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

# The verification harness

> The independent judge between agent intent and execution.

A guard that only checks hard limits can't read **intent**. It waves through an action that breaks the spirit of the rules while staying within the letter of them — and when it does block, it leaves only a log line you have to take on faith.

Chance is a **judge, not a guard**: it reads the intent and the proposed action and rules on whether they actually match — and the ruling is provable.

## PROPOSE → VERIFY → SETTLE

<Steps>
  <Step title="PROPOSE">
    Your agent constructs an action from your intent.
  </Step>

  <Step title="VERIFY">
    An independent model checks the action against the intent — both the hard constraints and the plain-English meaning — and returns a verdict with its reasoning. This is where Chance lives.
  </Step>

  <Step title="SETTLE">
    Only a passing verdict clears to execution. Fail-closed: a failed check executes nothing.
  </Step>
</Steps>

## The verdict

Every verification returns exactly one of:

| Verdict    | Meaning                                   | What to do       |
| ---------- | ----------------------------------------- | ---------------- |
| `ALLOW`    | The action faithfully matches the intent. | Execute.         |
| `BLOCK`    | The action violates the intent.           | Do not execute.  |
| `ESCALATE` | Genuinely ambiguous — needs a human.      | Hold for review. |

Each verdict comes with a plain-English **reasoning** trail, so a rejection is something you can read and audit, not just a label.

## What it catches that a numeric guard can't

* **Intent drift** — an action that passes every limit but breaks the strategy's theme.
* **Side-inversion** — an agent about to take the *opposite* action to its stated reasoning.
* **Spirit-vs-letter gaps** — technically-compliant actions that betray what you actually asked for.

## Beyond a single check

The same judge generalizes past one action: review a batch of recent actions for drift, grade risk on a payload, or resolve a disputed outcome. Every one runs through the same pipeline and produces the same kind of provable verdict — see [Provable verdicts](/concepts/proofs).