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

# Detection

> Flagging unreliable agent decisions, and the evidence behind each flag.

Airlock does not classify your content or score truthfulness. It applies
checks a reviewer can verify by hand, to the one place an agent's output has
consequences attached: its [decisions](/decision-events).

## Decision flags

Three checks, each a comparison rather than a judgment:

| Flag             | Raised when                                                 | Why it matters                                                       |
| ---------------- | ----------------------------------------------------------- | -------------------------------------------------------------------- |
| `no_evidence`    | The decision cites no sources                               | An unevidenced conclusion is indistinguishable from a fabricated one |
| `low_confidence` | Self-reported confidence is missing or below your threshold | The agent is telling you to look                                     |
| `disagreement`   | Two decisions on the same `target` do not agree             | One is wrong, and only a human can say which                         |

Flagged decisions land in the console's
[review queue](/dashboard#the-flags-queue) with the decision's sources,
quotes, reasoning, and originating session attached — and carry an
append-only review status with reviewer and note.

<Tip>
  `disagreement` is the cheapest cross-check available: run two independent
  resolvers against the same `target` and let their deltas surface
  themselves. No shared prompt, no shared failure mode.
</Tip>

## Access is logged, not classified

The other two event kinds — [access and chat](/capture) — are recorded in
enough detail that a reviewer can answer *who asked, through which
connector, and what came back* after the fact. Airlock does not inspect or
redact that content in the request path.

That boundary is deliberate. An agent answering a question it should not
have been able to answer is visible in the record within seconds, and
reviewable in the console; it is not blocked mid-flight. Enforcement is a
dial you turn later, at the [boundaries](/governance#action-gates) where an
action cannot be taken back.

## Evidence

Everything Airlock raises carries the material it rests on:

* **The failed check** — exactly what fired.
* **The cited sources** — url, title, and the verbatim quote the agent
  relied on.
* **The identity** — the user and agent session behind the event.

A flag is evidence about a *decision*, not a verdict on a person. The
disposition belongs to a reviewer, and that review is itself recorded.