Satellite active-fire products can repeatedly sample the same fire episode across sensors and overpasses, creating pseudo-replication and uncertain event-level labels. We developed an auditable workflow to reconstruct candidate fire events in the Tanjung Lago-Banyuasin II tropical peatland landscape, quantify Sentinel-1 corroborative support, and conduct an independent Landsat 8/9 optical audit while preserving unresolved evidence. MODIS and VIIRS detections were harmonized, confidence-filtered, and linked using prespecified pairwise limits of 1 km and 24 h. Of 437 harmonized detections, 415 were retained and reconstructed into 218 candidate events. A Sentinel-1 decision rule was developed on 18 clusters, frozen before evaluation, and applied unchanged to a stratified random sample of 60 clusters. Following robustness and leave-one-acquisition-out analyses, 15 events were provisionally corroborated, 24 partially corroborated, and 21 unresolved; 56 of 60 primary decisions were stable. The design-weighted estimated proportion with at least partial radar support was 57.58% (95% CI 44.41–70.76%). An independently specified Landsat 8/9 audit, with thresholds frozen before unblinding, identified seven REF-FIRE events, all within the Sentinel-1-supported group; no event met the conservative REF-NOFIRE criteria, and most optical outcomes remained unresolved because imagery was insufficient or inconclusive. The framework reduces repeated counting while keeping detection provenance, corroborative evidence, and downstream label eligibility distinct. It is intended for auditable event reconstruction and label governance rather than fire-detection or burned-area accuracy assessment.