| # MoraLink moral-grouping documentation |
|
|
| This directory packages the surviving evidence for the construction of MoraLink's moral groups. The frozen benchmark files remain the source of truth for the final data. The documents here explain how those files were produced and preserve the available decisions and moral edits in easier-to-review formats. |
|
|
| ## Files |
|
|
| - `PROTOCOL.md`: Plain-language description of candidate generation, joint review, decision criteria, consensus, and limitations. |
| - `DECISION_LOG.csv`: One row for each of the 89 final non-singleton groups, followed by 66 explicit rejected decisions found in the preserved review records. |
| - `MORAL_EDITS.csv`: All 23 morals that retain both `original_text` and final `text` in the processed corpus. |
| - `AUTHOR_CONFIRMATIONS.json`: Original-annotator confirmations for final decisions whose prose rationale was not preserved. |
| - `../validate_annotation_release.py`: Rebuilds the CSV files from the frozen data and preserved records, then validates their coverage and counts. |
|
|
| ## Sources of truth |
|
|
| - `analysis/clusters_full.json`: Final 558-group assignment. |
| - `data/clustered/cluster_mapping.json`: Released form of the final group assignment. |
| - `data/clustered/morals_unique_corpus.json`: Final 668 moral queries. |
| - `data/processed/morals_corpus.json`: Final moral text and the preserved original text for 23 edits. |
| - `analysis/cluster_review_progress.md`: Joint-review decisions and moral-edit reasons. Its summary count is stale and must not be quoted. |
| - `analysis/singleton_cluster_proposals.md`: Later accepted, split, and rejected candidate groups. |
| - Git history: Evidence for two finalization splits whose prose reasons were not recorded in the Markdown notes. |
|
|
| ## Current status |
|
|
| The package reconciles to 558 groups, including 469 singleton groups and 89 non-singleton groups, 668 unique moral queries, 709 fables, and 23 edited morals. The two finalization splits whose prose rationales were missing from the earlier notes were confirmed by an original annotator and are recorded in `AUTHOR_CONFIRMATIONS.json`. |
|
|
| In `DECISION_LOG.csv`, `query_ids` are the final `moral_unique_*` benchmark queries for which the listed fables are relevant. Accepted-row membership always comes from the frozen 558-group mapping. Reasons and source locations come from the surviving review notes or, for the two confirmed finalization splits, `AUTHOR_CONFIRMATIONS.json`. Rejected rows document the explicit decisions that survive in the notes; they are not presented as a complete action-by-action transcript of every discussion. |
|
|
| ## Rebuild and validate |
|
|
| Build a reviewable draft: |
|
|
| ```bash |
| ./run.sh review_response/23_camera_ready/validate_annotation_release.py --build --allow-unresolved |
| ``` |
|
|
| Validate the public release: |
|
|
| ```bash |
| ./run.sh review_response/23_camera_ready/validate_annotation_release.py |
| ``` |
|
|
| The public-release validation fails if an `author_confirmation_required` row remains. |
|
|
