How has the city’s record changed since the Archive began?

Method

Counted: every row in the Archive — the committed accumulation of every row any Snapshot has held (ADR-0004, ADR-0005), each stamped with the pull that first held it and, once it leaves the source, the pull that last did. "At first pull" is the rows stamped with the dataset's earliest pull day; "added" arrived on a later pull; "dropped" carry a last-seen stamp. "Pulls" counts the distinct days appearing in the stamps, not pulls: several pulls on one day count once, and a pull that changed nothing leaves no stamp; "First" and "Latest" are the earliest and latest stamp days — the last day a dataset changed, not its last pull. Days are UTC calendar days, which run ahead of Central time in the evening: a pull the Sources date by its Central day can appear here dated the day after. Excluded: Police Calls, which is archived by year and is not in this table; ADR-0006 covers its identity. Not charted: none — the evidence is two tables.

A row's identity in the Archive is its content with the city's OBJECTID removed, so a row that changes in any field is a new row and the old one is dropped. Where a layer declares an identity in the city's own fields — four Inspections tables do: a site by its SITE_INV, a violation by unit, inspection control number, and category, repeats of those three numbered in the city's row order — a dropped row with a live successor of the same identity is counted as edited. Those layers also publish a rolling window (ADR-0004: about a year of violations, the two most recent inspections per license within about a year), so a dropped row there can be the window closing, an edit, or a departure; only the edit is countable, and the rest are reported as rows without a successor, not as removals.

A pull is counted as a Re-publish when at least nine in ten of the rows the dataset held at the pull before it were dropped. The rows a Re-publish re-keyed — each dropped row paired with the arriving row of the same identity, or every row it moved where the layer declares none — are counted in the second table and not in the first; a row that arrived or left at a Re-publish without a partner is ordinary churn. A field is named as changed when it differs between partners in at least 50% of the pairs holding a value in it on either side, the city's OBJECTID aside, and a shift is named when every changed value moved by the same whole number of hours. A pair that also differs in a field the Re-publish did not name is counted as "also changed", and the answer calls a Re-publish mostly not a turnover while fewer than 50% of its pairs also changed. The differences include, among others: on the site layers, an inspection the previous pull had not held and an inspection more than a year old leaving the second slot; changes of format, such as the city's placeholder date for no inspection becoming an empty field, or a visit type relabeled; and, among repeated violations, which are paired in the city's row order, partners swapped by a reordering, so some differences are the pairing and not the record. The Re-publish of Aug 24, 2026, between the pulls of Aug 18, 2026 and Aug 24, 2026: The site layers began publishing LAT and LON in projected feet instead of degrees; the point geometry itself did not move. The same pull carried inspections the previous pull had not held, dated days to months before it — the layer publishes with a lag — so a site's newest inspection can be one dated before the previous pull and published only after it. The violations layers stopped filling VIOL_CODE and RANK (the violation code is still published, as TXT_VCODE), spelled out the establishment-type names in INV_TYPE, and began publishing INSP_TYPE values other than 1, the only one they had published before, relabeling some rows already published. The Re-publish of Oct 5, 2026, between the pulls of Sep 28, 2026 and Oct 5, 2026: Every Inspections date moved from midnight CST (UTC−6, year-round) to midnight UTC; the site formats dates in UTC, so no date it shows changed. Without a partner at the Re-publish of Aug 24, 2026: processing sites, 49 brought; processing violations, 167 dropped and 272 brought; service sites, 55 brought; service violations, 432 dropped and 721 brought. The Archive began on Aug 18, 2026 with no backfill, so nothing here reaches further back than that.