Every Perps Agent run goes through the same on-chain phases on Mantle Sepolia. There is no “off-chain track record.”
1. Commit
Before the grid engine is allowed to place a single order:
- The AI loop finalizes the strategy config.
- The config and the user’s risk caps are hashed.
- The hash is committed to
StrategyLedger under a runId.
From this point on, the agent is locked to that config for that runId.
2. Trade
The grid engine runs on Bybit through the exchange adapter, using the trader’s trade-only API keys. Fills, PnL, drawdown, and risk metrics are tracked in real time. Nothing about this phase touches the proof layer.
3. Attest
When the run ends (target hit, stop hit, or user-stopped):
- The verified outcome (
pnl, maxDrawdown, riskScore, status) is hashed.
- The attestation is written to
StrategyLedger, referencing the original commit by runId.
- The
regime → params → outcome record is appended to StrategyMemory.
- Fees and performance settle through
Vault in native MNT.
4. Recall
The next time the AI loop proposes params:
- It queries
StrategyMemory for past attested runs on similar markets / regimes.
- It reads only attested outcomes — in-progress runs and uncommitted backtests are ignored.
- Those attestations become training signal.
Commit + attest + recall is the whole loop. Everything else — grid engine, Telegram UX, AI loop — is implementation around this on-chain skeleton.
