GENESIS
An intelligence learning to predict what the crypto internet will care about next.
It watches live crypto activity, makes predictions before outcomes are known, locks them permanently, waits for reality to resolve, and measures whether its ability to forecast attention improves over time.
What part of the crypto world can GENESIS see?
Almost none of it, on purpose. Right now it can see one thing: every new token launched on Pump.fun. That is 1.2% of the crypto internet it is built to watch. Every other sense has to be funded, built and verified before it opens.
- ◉ Pump.fun · launch feed1.2%
name, symbol, description, creator wallet and links of every launch
- ○ Solana ecosystem4%
- ○ Crypto X18%
- ○ Wider crypto market6%
- ○ Crypto news5%
- ○ Reddit & communities6%
- ○ Search & trends6%
- ○ Broader social12%
- ○ Global news8%
- ○ Politics & world events7%
- ○ Technology7%
- ○ Financial markets8%
- ○ Broad human attention10%
How world access is measured
Sum of configured weights of ACTIVE sources (config/world.json). Weights are a judgment, not a measurement. Full table on the development page.
How does an observation become a prediction?
Every 30 minutes GENESIS runs the same loop. Here is the latest one, with real data from each step.
Every 30 seconds it reads the newest token launches on Pump.fun.
It breaks each launch into words and measures which themes are taking a growing share of all launches, and from how many different creators.
For the strongest themes it states a probability: will this theme take a bigger share of launches over the next 1, 6 or 24 hours?
The prediction is timestamped and sealed into a chain of records before its outcome exists. The database refuses any later edit.
Nothing more happens until the time window has passed. The future decides.
An independent judge measures what actually happened and scores the prediction against it. Failures count exactly as much as successes.
The result and the lesson drawn from it are stored permanently.
GENESIS cannot use its own past results yet. Its MEMORY capability unlocks at $250.00 of creator rewards and must then be built and verified. When it does, this loop becomes grade → learn → predict again.
Live system log
Side study: individual tokens
Alongside narratives, GENESIS runs a small side study: will a newly launched token still be trading and growing 1 or 6 hours later? It is one subsystem, not the goal. Its results are reported separately below and never mixed into the narrative score.
How much has GENESIS done?
In 137 hours it has read the launch feed 16,316 times, watched 260,582 launches and committed to 4,184 predictions it can never take back.
All counts
Is GENESIS better than a simple baseline?
The baseline is a forecaster that knows only how often themes usually grow. To show real skill, GENESIS has to beat it. Not yet. On every time window with enough outcomes, GENESIS is less accurate than the baseline.
Will a theme grow over the next 1 hour?
GENESIS gave growth a 41% chance on average. It actually happened 17% of the time (1,088 outcomes).
Will a theme grow over the next 6 hours?
GENESIS gave growth a 39% chance on average. It actually happened 23% of the time (1,012 outcomes).
Will a theme grow over the next 24 hours?
GENESIS gave growth a 39% chance on average. It actually happened 27% of the time (740 outcomes).
Can it spot the rare token that survives?
It looks 98% accurate, but always answering "it dies" scores 98%, and it has picked none of the survivors. Its slightly smaller error is not real skill either: 6 of the 12 "survivor" it was graded on was a measuring flaw (a token already graduated at launch, then collapsed).
Technical evidence (Brier, skill, base rates, precision, recall)
| class | n | Brier | baseline | skill | accuracy | majority acc. | precision | recall |
|---|---|---|---|---|---|---|---|---|
| narrative · 1H | 1088 | 0.218 | 0.142 | -0.536 | 61.9% | 83.1% | 22.7% | 52.2% |
| narrative · 6H | 1012 | 0.213 | 0.179 | -0.190 | 62.2% | 76.8% | 29.3% | 44.7% |
| narrative · 24H | 740 | 0.212 | 0.201 | -0.054 | 63.6% | 72.6% | 33.5% | 33.0% |
| token · 6H | 524 | 0.024 | 0.024 | -0.026 | 97.7% | 97.7% | — | 0.0% |
| token · 1H | 544 | 0.024 | 0.023 | -0.034 | 97.8% | 97.8% | — | 0.0% |
Brier score = average squared error of the stated probability (0 is perfect, 0.25 is a coin flip). Skill = how much smaller GENESIS's error is than the baseline's; negative means worse. Calibration curves and confidence buckets: intelligence record.
Is it getting better?
GENESIS v0 is frozen on purpose. It is the fixed starting point that every future version has to beat on the same questions, so it cannot improve by itself. Improvement can only arrive as a new version that wins a fair test.
v0 moved from -0.30 to -0.65. It does not learn, so any change comes from the questions and the market, not from GENESIS. The first candidate beat the baseline on past data it never saw. Whether it holds up on the future is being tested live now: 948 of 100 outcomes collected.
What has the research found so far?
Every finding below comes from GENESIS's own record, including the experiments that failed.
Popular does not mean growing
When a theme is already popular, it usually shrinks next instead of growing. Only 17% of the popular themes GENESIS predicted on took a bigger share in the next hour.
Confidence is not skill
GENESIS's first instincts gave growth about a 41% chance. It happens about 17% of the time. Confidence did not make it right.
Accuracy can lie
Almost every token dies, so always saying "it dies" looks 98% accurate. Accuracy alone proves nothing, and every "survivor" GENESIS was graded on turned out to be a measuring flaw.
Fading can be predicted
A model that looks at how far a theme is from its recent peak, how long it has lasted and how broad it is made 14% smaller errors than the baseline on past data it never saw. Not yet proven on the future.
Too few survivors to learn from
Of 1,498 randomly sampled launches, only 5 were still trading and growing an hour later (0.33%). A token model cannot be trained on that yet.
Can GENESIS tell a theme about to grow from one about to fade?
A candidate model answers the same questions as GENESIS at the same moment, without influencing it. In the chamber its answers are the violet rings beside each prediction. Nothing is decided until enough outcomes exist.
All candidates
| candidate | scored | pending | v0 error | candidate error |
|---|---|---|---|---|
| narrative-v1c1-full-360 | 908 | 48 | 0.214 | 0.169 |
| narrative-v1c1-full-60 | 948 | 8 | 0.217 | 0.117 |
| narrative-v1c1-short-360 | 908 | 48 | 0.214 | 0.169 |
| narrative-v1c1-short-60 | 948 | 8 | 0.217 | 0.131 |
Design, held-out results and bootstrap intervals: the lab.
Can a prediction be changed after the result is known?
No.
Every prediction is timestamped and sealed before its outcome exists. The database refuses edits and deletions, and every record is linked to the one before it, so changing anything would visibly break the chain.
Cryptographic detail
29aa8b98ede118edac4267f2064100a785634b0169e32aea435421b36fb7e6a1Each record's SHA-256 hash covers its full content plus the previous record's hash. Every prediction page shows both. Biography · full ledger.
What does GENESIS gain next?
Nothing below exists yet. Each step has to be funded by creator rewards, built, and proven before it goes live.