# Evidence

> Every answer carries the evidence it was grounded in.

**Evidence** is what an agent’s answer or finding rests on: the passages it was given, the claims it made from them, and how strong the source behind each claim is. The engine records it for every cited answer and every findings run; LightHouse shows the counts.

## Evidence tiers

Every claim carries the tier of its strongest evidence. The default ladder, for engineering work, runs from strongest to weakest:

| Tier | Source                                                               |
| ---- | -------------------------------------------------------------------- |
| T0   | Live: observed in the running system (monitoring, logs, a live call) |
| T1   | Declared deployment: what each environment is configured to run      |
| T2   | Code at the deployed version                                         |
| T3   | Code on the default branch: what is coming, not what runs            |
| T4   | Verified contracts: interfaces generated from code or checked in CI  |
| T5   | Governed decisions: approved designs, release notes                  |
| T6   | Descriptive documents nobody checks against the code                 |
| T7   | Conversations: meetings, chats, handover notes                       |

A pack may define its own ladder (a compliance pack might rank regulation first), with the same shape: lower number, stronger evidence.

## What is recorded

For each answer, finding or set of citations, the deployment sends counts: claims by tier, uncited claims, claims at the ladder’s strong tiers, and the strongest tier. Never a claim, a source or the ladder’s wording.

Separately, the **grounding** check records whether each answer was `grounded`, `flagged` or `refused`, with the number of claims and how many were ungrounded.

## Retrieved text is evidence, never instruction

Every passage an agent is given is fenced as untrusted material, and the instruction to treat it as material, not orders, sits in the system part of the prompt. Text in a document that says “ignore your instructions” reaches the model only inside that fence. What an agent may do is enforced at run time by the [gateway](/concepts/gateway/) and the tool gate, not by the prompt. See [MCP security](/security/mcp/).

## Evidence and approvals

When a person approves an output, LightHouse records the digest of the artefact they were shown. If the artefact changes before the decision reaches your deployment, the decision is refused (`evidence-changed`). See [Approvals](/concepts/approvals/).

## In LightHouse and the API

- Each agent’s detail page shows its evidence scores; the evidence pack (**Compliance → Evidence pack**) is the report an assessor reads.
- `GET /api/v1/evidence` and `GET /api/v1/grounding`: see the [Read reference](/api/reference/read/#evidence-read).
- MCP: [`evidence`](/mcp/tools/evidence/).

Source: https://docs.atlanticagenticlabs.com/concepts/evidence/