Skip to content

Every model, tested on launches it never saw.

One card per forecast class: what it reads, how many launches it was tested on, its base rate, its skill, and whether it is live. Measured 18 Nov 2026.

A young record: the first model went live on 18 Nov 2026 and each test covers weeks of launches, not years. Measured by CON, not audited by a third party.

Illustration, generated with AI

Cards

Skill is the Brier skill score: 0 is no better than the base rate, 1 calls every launch right. Hit rate: share that came true among the launches ranked highest, as many as catch a quarter of all cases.

  • crash24_m2

    Crash within 24 h, read at minute 2

    Skill
    0.59
    Hit rate
    88.7%
    Base rate
    4.1%
    Tested on
    63,528 launches
    Trained on
    148,232 launches
    Walk-forward
    trees better on 16 of 17 days

    In the gate: shown nowhere until it passes

  • crash24_m5

    Crash within 24 h, read at minute 5

    Skill
    0.63
    Hit rate
    83.7%
    Base rate
    3.8%
    Tested on
    55,247 launches
    Trained on
    128,909 launches
    Walk-forward
    trees better on 14 of 17 days

    In the gate: shown nowhere until it passes

  • crash24_m15

    Crash within 24 h, read at minute 15

    Skill
    0.68
    Hit rate
    90.4%
    Base rate
    2.5%
    Tested on
    44,766 launches
    Trained on
    104,453 launches
    Walk-forward
    trees better on 14 of 17 days

    In the gate: shown nowhere until it passes

  • crash24_m60

    Crash within 24 h, read at minute 60

    Skill
    0.55
    Hit rate
    75.2%
    Base rate
    1.0%
    Tested on
    34,431 launches
    Trained on
    80,339 launches
    Walk-forward
    trees better on 7 of 17 days

    In the gate: shown nowhere until it passes

  • dev_exit

    Dev sells out, read before launch

    Skill
    0.15
    Hit rate
    98.1%
    Base rate
    90.7%
    Tested on
    147,568 launches
    Trained on
    344,323 launches
    Walk-forward
    trees better on 15 of 17 days

    Live in the check since 18 Nov 2026

A rate describes a group of launches, never one coin. No class forecasts winners; that would read as a buy signal, so none is built.

How they are tested

The rules every class meets before it reaches a page.

  • HoldoutThe newest days are kept out of training and used only to score.Every night
  • Walk-forwardEach past day is scored again by models trained only on the days before it.17 days per class
  • Gate3 nightly runs in a row with skill above zero and better than the best single rule, then review.Before a class shows
  • ParityThe live code computes the same features and scores as the research code.Trees to 1e-9, features 98.6% to 99.99%
  • Drift watchEach live class scored on the day before; three days under half its holdout skill raise an alarm.Every night

Results

The figures behind the cards.

Every class on launches it never sawholdout
ClassLaunches testedBase rateSkillHit rate
Crash within 24 h, read at minute 263,5284.1%0.5988.7%
Crash within 24 h, read at minute 555,2473.8%0.6383.7%
Crash within 24 h, read at minute 1544,7662.5%0.6890.4%
Crash within 24 h, read at minute 6034,4311.0%0.5575.2%
Dev sells out, read before launch147,56890.7%0.1598.1%
Measured 18 Nov 2026 on the newest days, held out of training. Hit rate: share that came true among the launches ranked highest, as many as catch a quarter of all cases. Skill is the Brier skill score: 0 means no better than the base rate, 1 means every launch called right.
Skill per held-out day, crash within 24 h read at minute 5Brier skill score
Tree ensembleLogistic regression
0.800.53
31 Oct3 Nov6 Nov9 Nov12 Nov15 Nov
Show the 17 values
DayLaunchesTreesRegression
31 Oct 20267,9650.7540.709
1 Nov 20267,5930.6960.591
2 Nov 20268,8320.7150.651
3 Nov 20269,3820.7710.720
4 Nov 202611,7550.6760.654
5 Nov 20268,8590.7180.634
6 Nov 20269,4200.7400.680
7 Nov 20268,0270.7460.654
8 Nov 20268,6380.7170.622
9 Nov 20269,1530.7530.695
10 Nov 20268,2110.7470.682
11 Nov 20268,1230.6830.631
12 Nov 20268,2250.6490.585
13 Nov 20268,7310.7200.666
14 Nov 20267,8470.6640.667
15 Nov 20267,1700.5810.600
16 Nov 20264,9040.6180.563
Walk-forward: each day scored by models trained only on the days before it. Trees beat the regression on 14 of 17 days and lost on 0. Skill is the Brier skill score: 0 means no better than the base rate, 1 means every launch called right.
What holder data adds
ClassTested onSkill beforeSkill with holders
Crash, read at minute 266,4830.6240.660
Crash, read at minute 557,9280.6570.700
Crash, read at minute 1546,9900.6770.687
Crash, read at minute 6035,9810.5220.523
Top 20 holders, dev holding and largest holder added to the features, computed by the live code. Measured 19 Nov 2026 on a newer export with more launches and the first-minutes features, so the starting skill differs from the cards. Models with holder data go through the gate from scratch.

