annotation/PROTOCOL.md

MoraLink moral-grouping protocol

Purpose

MoraLink starts from 709 fables, each paired with a human-written moral in MORABLES. Several fables share identical or closely equivalent lessons. The grouping process prevents a retrieval system from being marked wrong when it returns a different fable that expresses the same moral.

The final benchmark contains 668 unique moral queries arranged in 558 groups. Of these groups, 469 contain one fable and 89 contain more than one fable.

Who made the decisions

Three annotators worked together: one PhD researcher, one PhD candidate, and one graduate student. All three have academic research backgrounds and are proficient in English. They discussed candidates jointly and produced one agreed grouping.

The process did not produce three separate sets of labels. Therefore, agreement between independent annotators cannot be calculated retrospectively.

How candidates were found

Candidate generation was broader than the final decision process:

1. Exact duplicate moral text was identified directly.
2. All moral pairs were compared using BGE-large-en-v1.5 sentence embeddings. High-scoring pairs were surfaced for review.
3. Claude Sonnet 4.5 was used as an assistant to propose possible merges and splits using the moral statements, fable texts, similarity scores, and current grouping state.
4. The team also read remaining singleton morals to find plausible groups that automated similarity searches had missed.

The embedding scores and model suggestions were candidate generators only. They did not determine the final labels.

How a decision was made

The annotators asked whether the morals expressed the same underlying lesson, not only whether they mentioned the same topic or used similar words.

A proposed group could be:

• accepted as proposed;
• split into narrower groups;
• expanded with additional fables;
• rejected, leaving its members as singletons or in their existing groups; or
• moved when a moral fit another group more closely.

The annotators discussed non-obvious cases together until they reached a shared decision. The candidate-review cycle was repeated while further changes were still warranted.

For example, the proposal based on the word "necessity" was rejected because its three morals made different claims: necessity can override rules, intensify effort, or encourage invention. In contrast, the adversity_reveals_friends group was accepted because both morals state that hardship exposes whether a friend is sincere.

Moral-text edits

During review, 23 moral texts were corrected or clarified. Every one of these records keeps the previous wording in an original_text field. The edits include punctuation and spelling fixes, small wording corrections, and a small number of changes made to better match the fable or the accepted group.

For example:

• Flatters are not to be trusted. was corrected to Flatterers are not to be trusted.
• False men cannot be trusted. was changed to Sometimes, those you support become your worst enemy. when the fable was moved to the support_becomes_enemy group.

MORAL_EDITS.csv contains all 23 original and final texts together with the preserved reason for each edit.

Preserved records and final authority

The final group assignment is analysis/clusters_full.json, mirrored in data/clustered/cluster_mapping.json. These files contain 558 groups and cover all 709 fables exactly once.

analysis/cluster_review_progress.md and analysis/singleton_cluster_proposals.md preserve many accepted and rejected decisions and their reasons. The progress document was not updated after the last finalization changes and still reports 563 groups. That number is obsolete. The release CSV is reconciled against the frozen 558-group mapping so the stale count is not propagated.

For final accepted groups whose review note survives, DECISION_LOG.csv points to that note. Exact groups are verified from identical final moral text. If a final near group has no cluster-specific prose rationale in the surviving notes, the row is marked author_confirmation_required rather than inventing a reason.

Limitations

• A language model helped propose candidates, although humans made the final decisions.
• The annotators worked through joint discussion and consensus rather than independent labeling.
• No separate independent validation of the final grouping was performed.
• Some review evidence survives as prose summaries rather than a complete action-by-action transcript.
• Two finalization splits lacked prose rationales in the earlier notes; their reasons were confirmed by an original annotator and are preserved in AUTHOR_CONFIRMATIONS.json.