Methodology
How the lab works, and what it cannot tell you
Scoring, simulation assumptions and limitations, written so that a stored result can be recomputed and questioned. The numbers on this page are the configuration in force for the paper arena you are viewing.
Data modes
The site has three explicit modes and never moves between them on its own. If real data is unavailable, the page shows a setup or error state; it does not substitute fixtures.
- Demo
- Deterministic fixtures replayed through the real engine in a separate in-memory database. A persistent “Demo data” indicator is shown. No market observation, no current performance, no signing.
- Paper
- Real observations from the chain when a database and a worker are running. Execution is simulated and labeled as simulated. Stale and missing data are shown as such.
- Live
- Only after an operator activates it and every readiness check passes. Confirmed transactions and their reconciliation, kept apart from paper results. Off by default; it has never traded in this build.
Agents and strategies
The population is exactly 24 agents. An agent is a strategy configuration: one of 6 rule families plus a vector of bounded parameters. The founding population is drawn from a seeded generator, so it can be regenerated. These are configurations to test, not claims that any of them is profitable.
Decisions are produced in four separate steps. Features are computed deterministically from stored observations. A rule policy turns them into a proposal: buy, sell, hold or skip. Where a model provider is configured, the model reviews that proposal for the agents labeled model-assisted and returns a decision in a strict schema; it can confirm, reduce the size or decline. Validation in code then rejects unsupported markets, invalid sizes, stale evidence and malformed output. Only an accepted decision reaches the simulator.
Without model credentials every agent runs on its rule and is labeled Rules-based. A rules-based decision is never shown as a model response. The arena does not use 24 different foundation models, and no model is retrained during a cycle: a “child” is a new configuration derived from parent configurations. Token names, symbols and links are untrusted on-chain text; they are never sent to a model and cannot change policy.
| Family | Enters when | Exits when | Parameters |
|---|---|---|---|
| Momentum MOM | Lookback return at or above the entry threshold, with a positive return over the most recent quarter of the lookback. | Lookback return at or below the exit threshold, or stop, target, or time limit. | 10 |
| Mean reversion REV | Price z-score against the lookback mean at or below the entry level. | Z-score back at or above the exit level, or stop, target, or time limit. | 10 |
| Breakout BRK | Mid price above the channel high (excluding the current tick) by the buffer. | Mid price below the exit-channel low, or stop, target, or time limit. | 10 |
| Liquidity-aware trend LQT | Fast average above slow average, pool depth at or above the minimum, and depth not falling over the slow window. | Fast average below slow average, or depth down by the drain threshold over the slow window, or stop, target, or time limit. | 11 |
| Volume confirmation VOL | Lookback return at or above the entry threshold and window volume at or above a multiple of its earlier average. Skips when swap volume is unavailable. | Lookback return negative, or stop, target, or time limit. | 10 |
| Defensive cash allocation DEF | Realized volatility at or below the ceiling, a modest positive lookback return, and the cash floor respected after the purchase. | Volatility above twice the ceiling, a negative lookback return, or stop, target, or time limit. | 11 |
Simulation
The arena is a persistent paper-trading engine on the server. Nothing about a result is computed in the browser, and there is no random P&L. Every tick records one observation per market: the pool mid price, the pool's quote-side depth, swap volume since the previous tick when the logs are available, and two executable quotes taken at the same block: buying the reference size, and selling that amount back.
No look-ahead. A decision is made on observation k. Its order is filled on a later observation, k+1 or after, at a quote requested for the order's exact size at that later block. An order that finds no usable observation within three ticks fails and produces no fill. Paper fills never create a transaction hash.
Stale data. When a market has no fresh observation, positions in it keep their last recorded value and are flagged stale. The agent cannot trade that market, and the flagged valuations count against its data-quality eligibility. The value is not presented as a price the position could be sold at.
- Tick interval
- 60 s
- Starting paper bankroll
- $10,000 USDG per agent
- Reference quote size
- $250 USDG
- Fill latency
- next valid observation, at least 1 s after the decision's observation
- Pool fees
- Inside the venue's quote; attributed per fill for reporting
- Slippage haircut
- 10 bps off the quoted output
- Gas
- (quoter estimate + 60,000 units) × observed gas price, in USDG at the observed ETH price
- Position marks
- Executable sell rate of the reference quote (a bid after fees), not the mid price
- Stale after
- 180 s without a usable observation
- Accounting
- Integer token units end to end; floats only for signals and display
What the simulation leaves out: it assumes the quoted size is still available one observation later, it does not model other traders reacting to the order, and a quote at the reference size understates the impact of a larger position. Real execution would be worse than these figures, by an amount this model cannot state.
