Implementation brief

## Who Uses It
A reviewer assessing submitted technical work against a published rubric.

## What You Will Build
Build an agent that organizes work samples, maps evidence to rubric criteria, and drafts review notes. Let human reviewers inspect the work, adjust criterion judgments, and publish their own feedback.

## Inputs
- Synthetic work submissions
- Published job-relevant rubric
- Reviewer comments

## Workflow
1. Remove irrelevant identity fields from the review view.
2. Extract evidence for each rubric criterion.
3. Draft criterion-level judgments and clarification questions.
4. Collect human edits and export an evidence-linked feedback report.

## Interface
A reviewer workspace with the submission, criterion evidence, editable judgments, and a feedback preview.

## Working Demo
Draft five points for R1 and zero for R2 with citations to the work, then let a reviewer edit the feedback.

## Deliverables
- A working local application with the specified review interface.
- Source code, dependency lockfile, and setup commands.
- A recorded end-to-end demo using the supplied starter data.
- Example outputs, action logs, and automated tests.

## Starter Data
These fictional records are a small development fixture, not a production dataset. Implement the mock services described in the brief.

```json
{
  "submission": {
    "id": "S1",
    "synthetic_name": "Person A",
    "text": "The solution returns the correct total. It includes no tests."
  },
  "rubric": [
    {
      "id": "R1",
      "criterion": "Correct result",
      "maximum_points": 5
    },
    {
      "id": "R2",
      "criterion": "Regression tests",
      "maximum_points": 5
    }
  ],
  "evidence": {
    "correct_result": true,
    "tests_present": false
  }
}
```

## Constraints
Use synthetic work. Keep judgments reviewable and do not automate real employment decisions.
Choose an available model that supports your implementation. Record its identifier and version.

## Evaluation
Review the complete application, not only the starter-case answers.
- The draft total is five and each judgment has evidence.
- Identity fields are absent from the scoring view.
- No hiring decision or applicant ranking is generated.
Add one normal case, one incomplete-input case, and one failure-recovery case.