Studies

Questions the research runs answered on the way, with the data that answered them.

Crash within 24 h, by flag at minute 2share of launches
  • All launches4.3%
  • Bundled first block43.1%
  • Top 10 hold over 35%38.7%
  • Top 10 hold 22 to 35%20.9%
  • Largest holder over 10%19.5%
  • Dev holds over 10%13.8%
  • 3+ snipers in the first second7.2%
  • Largest holder 5 to 10%6.4%
FlagLaunches with itCrashed with itCrashed withoutTimes the base rate
Bundled first block3.4%43.1%2.9%15×
Top 10 hold over 35%2.3%38.7%3.4%11.3×
3+ snipers in the first second54.4%7.2%0.7%10.1×
Top 10 hold 22 to 35%7.2%20.9%3.0%7.1×
Largest holder over 10%3.0%19.5%3.8%5.1×
Dev holds over 10%2.9%13.8%4.0%3.5×
Largest holder 5 to 10%14.8%6.4%3.9%1.7×
237,686 recorded launches, scored on held-out days, measured 18 Nov 2026. Shown are the flags that separate; 5 of 12 did not and carry no weight on fresh coins.
Minutes from the reading to the crashlaunches that crashed
  • Read at minute 2
  • Read at minute 5
  • Read at minute 15
  • Read at minute 60
1 min10 min1 h10 h
Read atCrashed10%25%Median75%90%
minute 28,9280.8 min2.5 min8.1 min23.1 min68.3 min
minute 57,1091.8 min4.6 min11.7 min32.9 min103.2 min
minute 153,8633 min7.4 min20.7 min60 min175.5 min
minute 601,2838.6 min27.6 min79.8 min251.7 min763.3 min
Launches that crashed within 24 h, split by when they were read. Thin line: 10 to 90% of them; bar: the middle half; red tick: the median. Measured 18 Nov 2026.
Similar launches: forecast against outcome50 nearest
Forecast 0%Forecast 100%

Up: share that crashed. Dashed: forecast equals outcome.

Show the 6 values
Forecast bandLaunchesForecastCrashed
0% to 1%56,0290.4%0.1%
1% to 3%1,6372.2%2.4%
3% to 10%1,4806.0%7.0%
10% to 30%1,43119.3%23.4%
30% to 60%1,40643.0%45.4%
60% to 100%1,90576.4%83.5%
Each held-out launch gets the crash share of its 50 most similar earlier launches. 63,888 launches, measured 19 Nov 2026. Close to the diagonal means the share can be taken at its word; it is shown as a reason, the trees give the rate.

What did not work

Tried, measured, dropped. Kept here so nobody has to take our word for what is in the models.

Coordinated buyer groups
Wallets that bought the same launches early again and again were grouped. Launches bought by a group whose earlier launches crashed 30% of the time or more crashed in 34% of cases, against 0.22% without such a group. Strong alone, but the models already see it in the first minutes of trading: added to them, the groups raised no score on held-out days. Not in the models.
Operator trace
Dev wallets were linked through shared funding to find one operator behind many launches. Launches marked this way did worse on every measure than a plain control: dev sell-outs 38% against 43% for bundled launches without any history, crashes 8.3% against 10.3%. The dev wallet's own record alone was more precise, 78% against 67%. Dropped.
Narrower time windows
A model was trained to give each launch its own window for the time to a crash. Its windows hit as often as promised but were no narrower than the fixed spread per reading time. The fixed spread is what the pages show.

Research runs of 18 and 19 Nov 2026 on the CON market recording.