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

# Run the eval loop

> Carry one failure from trace to signed report.

# Run the eval loop

The Unified Eval Loop (`/eval-loop`) carries a single production failure through
14 evidence stages, ending in a signed report. It does not manufacture success
from aggregate counts — every stage must be backed by a real platform record.

## Step 1 — Start from a real trace

Open `/eval-loop`, click **Start from a trace**, enter a workflow name and the
numeric trace ID, and click **Create evidence chain**. The source trace becomes
version 1.

<img src="https://mintcdn.com/evalgate/c7l4ztAi8W3wgmvV/help/images/run-eval-loop-start.png?fit=max&auto=format&n=c7l4ztAi8W3wgmvV&q=85&s=899e9295ac98a6c070fe4bd2e9d53042" alt="A screenshot of the eval-loop start dialog with a trace ID entered." width="1440" height="900" data-path="help/images/run-eval-loop-start.png" />

## Step 2 — Advance each stage

Each stage requires canonical evidence from another surface. You do not type
evidence in — you create it on the surface it belongs to, then the loop
validates and advances.

| Stage                   | Where the evidence comes from                          |
| ----------------------- | ------------------------------------------------------ |
| Monitor sample          | Continuous Eval monitor                                |
| Scorer result           | A scorer run with a Model Gateway call and cost record |
| Alert and failure topic | The alert and topic the trace was grouped into         |
| Review item             | The review queue                                       |
| Candidate case          | Promoted from the trace                                |
| Human approval          | A review decision                                      |
| Active golden gate      | A passing gate run on the promoted case                |
| Playground reproduction | A Playground run with a selected winner                |
| Experiment winner       | An experiment with an accepted winner                  |
| Published artifact      | A deployable artifact version                          |
| Active deployment       | A deployment at a recorded revision                    |
| Online verification     | A fresh monitor run on new traffic                     |
| Signed report           | A shared, signed report                                |

## Step 3 — Handle a stuck stage

If a stage will not advance, the page tells you what evidence is missing or
invalid — stale, skipped, partial, foreign, mismatched, expired, revoked, or
provenance-incomplete. Fix the referenced record on its native surface, then
retry.

<Note>
  A completed stage is **immutable**. The workflow version increases by exactly
  one per accepted transition.
</Note>

## Step 4 — Seal the report

The final stage is a non-expired, non-revoked signed report that names the
workflow and seals the evidence body. Until that report exists, the loop is not
complete.

## Up next

[Share the signed report](./share-report) with reviewers.
