Boston Food Establishment Inspection Results and Findings in OCEL 2.0

This dataset contains two object-centric event logs of food-establishment inspection results in Boston, Massachusetts, USA. Both logs are provided as gzip-compressed OCEL 2.0 JSON files (.json.gz) for process-mining research and teaching. They connect recorded inspection outcomes to food-establishment licenses and properties; the detailed log additionally links individual finding observations to their inspections. The logs were extracted from the City of Boston's Food Establishment Inspections dataset on Analyze Boston, maintained by the Inspectional Services Department. The extraction uses a frozen snapshot of 907,385 source rows collected on 2 November 2026. The two files provide complementary views of the same source: boston-inspections-history.json.gz contains 220,270 events spanning 4 April 2006 to 31 October 2026, with 18 event types and 236,108 objects: 220,270 Inspection objects, 11,324 FoodLicense objects, and 4,514 Property objects. It supports analysis of longer result sequences and relationships between licenses and shared properties. boston-inspections-findings-2022-2025.json.gz contains 60,304 events recorded during 2022–2025, with 13 event types and 239,811 objects: 60,304 Inspection objects, 8,341 FoodLicense objects, 3,911 Property objects, and 167,255 Finding objects. It supports analysis of inspection outcomes together with recorded finding codes, descriptions, severity strings, and snapshot statuses. The logs overlap and must not be concatenated as independent datasets. The detailed log contains the historical log's eligible events for 2022–2025, enriched with finding objects and relationships. Common events and objects retain the same identifiers. The public source does not expose an inspection identifier. Each event is therefore reconstructed by grouping rows with the same license number and exact result timestamp. Multiple finding rows become one result event rather than separate inspections. Activities preserve the source outcome code in the form Inspection result: <code>. Event timestamps are normalized to UTC, retaining fractional seconds. Qualified event-to-object and object-to-object relationships express which license was inspected, the recorded property, and which findings belong to the inspection. Each event and finding retains its contributing source row identifiers for traceability. License metadata and finding attributes are explicitly represented as snapshot observations at extraction time. They are not assumed to describe historical state at every inspection. Finding objects represent individual recorded observations; the source provides no stable identifier for linking a violation episode across inspections. Source row and finding identifiers are specific to the frozen snapshot. Quality controls excluded 6,528 source rows (0.72%): 6,398 with missing or unusable result timestamps, 88 belonging to 20 groups with conflicting outcomes at the same license and timestamp, and 42 carrying the explicit DATAERR result code. The remaining 900,857 rows form the historical log. Missing property identifiers do not exclude events; those events retain their license relationships without an inferred property link. Validation includes full source-to-export reconciliation, checks of declared attribute types, timestamps and graph integrity, and PM4Py import and object-centric directly-follows discovery. Schema validation uses a documented correction to the legacy OCEL schema's string-only attribute-value constraint. Independent offline rebuilds produced byte-identical compressed exports. The logs support outcome-sequence discovery, analysis of intervals to the next observed result, comparisons of license and property perspectives, and study of recorded finding composition. Administrative outcomes such as HE_NotReq are retained and include simultaneous batches across licenses. Accordingly, recorded events should not all be interpreted as on-site visits. Inspection identities are inferred, observation windows truncate traces, and a subsequent pass does not establish a remediation episode or repair duration. These data do not establish food-poisoning incidence or causality. The companion repository at the extraction revision provides extraction and validation code, the checksummed source snapshot, quarantined-row provenance, reproducibility instructions, and a process-mining quality report. The frozen source metadata identifies the underlying municipal dataset's license as the Open Data Commons Public Domain Dedication and License (PDDL); the repository's software license is separate. Acknowledgements We thank Wil M.P. van der Aalst ([e-mail]) and Jean-Christophe Schmitt ([e-mail]) for suggesting the source data for these event logs. Jean-Christophe Schmitt also performed an initial extraction.