiliate of one deployer. Per day = call rate; hundreds = a bot. Own P&L = what the wallet itself made. Flags list these problems; flagged wallets cannot be Golden, and any wallet ever seen as an insider is excluded. ↗ opens the wallet on Solscan.
Recent launches
Newest tokens with age, curve type (std/odd), peak so far, trade count and how many sources called them. 🎓 marks a token that completed the curve and moved to PumpSwap.
Detail pages
- Channel: scoring summary and every scored call with our result next to the guide's ladder result, plus the price at 1/2/5/15/60 s after each post (how fast the edge decays).
- Wallet: our result of copying each buy next to what the wallet itself made and its exit quality; linked channels = channels that keep posting right after this wallet buys.
- Token: creation time, curve, progress, creator and its past launches, insider wallets, every source that called it in order, and the safety checks with a "run now" button. Each check shows pass / fail / unknown, the measured value, the limit and a plain explanation.
Studies
The Phase 2 analysis over all scored calls: price decay after a post, time to peak, the dip before the peak with a stop-loss table, our exits vs the guide's ladder, results by seconds after launch, curve progress, hour and source type, a grid of entry latency × hold time, and which safety checks separate winners from losers. It says plainly when there is too little data to trust.