July 9, 2026

Six new Codex research-tier datasets on Hugging Face

Codex now publishes six additional research-tier datasets on the Axiom AI Hugging Face organization, each a deterministic 100K-record stratified sample in Parquet under CC-BY-4.0:

• AIS Maritime
• Civic Intelligence — Events Timeline
• OSHA Safety
• LEHD Commuter Flows (workplace jobs)
• POI Intelligence
• Permit Signals

Every dataset ships with a README, schema documentation, and the standard APRS envelope — so the loading patterns in the Codex notebooks work against any of them without changes.

Fixes

• Locus: The building-inspection records dataset now carries the correct source_uri on every row across all four covered metros — Chicago, Los Angeles, New York City, and San Francisco. source_uri is now derived per-row from each row's source_name (its Socrata domain/dataset_id) rather than assumed to be a single hardcoded URL, so citations, re-fetches, and downstream joins from Codex resolve to the correct metro's Socrata endpoint. Today's 4,299 rows (all Chicago at the time of the fix) are unchanged; the correction is forward-looking — LA, NYC, and SF rows landing from the weekly --all scout run will be tagged with their own metro's source URL from ingest onward.
• Locus: USDA SNAP retailers — the nationwide food-retailer roster that feeds the food access sub-signal in the amenityDemand group — is now deduplicated and kept fresh incrementally. A silent failure in the previous monthly reload had let each snapshot stack on top of the last instead of replacing it, leaving 544,419 excess rows across 256,669 duplicate retailer IDs. The table has been cleaned to one row per retailer, a uniqueness constraint on the retailer ID prevents recurrence, and the refresh job now upserts by retailer ID with resumable checkpoints instead of truncate-and-reload — so partial-run failures no longer leave the table empty or duplicated. No action is required on your part; food-access scoring is now driven by an accurate retailer count and the roster stays current between monthly USDA publications.
• Overwatch: The analytical charts on Forecasts, Estimates, Fleet (age histogram and age-trend), Berth utilization, seasonality panels, ownership-graph visualizations, and the Locus score-history panel on the locus-validation view now render cleanly during layout transitions. Three earlier passes had guarded the terminal-panel charts against a Recharts zero-dimension render warning, but ten charts elsewhere in the app were still unguarded and could briefly render at negative width or height (or flash a console warning) whenever their container measured before layout settled. All chart surfaces are now guarded uniformly, and a repository-wide check keeps new charts from regressing.