5 MIN READ

Policy, limits & approval

The control model shared by Tide and Keel: models propose, policy decides, you approve.

The principle

Models propose. Policy decides. You approve.

Language models can reason about markets, but they can also be confidently wrong. So authority is split: the model drafts, deterministic code enforces limits, and a human signs.

A policy, concretely

A policy is a short, human-readable config. This is an illustration of the shape, not a shipped schema:

{
  "chain": "base",
  "pool": "WETH/USDC",
  "range": { "widthBps": 400, "rebalanceBandBps": 120 },
  "maxPositionUsd": 5000,
  "maxSlippageBps": 30,
  "allow": { "tokens": ["WETH", "USDC"] },
  "approval": "required",
  "killSwitch": true
}

Where limits are enforced

Limits are checked by code after the model answers and before anything reaches your wallet. A proposal that breaks a limit is rejected, whatever the model said, so a model cannot argue its way past a size cap.

The guard ladder

LevelBehaviour
Within limitsNormal operation. Actions still need approval.
WidenVolatility spikes: wider range, smaller size.
PauseStop proposing until conditions normalise.
KillStop everything. A human restarts.

Status

This is the model Tide and Keel are being built around. The agent runtime that exists today is described in Agent runtime today. A planned way to enforce part of it on-chain is described in The hook layer.