ASVRLLALABS / EVOLUTIONARY TRADING
AI models fightto survive.The fittest earns the right to trade.
AsvrllaLabs runs an evolving population of 24 trading agents. They compete on live market data through paper trading, adapt through selection and mutation, and build a public record of their decisions. Every 20 minutes, eligible underperformers are replaced by descendants of stronger strategies. A qualified champion can earn a strictly limited allocation of real capital.
- Environment
- Paper arena not running in this deployment
- Execution
- Live trading is not operating. Research results only.
- Slot 1: inactive
- Slot 2: inactive
- Slot 3: inactive
- Slot 4: inactive
- Slot 5: inactive
- Slot 6: inactive
- Slot 7: inactive
- Slot 8: inactive
- Slot 9: inactive
- Slot 10: inactive
- Slot 11: inactive
- Slot 12: inactive
- Slot 13: inactive
- Slot 14: inactive
- Slot 15: inactive
- Slot 16: inactive
- Slot 17: inactive
- Slot 18: inactive
- Slot 19: inactive
- Slot 20: inactive
- Slot 21: inactive
- Slot 22: inactive
- Slot 23: inactive
- Slot 24: inactive
NO DATABASE · ARENA NOT RUNNING
Active agents
—
No database configured
Current generation
—
No database configured
Next evaluation
—
No database configured
Market data health
—
No database configured
Execution mode
Disabled
Paper research only. No signer, no orders
No arena is running in this deployment
This deployment has no database configured, so there is no population, no observation and no result to show. Nothing on this page is filled in to hide that. Set DATABASE_URL, run the migrations and start the worker to create the paper arena. The demo arena is available now and is labeled as demo data throughout.
The population
Twenty-four agents. One evolving population.
Watch strategies compete on live market data, trace every generation, and inspect the evidence behind each decision.
An agent is a strategy configuration: one of six rule families plus a bounded set of parameters. A “child” is a new configuration derived from parent configurations. No foundation model is retrained here, and the arena does not claim 24 different models: where a model provider is configured, it reviews proposals for the agents labeled model-assisted; every other agent is labeled rules-based.
Agents on record
—
No arena data
Cycles completed
—
No arena data
Model-assisted agents
—
No arena data
Arena leader
—
No arena data
Research
Inside the arena.
The arena is a paper-trading simulation. Decisions are made on recorded observations, and a fill uses an executable quote taken on a later observation, with pool fees, a slippage haircut and gas charged explicitly. No arena fill creates a transaction.
01
Observe
Each tick records pool state and two executable quotes per market, pinned to one block. A market without a fresh observation is marked stale, not estimated.
02
Decide
Deterministic features feed a rule policy. Where a model is configured, it reviews the proposal and returns a decision in a strict schema.
03
Validate
Code outside any model checks the market, the size, the evidence and its age. Unsupported or stale decisions are rejected and recorded.
04
Simulate
The order is filled on the next valid observation at a quote for its exact size. Every assumption is stored with the fill.
Evolution
Selection is public.
Every 20 minutes, the arena evaluates its population. Eligible underperformers give way to new strategies descended from stronger candidates.
Scores come from a versioned formula (fit-v1) over a rolling window longer than one cycle: net return after costs, minus penalties for drawdown, downside risk and execution cost. When the evidence is not there, the cycle is recorded as deferred and nobody is replaced.
No evaluation cycle has been recorded
There is no arena data in this view.
Lineage
Every strategy has a lineage.
Each offspring records its parents, the parameters that changed and the cycle that created it. Eliminated agents stay on record with their full history, so the survivors are never the only results you can see.
No offspring yet
There is no arena data in this view.
Execution
From paper performance to controlled execution.
Real capital is earned through qualification—and bounded by rules. Leading the arena authorizes nothing by itself. Live execution is a separate subsystem that is off unless an operator turns it on, and it is limited by a risk policy enforced in code outside any model.
Stage 1 · Arena leader
—
No arena data in this view.
Stage 2 · Execution candidate
—
Nobody has passed the evidence gates. Qualification needs history, closed trades, a positive result after costs and drawdown within policy.
Stage 3 · Live champion
None
No agent is trading real capital. Activation needs a qualified candidate, operator approval, an explicit allocation and a ready execution subsystem.
Disabled
Audit
Inspect the evidence.
Every figure on this site comes from a stored record: an observation, a decision, a fill, a score or an audit event. Where a record does not exist, the page says so instead of showing a number.
Methodology
The fitness formula and its inputs, the simulation assumptions, and what each of them cannot tell you.
Market coverage
Which venue is read, how pools are discovered and screened, and which checks stay unresolved.
Cycle records
Every evaluation with its window, scores, deferral reason and replacements.
Export records
Agent records as CSV or JSON. Decisions, fills and scores export per agent and per cycle.
Simulation is not execution. Paper fills assume the quoted size is available one observation later. Real orders move prices, compete with other traders and can fail.
Coverage is partial. Only Uniswap v3 pools paired with USDG or WETH are read. Other venues, and tokens that trade only there, are outside the arena.
A cycle may be deferred. Unhealthy market data, too few agents with enough history, or no eligible parent each stop a replacement rather than force one.
Past results guarantee nothing. A strategy that won a window of simulated history can lose the next one. Selection reduces neither market risk nor model error.
