Omega is a private research stack that builds, breeds and verifies AI trading models. Every number on this page is computed from live records or a pre-registered test — including the ones that don't look good.
loading live numbers…
Two champions run with real money. Each one earned its slot by passing tests on data it had never seen. Their live records are shown exactly as they stand.
Watches brand-new pump.fun pools. A learned filter decides in the first ~15 seconds which launches to buy; it sells at +24.8%, at −50%, or after 10 minutes — and a live rug guard sells early when a pool starts draining.
Daily-bar trend following on 24 spot pairs, only trading while a Bitcoin regime gate is open. Built and walk-forward tested on 2019–2026 data.
Profit factor (PF) = money made on winning trades ÷ money lost on losing trades. PF 1.0 breaks even; PF 2.0 makes two dollars for every dollar it loses. A PF is only worth something if it was measured on trades the model could not have learned from — so we report where each PF comes from.
The gap between the bot's own picks (PF 8.29) and the launches it skipped (PF 0.78) is the whole point: the edge is the filter, not the exits. The live PF is lower than the replay because live trading has already paid for two rugs — the losses the rug guard now targets.
Every closed live trade, in order. The line is cumulative profit in SOL at 0.05 SOL per trade; each dot is one trade.
Most of the losses come from rugs: a creator drains the pool and the price gaps straight through the stop to −80% or worse. Our on-chain tap sees every pool trade about half a second after it happens; the guard sells a position the moment its pool loses 10% of its SOL within 10 seconds.
Screenshots prove nothing. Every live decision CHAMPION MINI v2 makes — every buy and every skip — plus every live entry and every result is hashed and chained within 15 seconds, and the chain is timestamped on the Bitcoin blockchain through OpenTimestamps. Decisions are locked in minutes before their outcome exists. Nothing can be added, dropped or edited later without breaking the chain and its Bitcoin timestamp, and every skip is counted next to every win.
decision (BUY/SKIP), entry, outcome, model version → canonical JSON → SHA-256 leaf
every 15 s: Merkle root of new leaves, chained to the previous batch hash
each batch hash goes to public OpenTimestamps calendars → committed to a Bitcoin block
chain, batches and Bitcoin receipts mirrored here every 5 minutes
wget -r -np -nH proof/ (or download the files below) python3 verify.py chain.jsonl batches/ pip install opentimestamps-client ots verify batches/0000001.hash.ots
Files: chain.jsonl · batches/ · verify.py · summary.json. The proof starts on 25 Oct 2026, when this ledger went live; earlier trades are shown on this page but are not claimed as proof.
We don't train giant models — we breed agents on top of them. Each agent descends from a frontier model (Anthropic Claude Opus 5.5, Fable 5.1, Sonnet 5.5; OpenAI GPT and Codex; open models) and adds what we control: traits, tools, memory, a budget and a lineage. Every agent has one job tied to profit: propose a filter that makes CHAMPION MINI v2 skip the launches that lose.
Ancestor model + traits + tools + memory + budget. DNA is hashed and written to a tamper-evident ledger.
Each agent studies frozen history and proposes one rule (1–3 conditions) for which launches to skip.
Rules are validated and frozen together into a cohort. Its fingerprint is published before any result exists.
Every rule in the cohort is graded on the same next 150 unseen launches — paired, once, against a random placebo.
Agents whose rules make money breed (traits cross, ancestors pass down); agents that keep losing retire.
A passing rule is replicated by a second model family, then approved by a human, then goes live.
| filter | agent | rule (buy only if…) | cohort | verdict |
|---|
In our research engine (GENESIS) every strategy is an organism called an OPAL. Its DNA is data — an entry time, an exit, and a few conditions — never code. Each piece of DNA is a gene with its own fingerprint, so we can trace exactly which parent every gene came from as organisms cross and mutate.
Genes still carried by living organisms across several generations — the traits selection keeps choosing.
Every birth is appended to a hash chain: each record includes the previous record's hash, so changing any past birth breaks every later link. The chain head is posted daily to a signed message channel.
The easiest way to fool yourself in trading research is to test an idea on the data you used to find it. These rules exist to stop that.
Hypothesis, data window, costs and pass bar are written and hashed before the test.
Rules are chosen on the first 60% of history and tested once on the last 40%.
A result must beat random picks of the same size, not just zero.
Measured fees, price impact and execution delay — the same engine for every test.
A second model family re-runs every pass with its own code.
Live trading on the future is the final judge. Live numbers here update automatically.