Fitness
Formula fit-v1. Every eligible agent is measured over the same rolling window of 120 min, on returns relative to its own equity at the window start. The window is longer than one cycle, so a single lucky 20 minutes does not decide anything.
score (bps) = 10,000 × ( R − 0.5·DD − 0.25·DS − 0.5·C )
- R · net return
- Equity at the window end over equity at the window start, minus one. Already net of pool fees, slippage haircut and gas.
- DD · drawdown
- Largest peak-to-trough fall inside the window, with the peak starting at the window-start equity.
- DS · downside risk
- Root mean square of negative per-tick returns, scaled to the window by √n.
- C · execution cost
- Modeled costs paid in the window over average equity. Half of them are charged again as a margin for simulation error, which also penalizes turnover.
Eligibility comes first. An agent is scored only if it was alive for the whole window (so a new agent warms up for a full window), has at least 40 valuation observations in it, and no more than 20% of them stale. An agent with a short clean history or a freshly reset bankroll therefore earns nothing from it. Current balance, lifetime results and these window metrics are stored and shown separately. Every cycle page lists each agent's inputs so a score can be recomputed by hand.
Evolution
A durable worker evaluates the population every 20 min from a timestamp persisted in the database. The countdown in the interface is only a display. Each cycle has a unique key and runs under a database lock, so the same slot cannot be evaluated twice.
- Check market-data health. Unhealthy data defers the cycle.
- Determine which agents are eligible, and compute their scores.
- Require at least 8 eligible agents; otherwise defer.
- Select up to 2 of the lowest-scoring agents, but only those at least 5 bps below the eligible median. If nobody is, nobody is replaced.
- Select parents: agents above the median with at least 2 fills in the window. No eligible parent defers the cycle.
- Derive offspring. The primary parent supplies the family and base parameters; a second parent of the same family may contribute parameters by crossover; each parameter then mutates with probability 0.35 by at most 20% of its range, clamped to its hard limits.
- Record lineage and the exact parameter differences, replace atomically, and verify the population is still 24. If that check fails, the whole cycle is rolled back.
An offspring starts with a fresh paper bankroll and must warm up before it can be scored. Eliminated agents keep every record: decisions, fills, valuations and scores. Their positions are marked at the last observation and are not liquidated in simulation. Survivors are never the only results visible, which is what keeps the record free of survivor bias. Missed cycles (the worker was down) are logged and not replayed.
Champion qualification
The arena leader, a qualified execution candidate and an active live champion are three different things. Promotion is evidence, not proof of future profitability.
- History
- 6 h alive and 300 valuation observations
- Closed trades
- at least 8
- Performance after costs
- net return above 0 bps over 4 h
- Drawdown within policy
- at most 600 bps
- Data quality
- at most 10% stale valuations
- Validation on unseen data
- 120 min and 100 observations recorded after candidacy, return not below -50 bps
- Minimum tenure
- 120 min before a challenger can take the seat
- Challenge rule
- lead by 25 bps for 3 consecutive cycles
- Activation
- Healthy execution infrastructure, operator approval and an explicit capital allocation
An agent that is eliminated from the arena, or whose drawdown since candidacy exceeds policy, is disqualified. When a champion changes, live positions are not reset or abandoned: they belong to the execution account and stay in the ledger.
Live execution
Live trading is a separate subsystem and is disabled by default. Three independent switches must be on: an environment flag on the worker, an operator activation recorded in the database, and an approved champion with a capital allocation. The scope of this version is a project-owned account with limited capital. There are no public deposits, no pooled investments and no deposit button.
The worker never holds a key. It sends an unsigned transaction to an isolated signer, which decodes the calldata itself and signs only an exact-amount approval to the verified router or a swap whose output returns to the account. No seed phrase is requested or stored, and no key material enters the browser bundle, the logs, the database or a public environment variable. The model never sees any of it.
Orders follow a durable state machine: proposed, rejected, approved, submitted, confirmed, failed, or unknown while reconciling. The transaction hash and nonce are stored before the broadcast. After an uncertain submission nothing is retried: reconciliation settles the order from receipts and the account's nonce first, and only one transaction is in flight at a time. A pause stops new risk-taking and never sells anything by itself.
