> ## Documentation Index
> Fetch the complete documentation index at: https://docs.ameo.agiwithai.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Policy specification

> Formal guardrail predicates evaluated outside the LLM.

Predicates are evaluated **outside** the LLM; the model cannot bypass them. Each [policy predicate](/glossary#policy-predicate) is a boolean rule enforced before any transaction is signed.

## System diagram

```mermaid theme={null}
flowchart LR
  O[Observation] --> P[Plan]
  P --> G{Guardrail Predicates}
  G -- all true --> X[Execute]
  G -- any false --> R[Reject + log PolicyRejected]
  X --> S[Settle on Mantle Sepolia]
  S --> Z[0G Receipt]
```

Settlement happens on [Mantle Sepolia](/glossary#mantle-sepolia); reasoning traces anchor to [0G Storage](/glossary#0g-storage).

## Predicate set

| Predicate                      | Invariant                                                        | Code reference                                                 |
| ------------------------------ | ---------------------------------------------------------------- | -------------------------------------------------------------- |
| `drawdown_ok(state, plan)`     | `pnl_24h_pct ≥ -MAX_DRAWDOWN_PCT`                                | `policy.py` — drawdown check in `PolicyEngine.validate`        |
| `whitelist_ok(plan)`           | `∀ token ∈ plan.actions, token ∈ ALLOWED_TOKENS`                 | `policy.py` — `asset_in_not_allowed` / `asset_out_not_allowed` |
| `trade_size_ok(plan)`          | `∀ a ∈ plan.actions, a.usd_notional ≤ MAX_TRADE_SIZE_USD`        | `policy.py` — `max_position_exceeded`                          |
| `slippage_ok(plan)`            | `∀ a ∈ plan.actions, a.expected_slippage_bps ≤ MAX_SLIPPAGE_BPS` | `settings.py` — `DEX_SLIPPAGE_BPS` at execution adapter        |
| `daily_volume_ok(state, plan)` | `today_volume + plan.notional ≤ MAX_DAILY_VOLUME_USD`            | `graph.py` — daily volume cap in `act`                         |

## Composition

```
plan_approved ⇔ drawdown_ok ∧ whitelist_ok ∧ trade_size_ok ∧ slippage_ok ∧ daily_volume_ok
```

Additional runtime checks in `GuardrailService.check_plan`: observation quality, balance sufficiency, gas spike guard, protocol whitelist.

## Failure mode

Any predicate false → plan rejected → `guardrail_evaluated` event with `violations[]` → surfaced in replay Policy validation node. No on-chain execution proceeds.

## Design principle

Most agent stacks treat policy as an LLM judgment call. AMEO separates **cognition** (LLM or rules planner) from **authorization** (deterministic predicates). A fluent rationale cannot override a failed drawdown or whitelist check.

To add a custom predicate, see [Write a policy](/guides/write-a-policy).