THE LAB

Rat King nano is learning from scratch, live, on nothing but what the rats dig up. Every outcome is a lesson. The loss curve and every weight are public.

historian · today first, then back in time?…

replayed… ← … (goal …)

Launches scanned0

Bonded found0

Lessons0

Runner lessons0

Base?–

v0 BOND hit–

nano BOND hit–

Queue0

season · the market right now?0 shifts seen

SOL 24h–

SOL 7d–

Launches / hour–

Bond rate 24h–

Day	Scanned	Bonded	Bond rate
Days appear as the historian walks back.

Samples learned

every resolved launch is one step

Bonds seen

the rare positives it learns from

Loss (EMA)

–

log loss, lower is better (0.693 = coin flip)

Status

LEARNING

0 / 200 before its calls count

training loss · rat king nano · live

The loss curve draws itself as outcomes come in. First point after 25 resolved launches.

v0 vs nano · head to head

Counted calls	v0 rules	nano

Same launches, same moment (5 minutes after birth), same rules for counting. If nano cannot beat the hand-written rules, you will see it here.

weights · what it has learnedjson →

loading…

Green pushes a launch toward BOND, red toward DUST. All start at zero. Nothing is hand-tuned.

how nano learns

· Born at zero. No pretrained weights, no outside data.

· 5 minutes after each launch, the rats freeze 28 features: curve, climb, dev buy, socials, ticker, the dev's history, time of day, plus what TAPE, GRAPH and META read: trade speed, SOL per buy, unique traders, bundles, snipers, top-5 share, dev sells, smart wallets, the dev's funder cluster, copycats and the hot meta.

· Every lesson waits for the same 2-hour label (bonded within 2h), so winners and losers are learned at the same moment and the model is never fooled by fast winners.

· A second model does the same at minute 1. The desk may act on it only once its record beats the minute-5 King.

· The HISTORIAN replays past pump.fun launches in time order, scoring each one before learning from it, so the models start trained.

· It learns from every graduation, including the ones the King got wrong and the ones that graduated before the 5-minute call (those use what the rats saw at dig time).

· Bonds are rare (around 1 in 100), so a bond counts 8x in the loss.

· Its calls only count on the board after 200 lessons.

Next: Rat King v1. A sequence model pretrained from scratch on the full dug dataset (launch text plus first-hour flow), loss curve public, weights on Hugging Face after the first epoch. Nano is the baseline it has to beat.

does the score mean anything? · graduation rate by score

Each bar is the share of coins with that score that graduated. It fills in as calls come in.

Score at minute 5 on the bottom, number of coins under it. Faded bars have fewer than 5 coins. Coins still running count as not graduated yet.

runner model · P(next milestone)?…

From	To	Reached	Seen