Pre-launch· Token not live yet — no rewards or treasury until launch
reading the ledger
Forward test
Predictions written down before the result
Model Zero scores new Pump.fun launches 5 minutes after they start and writes the scores into a public, tamper-proof ledger before anyone knows what happens. Six hours later the real outcome is written next to them. This page shows how those predictions turned out. Use the arrows to step from the newest model back to the original frozen one; hover any ? for what a number means, and click a top-picks number to see the coins in it.
Model Zero v0.67
newest · 1 of 4
Scored afterwards on the exact same locked inputs the frozen model saw, for the same coins. It never trained on any of these coins, but its scores were not locked in advance. Trained 2026-10-14 on launches up to 2026-10-12. 5,899 coins scored.
Of the 5% of coins this model liked most, 25% doubled within 6 hours, against 3% of all coins: 7.2× better than picking at random.
What happenedAll coinsHow often this happened across all the coins scored here. Picking coins at random gets you this, so it is the level to beat.RankingRanking skill. Take one coin that did this and one that didn't: how often did the model rank the right one higher? 0.50 is a coin flip, 1.00 is always right.PrecisionHow well the model packs its highest-ranked coins with real hits, from 0 to 1. Picking at random scores about the same as the base rate, so the further above the base rate, the richer its top picks are in hits.Top 0.5%The 0.5% of coins the model ranked highest for this, and how many of them actually did it. Click a number to see the coins.Top 1%The 1% of coins the model ranked highest for this, and how many of them actually did it. Click a number to see the coins.Top 5%The 5% of coins the model ranked highest for this, and how many of them actually did it. Click a number to see the coins.
Doubled (2×) within 6 hoursranked by its 2× chanceall coins 3.5% (204/5899)ranking 0.82precision 0.199
5× within 6 hoursranked by its runner score (no separate 5× chance)all coins 1.0% (60/5899)ranking 0.74precision 0.054
10× within 6 hoursranked by its 10× chanceall coins 0.3% (16/5899)ranking 0.81precision 0.076
20× within 6 hoursranked by its 20× chanceall coins 0.1% (7/5899)ranking 0.84precision 0.148
Crashed 70% within 1 hourranked by its −70% chanceall coins 3.7% (219/5899)ranking 0.96precision 0.548
How each runner-score band did
The runner score is the model's 10× chance, ranked from 0 to 100 among these coins99 – 100 means the top 1% of coins by the model's 10× chance. Below 50 is the less promising half.. Each row shows what really happened to the coins in that band.
Runner scoreCoins2×Share of these coins whose price reached 2× the price at the prediction, at any point within 6 hours.5×Share of these coins whose price reached 5× the price at the prediction, at any point within 6 hours.10×Share of these coins whose price reached 10× the price at the prediction, at any point within 6 hours.20×Share of these coins whose price reached 20× the price at the prediction, at any point within 6 hours.CrashedShare that fell 70% or more below the price at the prediction within the first hour.1 in 10 reachedThe best price within 6 hours that one coin in ten from this band reached or beat, as a multiple of the price at the prediction. Most coins never rise at all, so this shows how good the better ones got.Average bestThe average of every coin's best price within 6 hours, as a multiple of the price at the prediction. A few big runners pull this up.
99 – 1005824%5%1.7%1.7%38%4.52×11.94×
Runner score 99 – 10058 coins
2× 24%5× 5%10× 1.7%20× 1.7%crashed 38%1 in 10 reached 4.52×average best 11.94×
95 – 9923611%6%1.7%0.4%12%2.49×1.65×
Runner score 95 – 99236 coins
2× 11%5× 6%10× 1.7%20× 0.4%crashed 12%1 in 10 reached 2.49×average best 1.65×
80 – 958858%2%0.7%0.5%6%1.59×1.67×
Runner score 80 – 95885 coins
2× 8%5× 2%10× 0.7%20× 0.5%crashed 6%1 in 10 reached 1.59×average best 1.67×
50 – 801,7703%1%0.2%0.0%4%1.12×1.12×
Runner score 50 – 801,770 coins
2× 3%5× 1%10× 0.2%20× 0.0%crashed 4%1 in 10 reached 1.12×average best 1.12×
below 502,9501%0%0.1%0.0%2%1.00×1.14×
Runner score below 502,950 coins
2× 1%5× 0%10× 0.1%20× 0.0%crashed 2%1 in 10 reached 1.00×average best 1.14×
Are its percentages honest?
When the model says a coin has, say, a 10% chance, about 10 in 100 such coins should really do it. Each row groups the coins by the chance the model gave them and shows how often it actually happened.
2× within 6 hours
Model saidAverageHappenedCoins
0–1%0.5%0.7%1,792
1–2%1.4%1.4%1,690
2–5%2.9%2.4%1,697
5–10%6.8%12.4%315
10–20%13.7%14.9%175
20–35%26.5%26.7%131
35–50%41.4%23.8%84
50–100%56.8%46.7%15
10× within 6 hours
Model saidAverageHappenedCoins
0–1%0.3%0.1%4,787
1–2%1.4%0.6%509
2–5%3.1%0.8%367
5–10%7.2%2.6%152
10–20%14.1%0.0%64
20–35%24.5%0.0%15
35–50%42.7%0.0%4
50–100%52.2%100.0%1
−70% within 1 hour
Model saidAverageHappenedCoins
0–1%0.4%0.2%4,423
1–2%1.4%2.0%692
2–5%3.0%7.4%349
5–10%7.0%18.2%154
10–20%14.3%27.1%96
20–35%27.4%51.0%51
35–50%42.5%53.1%32
50–100%79.3%73.5%102
Highlighted: groups of 20 or more coins where reality was far from what the model said.
