Annotation, QC, And Audit Evidence
This page summarizes the evidence that is directly inspectable from the workbook, manifest, and bundled documentation. It stays within what is actually documented in the public release.
Auditable evidence
| Release aspect | Evidence in bundle | Count or status | How to interpret it |
|---|---|---|---|
| Bundle identity | cantharm_release_manifest.json, cantharm_release_readme.md |
2026-04-02 public release |
Defines the single authoritative public release line used by the paper and this site. |
| Form-level manual label review | label_review_flag |
1376 form rows | Shows that a substantial subset of released forms was manually reviewed rather than untouched source carry-over. |
| Variant / normalization review | variant_rule_manual_review |
67 sense rows | Confirms that some variant handling decisions were explicitly reviewed. |
| Multi-source provenance fields | source_dictionary, source_status |
Present in released sense rows | Source information remains visible in the public workbook for audit and interpretation. |
| Form and sense layers | Finalized form targets and linked sense fields | Both released | The benchmark provides a finalized form layer for scoring and a linked sense layer for interpretation. |
| Pronunciation normalization | jyutping, jyutping_status, ipa |
Present in released form rows | Pronunciation is part of the released benchmark object. |
| Documentation pages | Site pages, README, reproducibility note | Included | The public bundle includes written guidance for scope, protocol, access, ethics, and release interpretation. |
Severity note
Severity is released as a scalar target for benchmark evaluation, but it should not be read as a universal or context-independent measure of harm intensity. It is a rubric-driven score aligned to this benchmark object.
What is intentionally not claimed