| Control | Enforcement | Description |
|---|---|---|
| Explicit activation | Backend code | Needs both an environment flag on the worker and an operator activation stored in the database. |
| Token and venue allowlist | Backend code | Only allowlisted markets on allowlisted venues; the signer independently checks the router, the token and the recipient. |
| Maximum position size | Backend code | Cap on the value held in one market. |
| Total exposure cap | Backend code | Cap across markets, never above the capital allocated to the champion. |
| Per-trade slippage limit | Backend policy; the limit itself is enforced by the router on chain | Sent as the swap's minimum output, which the router enforces on chain; the size of that limit is set by backend policy. |
| Daily loss limit | Backend code | Realized plus unrealized loss of the UTC day; when reached, no new risk is taken. |
| Gas spending cap | Backend code | Daily budget for transaction fees. |
| Quote age limit | Backend code | An order built on an old quote is rejected instead of being sent. |
| Circuit breaker | Backend code | Opens after consecutive failed or uncertain orders and stays open until an operator resets it. |
| Manual pause | Backend code | Stops new risk-taking. It never sells anything on its own. |
| Transaction simulation | Backend code | Each transaction is executed with eth_call against current state before it is signed; a revert cancels the order. |
No custom smart contract enforces these limits. Do not read them as on-chain guarantees. The route adapter for Uniswap v3 is unit-tested and checked by read-only simulation; it has never broadcast a transaction, and it is not described as production-ready. See the champion page for the current status.
Limitations
- Simulated results are not trading results. Costs, latency and liquidity are modeled. Real fills would differ and would generally be worse.
- Winning historically does not guarantee future returns. Selection picks what worked in a recent window. Markets change, and a short window on a few markets is weak evidence.
- Coverage is partial. Only Uniswap v3 pools paired with USDG or WETH are discovered and priced. Uniswap v2, Uniswap v4, Aggregators and RFQ are not covered, so prices here can differ from the best available price.
- Screening is not a safety audit. Holder concentration and token contract privileges are not checked. A simulated buy and sell shows behavior at one block.
- A cycle may be deferred when market data is unhealthy, when too few agents have enough history, or when no parent qualifies. A deferral is a recorded outcome, not an error.
- No champion may be active for long stretches. Qualification needs history, closed trades, a positive result after costs, validation on later data and human approval.
- Short windows are noisy. With a 20-minute cadence, many replacements are decided by small differences. The separation margin and the warm-up reduce this; they do not remove it.
- Sequencer finality. Observations use sequencer-confirmed blocks, which can be reordered until a batch is posted to Ethereum. Discovery stays behind the head and re-verifies after a reorganization.
Network sources
Every network value was read from a primary source and confirmed against the chain. Nothing is a remembered default. The check can be repeated with pnpm verify:network.
| Item | Value | Source | Verified |
|---|---|---|---|
| Chain ID | 4663 | docs.robinhood.com (opens in a new tab)eth_chainId returned 0x1237 from the public endpoint and from publicnode | 2026-11-16 |
| Gas token | ETH | docs.robinhood.com (opens in a new tab) | 2026-11-16 |
| Public RPC | https://rpc.mainnet.chain.robinhood.com | docs.robinhood.com (opens in a new tab)eth_blockNumber and eth_getLogs answered on 2026-11-16 | 2026-11-16 |
| Block explorer | https://robinhoodchain.blockscout.com | docs.robinhood.com (opens in a new tab) | 2026-11-16 |
| Observed block time | 0.104 s (measured) | measured from block headers (latest and latest − 6000) | 2026-11-16 |
| USDG | 0x5fc5360D0400a0Fd4f2af552ADD042D716F1d168 | docs.robinhood.com (opens in a new tab)decimals() = 6, symbol() read on chain | 2026-11-16 |
| WETH | 0x0Bd7D308f8E1639FAb988df18A8011f41EAcAD73 | docs.robinhood.com (opens in a new tab)decimals() = 18, symbol() read on chain | 2026-11-16 |
| Uniswap v3 factory | 0x1f7d7550B1b028f7571E69A784071F0205FD2EfA | developers.uniswap.org (opens in a new tab)eth_getCode: 24,535 bytes; PoolCreated logs served | 2026-11-16 |
| Uniswap v3 QuoterV2 | 0x33e885eD0Ec9bF04EcfB19341582aADCb4c8A9E7 | developers.uniswap.org (opens in a new tab)eth_getCode: 8,273 bytes | 2026-11-16 |
| Uniswap v3 SwapRouter02 | 0xCaf681a66D020601342297493863E78C959E5cb2 | developers.uniswap.org (opens in a new tab)eth_getCode: 24,497 bytes | 2026-11-16 |
| Multicall3 | 0xcA11bde05977b3631167028862bE2a173976CA11 | www.multicall3.com (opens in a new tab)eth_getCode: 3,808 bytes | 2026-11-16 |
AsvrllaLabs is an independent project. It is not affiliated with, endorsed by or sponsored by Robinhood.