▸Full audit of the original test: model fingerprint, coverage, cohorts, baselines, by day, paper P&L
over artifact, feature names and order, preprocessing, calibrators, config, evaluator version, code commit, evaluation set and the percentile table
Frozen at / forward test since
08 Oct 17:57 / 08 Oct 17:57
Training data to
07 Oct 21:38
launches after this date are prospective for the model
Chosen on
clean-eval-v2 · 7222af207381…
Sample
25% of tracked launches
sha256(mint) bucket, salt forward-test-v1; fixed before any outcome
Outputs it cannot give
p5x6h: no head for this outcome in any version (recorded as null)
Coverage and audit
ledger chains verified
Launches observed
184,227
all launchpads, since the freeze
Eligible (tracked)
79,570
Sampled
19,880
by hash, before any outcome
Ledger entries
59,581
chain head 1aee4c13cd1c…
5m predictions
19812
18925 timely · 887 late · median 43 s after the snapshot
Missed
51
Failed query:
with eb as (
select distinct on
Data-quality flags
6601
capped pages, partial history or no price
6h windows
11927 done
926 open · 1413 awaiting history · 5546 incomplete
Evaluated below: timely predictions on active launches (≥10 trades, volume in the last 5 minutes) with a complete 6h window — 5957 at 5m. Late, missed and incomplete entries stay in the ledger and are counted above, never scored. Cohorts: pre-fix seq 1–820 (820 entries) and post-fix seq 821+ (58761 entries) are reported apart below.
LIVE PROSPECTIVE RESULTS — post-fix cohort, 5-minute checkpoint (primary)
seq 821+ · 58761 entries
Corrected pipeline since 08 Oct 20:03: the history fetch pages incrementally from the stored coverage (up to 20 pages), so hot launches are read in full. This is the cohort to read.
223 evaluated · 2× 6 · −70% 7 · top-5% by P(2×): 2/11 · mean P(2×) 4.5%
2026-10-09
1577 evaluated · 2× 59 · −70% 55 · top-5% by P(2×): 11/79 · mean P(2×) 5.1%
2026-10-10
1343 evaluated · 2× 62 · −70% 68 · top-5% by P(2×): 13/67 · mean P(2×) 4.9%
2026-10-11
1144 evaluated · 2× 28 · −70% 43 · top-5% by P(2×): 11/57 · mean P(2×) 3.5%
2026-10-12
1177 evaluated · 2× 34 · −70% 26 · top-5% by P(2×): 12/59 · mean P(2×) 3.8%
2026-10-13
435 evaluated · 2× 15 · −70% 20 · top-5% by P(2×): 5/22 · mean P(2×) 3.9%
▸Paper P&L — economic edge, kept apart from the ranking edge
Net multiple per trade: entry at the first minute after the snapshot, 3% cost per side, stops assumed hit before targets within a minute, exit rules fixed before any outcome was seen.
▸PRE-FIX COHORT — seq 1–820 (820 entries): potentially affected by the hot-launch pagination defect; shown apart, never merged
Pre-fix cohort, 5-minute checkpoint
seq 1–820
Entries made while the history fetch had a 6-page cap that hot launches exceeded: some busy launches were recorded as having no trades (inactive, never evaluated), so the evaluated sample here under-represents hot launches. No entry was changed or removed (DATA-INTEGRITY-CHANGELOG §11).
58 evaluated · 2× 1 · −70% 3 · top-5% by P(2×): 1/3 · mean P(2×) 5.6%
▸Paper P&L — economic edge, kept apart from the ranking edge
Net multiple per trade: entry at the first minute after the snapshot, 3% cost per side, stops assumed hit before targets within a minute, exit rules fixed before any outcome was seen.
▸HISTORICAL CLEAN BENCHMARK (reference only, not combined with the live numbers)
v0.41 on the locked set clean-eval-v2 (7222af207381…), 2343 active 5m rows from held-out launches of 2026-10-05 → 10-07. The model's training data overlaps that period (held out by coin, not by time), which is why the live test above exists.
2× within 6h
AUC 0.917 · AP 0.256 · top-5% 27/117 (23%) · base 2.5% (58) · reach_2x_6h (direct)
5× within 6h
AUC 0.933 · AP 0.112 · top-5% 11/117 (9%) · base 0.7% (16) · upside_6h:ge2 (proxy)
10× within 6h
AUC 0.901 · AP 0.041 · top-5% 3/117 (3%) · base 0.3% (7) · reach_10x_6h (direct)
20× within 6h
AUC 0.985 · AP 0.027 · top-5% 1/117 (1%) · base 0.0% (1) · reach_20x_6h (direct)
−70% within 1h
AUC 0.991 · AP 0.902 · top-5% 83/117 (71%) · base 3.9% (91) · collapse_1h (direct)