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

# Results Vault, Run Inspector, and the quality flywheel

> Inspect workflow run traces, approve or reject paused runs, download artifacts, and improve outputs with the evaluate and refine quality loop.

Every workflow run in Agent Lab produces verifiable results: step-by-step execution traces, tool call telemetry, and tangible output artifacts such as posts, briefs, and reports. Three surfaces expose them:

* **Results Vault**: the artifact library in the Command Center, including a content calendar view of scheduled posts.
* **Run Inspector**: a modal that shows the full execution trace for any run. Open it from the Dashboard, the Auditing page, or the Command Center.
* **Quality flywheel**: an evaluate-and-refine loop that scores every artifact against a quality rubric and rewrites weak outputs with AI.

Use this page to verify what a run actually did, gate runs that wait for approval, and raise artifact quality before anything ships.

## Inspect a run

Click **Inspect Run & Tools** on a run in the Dashboard, **Inspect Execution Run** on an audit entry in the Auditing page, or **Inspect** on a run in the Command Center. The Run Inspector opens with four tabs:

| Tab                  | What it shows                                                                                             |
| :------------------- | :-------------------------------------------------------------------------------------------------------- |
| **DAG Steps**        | Every step in execution order with its title, step type, assigned agent, and status.                      |
| **Tool Evidence**    | Each tool call the agents executed during the run, so you can verify claimed work against real telemetry. |
| **Output Artifacts** | Artifacts the run produced, each with a **Download** action.                                              |
| **Context JSON**     | The raw context payload passed through the run.                                                           |

The inspector also aggregates run-level metrics: total tool calls, total cost, and total latency across all steps.

To fetch the same data programmatically, call `GET /api/runs/:runId`. See the [Runs API](/api-reference/runs).

## Approve or reject a paused run

Workflows that include an approval gate pause with the status `paused_for_approval` and wait for an operator decision. From the Run Inspector (or the run list in the Command Center):

* **Approve & Resume** sets the run back to `pending` and immediately resumes execution.
* **Reject Run** marks the run `failed` with the error message `Rejected by user` and ends it.

Both actions only apply to runs in `paused_for_approval`. Approving or rejecting a run in any other status returns HTTP `400`.

Programmatic equivalents: `POST /api/runs/:runId/approve` and `POST /api/runs/:runId/reject`.

## Browse the Results Vault

The Results Vault lives in the Command Center under **Content Calendar & Scheduled Posts**. It lists workflow artifacts across your workspace and tracks each artifact through its lifecycle:

* `draft`: generated but not yet queued.
* `scheduled`: queued with a `scheduledFor` timestamp.
* `published`: shipped to its target platform.
* `archived`: retired.

Each artifact carries an `artifactType` (`post`, `calendar_entry`, `document`, `file`, `crm_diff`, or `csv`), a `targetPlatform` (for example `linkedin`, `blog`, `newsletter`, `hubspot`, or `internal`), and its quality score.

The content calendar view pulls all `post` artifacts ordered by scheduled date and totals drafts, scheduled items, and published items. Fetch it directly with `GET /api/artifacts/content-calendar`.

### Download an artifact

Click **Download** on any artifact in the Run Inspector's **Output Artifacts** tab or in the Results Vault. Agent Lab serves the raw content as a file: `csv` artifacts download as `.csv`, everything else as `.md`. The API equivalent is `GET /api/artifacts/:id/download`.

### Change artifact status

Move an artifact through its lifecycle (for example `draft` to `scheduled`) from the Results Vault, or call `PATCH /api/artifacts/:id` with a new `status` and optional `scheduledFor` timestamp.

## The quality flywheel

Every artifact is scored 0-100 against a four-part rubric:

| Rubric dimension  | What it checks                                                              |
| :---------------- | :-------------------------------------------------------------------------- |
| Brand alignment   | Concise, authentic voice with no corporate boilerplate or filler phrases.   |
| Actionable CTA    | A clear next step for the reader.                                           |
| Factual integrity | No unpopulated template placeholders like `[INSERT NAME]` or `{{company}}`. |
| Formatting        | Structure appropriate for the target platform.                              |

The composite score maps to a grade: `A` (88+), `B` (75-87), `C` (60-74), `F` (below 60). An artifact passes when it scores 70 or higher and contains no placeholders. Scores, grades, feedback, and suggestions are stored on the artifact in `verificationNotes`.

### Evaluate

Click **Re-evaluate** on an artifact in the Command Center, or call `POST /api/artifacts/:id/evaluate`, to re-score it on demand. Evaluation updates the artifact's `qualityScore`, `qualityGrade`, and `verificationNotes` in place.

### Refine

Use **Refine with Agent** in the Command Center's **Agent Self-Correction & Refinement** panel, or call `POST /api/artifacts/:id/refine`, to rewrite an artifact. You can pass operator instructions, or leave them empty to let the flywheel refine based on the artifact's own quality feedback.

Refinement:

1. Evaluates the original content.
2. Rewrites it with Gemini (`gemini-2.5-flash`) when a Google AI API key is configured. Without a key, it applies deterministic cleanup that strips placeholders and filler phrases.
3. Re-evaluates the rewritten content.
4. Saves the result as a **new artifact revision** with an incremented `revisionVersion` and a `parentArtifactId` pointing at the original. The original artifact is never overwritten.

Repeat evaluate and refine until the artifact reaches the grade you want, then schedule it from the Results Vault.

## API access

All Results Vault and Run Inspector data is available over the REST API:

* [Runs API](/api-reference/runs): list runs, fetch full run details, approve and reject paused runs.
* [Artifacts API](/api-reference/artifacts): list, download, evaluate, refine, and update artifacts, plus the content calendar feed.