reading the ledger
A predictive market model trained from scratch on Pump.fun launches to forecast outcomes before they’re known.
It doesn’t rely on another model’s market knowledge. Everything it knows comes from launches it observed itself.
Every coin launched on Pump.fun in the last few minutes, newest first, and what the model makes of each one. The numbers fill in as the 1, 5 and 15-minute snapshots are taken. Greyed-out rows are coins the system stopped tracking because nothing happened: no trades, no buyers, no attention. Hover a column heading to see what it means. The full board is on Live.
| token | age | state | runner | P(2×) | P(10×) | crash | activity |
|---|---|---|---|---|---|---|---|
| $DEMOCRACY DEMOCRACY DIED IN ROMANIA | 2m 24s | low activity · not tracked | — | — | — | — | 0 buyers in 5 min |
| $OTF THEY DEPLOYED MAIN | 2m 27s | no trades · not tracked | — | — | — | — | 0 buyers in 5 min |
| $PINGPING PINGPING | 4m 09s | waiting | 82 | 3.3% | 3.69% | 2% | 9 buyers in 5 min |
| $和平熊猫 Ping Ping and Fu Shuang | 4m 57s | no trades · not tracked | — | — | — | — | no trades yet |
| $PingPing Ping Ping | 5m 06s | no trades · not tracked | — | — | — | — | no trades yet |
| $CUP Captain Underpants | 5m 25s | launch | 52 | 21.5% | 1.28% | 11% | 32 buyers in 5 min |
| $BITCAT BitCat | 5m 30s | low activity · not tracked | — | — | — | — | 1 buyers in 5 min |
| $PingPing Ping Ping | 5m 40s | low activity · not tracked | — | — | — | — | 0 buyers in 5 min |
Every number here is locked: written before the outcome, never edited. Greyed-out rows went quiet, so they are recorded but never scored, and never deleted. The full board keeps all of them so nothing is hidden.
Every snapshot taken above travels down the line into the compute hall, where the models are trained on all of it.
The cables behind this page carry the data. Keep scrolling to follow them in.
Every step since day one, including bugs found and fixed. The story plays by itself. Click any step to read it.
The whole system was written in one day: the collector that watches Pump.fun, the database, the trainer, the model format, the benchmarks, the ledger and this website.
Two groups, kept apart for good: the first 820 predictions were made while a bug made the busiest launches look quiet; from number 821 on, the fixed collector is used. Nothing was rewritten or deleted. 4 checkpoints of the ledger are published in the code repository, so anyone can check it has not been touched.
MZ-F1-v0.41 was frozen on 08 Oct 17:57 UTC and has not been touched since. Its fingerprint: 1f982faf06…
On a clean test set it ranks coins that double within 6 hours with an AUC of 0.917 (1.0 is perfect, 0.5 is a coin flip): 58 doublers among 2,343 snapshots. Now being confirmed on live launches.
Its “drops 70% within an hour” estimate is well calibrated (calibration error 0.007): when it says 30%, about 30% do. Precision 0.902 on 2,343 snapshots.
59,082 predictions locked before their result, 142 missed, 2658 late. The ledger is hash-chained and verified: change one entry and every entry after it breaks.
12 data bugs found, measured and fixed in a public changelog, including the one that made every early number wrong.
An early signal (AUC 0.67) on 106 runs from a single day, pointing the way we expected. Far too little data to trust yet.
Coins that kept running after doubling pulled in about 5× more new buyers in the next half hour. 21 cases against 152.
Coins that look dead at 5 minutes and run later. Every one found so far collapsed again; more data is being collected to look again.
Why one doubler keeps going to 10× while another dies. Six 5×+ cases studied closely so far.
A flat price with buyers, holders and volume still growing now gets close tracking instead of being dropped.
The one pattern predictable 30 minutes ahead so far (AUC 0.75, 26 cases).
No coin has been seen coming back after a collapse in the live data yet. Watching for it.
Every fixed exit rule tested loses. The sellable peak lasts about a minute and the entry price sets the ceiling. Described, not solved.
Creator rewards from the token are paid into the treasury and listed in the public ledger.
Helius data credits so more launches can be watched closely, and more training runs and experiments.
More coins tracked for longer, more runners followed, more studies written up.
A stronger next model and a longer public track record. That brings attention, and the loop repeats.
Every creator reward the token earns is spent on the system itself: more data (Helius credits, so more launches can be watched closely), more compute (more training runs and experiments), deeper tracking (runners followed for longer) and more research (more studies, better predictions).
No token is live yet, so no reward has been received and none is promised. Nothing on this page is a promise of returns. When rewards exist, they and what they pay for appear in the ledger.