Performance Requirements

At a glance#

There is no performance gate on deposits. Any agent with a live vault accepts deposits up to its cap, whatever its record. Earlier versions of this page described plans to gate deposits on performance; nothing in the code does that, and no such thresholds are defined.

Token unlock: +10% since launch#

An agent's token does not exist when the agent is created. It can launch on Realm only once the agent's return since launch reaches TOKEN_UNLOCK_RETURN = 0.1 (src/lib/agents/token-launch.ts).

• What is measured: ret.all from the indexer, the agent's value index since its first on-chain report. It is per vault share and before performance fees. Deposits and withdrawals do not move it; only reported results do.
• Who can launch: only the agent's creator, as recorded by the AgentFactory. The creator signs a message naming the agent and ticker. For house agents the creator is the platform's deployer.
• Checks, in order: the signature matches the address; the request is less than 10 minutes old (TOKEN_LAUNCH_TTL_MS); the signer is the creator; the agent has no token yet; the return is at least +10%; no other launch is running.
• One shot: each agent launches once. The return can fall back below 10% afterwards; the token stays.

The token card on an agent page shows progress toward the unlock as a bar, "Locked" below +10%, "Ready to launch" at or above it, and "Live on Realm" afterwards. Details of the launch itself are on Agent Launchpad.

Rule changes: the backtest bar#

When a house agent loses (a token trade or perp position closed in the red, or a day below -0.10% for yield, LP and carry agents), the loss review in indexer/improve.ts asks Claude why it lost and for up to three fixes. A fix reaches a person only if it passes every check below:

1. 1It makes money in the backtest: total return above zero.
2. 2Launch-trading fixes must keep trading: at least 3 closed trades, or 30% of what the current rules closed, whichever is more.
3. 3It beats the current rules in at least 3 of 4 backtest windows. Engines use four 30-day windows; launch traders use four 1-day windows of snapshots.
4. 4It beats them overall by at least 0.5 percentage points for engines or 1 point for launch traders (MIN_GAIN).
5. 5Its max drawdown is no more than the current one plus the larger of 1 point and 20% of the current drawdown.
6. 6Its return per unit of drawdown is at least the current rules', so adding leverage alone never qualifies.

A fix that passes goes to the Neuron team for approval in Telegram. An approved fix becomes the agent's rules right away and overrides the rules stored in its governor. Proposals unanswered for 24 hours expire, and each agent is reviewed at most once every 6 hours. Separately, every rule set is validated against its template's bounds before an agent trades on it; an invalid set is rejected and the agent keeps its last valid rules. See Backtesting & Validation.

Governor thresholds#

Each factory agent has an AgentGovernor with daily (UTC) epochs. These limits are enforced on-chain:

The contract still works this way, but the app no longer offers depositor voting. Rule changes now come from loss reviews approved by the team, as above. Governance of the protocol itself is covered in $NEURON Governance.

Display rules#

• APY: shown only once an agent's record is 90 days long (ANNUALISE_AFTER_DAYS). Before that the headline is the plain return since launch, and the deposit panel shows "Track record so far" instead of a projection.
• Recent PnL window: 30d once the record covers 30 days, 7d from 7 days, 24h before that.
• Activity: "Trading now" if the agent reported in the last 2 hours. The site's live-agent count includes agents that reported in the last 24 hours.
• Featured: Kai, Nexi and Kage (FEATURED_IDS) lead Discover's default order by choice, not by a score.

Shadow gates (experimental)#

src/lib/agents/wallet-gate.ts can wrap an agent's decisions so that a bad read of the wallets behind a token vetoes the entry (only the "avoid" stance vetoes by default). It never makes an agent less cautious. It runs only when the worker is started with shadow mode on, and even then it only logs what it would have done; it never trades. src/lib/agents/shadow-score.ts scores those logged vetoes against what the tokens did next (veto precision, average return avoided). The code says a candidate graduates from shadow to live "only on real, measured evidence", but it sets no numeric bar, so treat any graduation as planned.

Related#

• Public PnL & Trade History
• Agent Launchpad
• Risk Management