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

# Review and improve

> Understand the queue, inspect a flight, and make a controlled improvement.

## Find your flight

Open [Flights](https://irl.narev.ai/dashboard/flights/). A queued request is waiting for the simulator; it is not a completed flight yet. When a run appears, open it to see its status and available results.

After it finishes, start with the outcome and recording, then inspect the score and checklist. If it failed before flying, check the request error first. A failed flight can still help you identify the next improvement.

## Follow the request

After submission, return to your flights. Requests and runs have different lifecycles: a request can be queued before a simulator run exists. Request states include pending, queued, running, done, failed, and cancelled. A failed request can reflect setup or execution failure; it does not necessarily mean the aircraft crashed.

## Read the evidence

Open a flight to inspect the saved configuration, outcome, recording when available, and checklist results. Start with the first point where observed behavior diverged from the intended procedure. Compare the relevant cockpit evidence with the manual and the agent's recorded activity.

Recordings and scores depend on the artifacts saved for that run. An unavailable result is not a zero score. See [Flight scoring](https://www.narev.ai/docs/irl-gym/scoring.md) for checklist weighting, landing checks, crash caps, and ranking calculations.

## Diagnose the cause

| Problem | Next check |
| - | - |
| Flight never starts | Request status, harness/model selection, repository accessibility, and skill layout |
| Agent cannot find a reference | `search_docs` query, published page path, and `read_docs` response |
| Tool call is rejected | Required fields, enums, units, and the active tool schema |
| Action appears ineffective | Verification fields and a fresh reading of the relevant group |
| Checklist stays incomplete | The active execution, checkable item IDs, and explicit acknowledgement |
| Score differs from expectations | Final observed state and the published scoring rules |

## Improve one thing at a time

Keep the scenario, harness, and model constant while changing one skill. Select a fixed repository commit for each comparison. Repeat flights before drawing conclusions from a single outcome. Use recordings and checklist details to explain a change in score, rather than optimizing for acknowledgement alone.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.