Limitations

Reproduction boundary

The original adjudicated record- and lead-level labels were unavailable. The project therefore reconstructs the main SQI-fusion pattern using public Set-A labels rather than claiming exact replication of the original evaluation.

Generated support is not native data

Support-aware construction reduces, but does not eliminate, source separability. A finite candidate bank cannot recover absent poor-quality mechanisms, and agreement in selected features does not establish equality in waveform distribution.

Architecture attribution is dataset-dependent

The Conformer separated from CNN and matched ResNet controls on the multi-lead Set-A task, but the matched ResNet was comparable on single-lead BUT. The general result concerns temporal localisation and continuous waveform evidence; architecture-specific gains remain conditional.

Diagnostic scope

Some local-degradation examples are illustrative and restricted to a small number of records. They identify plausible mechanisms but do not establish cross-subject generality.

No clinical-readiness claim

The experiments do not establish clinical deployment readiness. Required next steps include subject-independent and cross-device validation, broader patient-group coverage, repeated local-degradation audits, and evaluation of whether quality gating improves downstream clinical tasks.

Interpret results with the report

This site is an explanatory companion. The submitted report is authoritative for experimental definitions, uncertainty, and claims.