# CHORE: eval-feedback-cluster — Evaluation Feedback Clustering

**Version:** 1.0.0
**Lane:** Medium
**Slug:** `eval-feedback-cluster`

---

## Overview

Periodically scan recent `adr-evaluation` ledger events and `gz-justify`
artifacts, cluster by recurring weak-dimension or confusion-shape patterns,
and emit structured proposal records when a pattern recurs ≥3 times across
distinct artifacts (ADR-0.0.26 Decision §3).

## Policy and Guardrails

- **Lane:** Medium — reads ledger and justify artifacts; no network required.
- **Read-only** at `.gzkit/ledger.jsonl` and `docs/design/adr/**`; only writes 
to its own proofs directory.
- No duplicate proposals: idempotent by content hash over `(cluster_key, sorted
source_artifact_ids)`.
- Threshold configurable via `data/eval_feedback_thresholds.json` 
(`cluster_min_recurrence`, default 3).

## Workflow

### 1. Run clustering

```bash
uv run -m unittest tests/chores/test_eval_feedback_cluster.py -q
```

### 2. Review proposals

```bash
ls .gzkit/chores/eval-feedback-cluster/proofs/
```

Proposals are JSON files: `proposal-<timestamp>.json` with schema:
`cluster_key`, `recurrence_count`, `source_artifact_ids`, 
`source_artifact_paths`,
`summary`, `proposed_rule_target`.

### 3. Validate layout

```bash
uv run gz validate --chores-layout
```

## Acceptance Criteria

| Type | Command | Expected |
|------|---------|----------|
| exitCodeEquals | `uv run -m unittest 
tests/chores/test_eval_feedback_cluster.py -q` | 0 |
| exitCodeEquals | `uv run gz validate --chores-layout` | 0 |

## Evidence Commands

