Crossfield website example evidence — October 24, 2026

Environment
  Linux ARM64 EC2, invoked from the maintained Crossfield ./dev shell workflow.
  DuckDB source base d54cbbe3df17e8e5ce68312f483b4155836112aa plus eight Crossfield patches.
  Matching current native Crossfield extension. No engine build or test ran locally.
  Website branch: feat/crossfield-content-strategy.

Fresh checks
  Each downloaded ZIP was transferred and extracted into its own directory.
  The actual setup.sql and run.sql were executed with the native DuckDB CLI,
  in fresh connections with the optimizer enabled and disabled. Eight successful sessions.
  Each CLI process had a 45-second deadline and five-second kill grace.

  Standard native DuckDB SQLLogicTests independently assert the published values:
    retail:       38 assertions per optimizer mode
    ossie:         8 assertions per optimizer mode
    measurements: 19 assertions per optimizer mode
    pipelines:    11 assertions per optimizer mode
  Total: 8 successful cases, 152 assertions. No skipped cases counted.
  Each native test process had a 60-second deadline and five-second kill grace.
  Website regression files: tests/crossfield/native/*.test.

  Included rejection checks: strict raw-source aggregation, interval mean without
  an approved reduction, missing callable signature. Empty pipeline: 0 count / NULL sum.
  Successful discovery and approved results are separate from the deliberately
  failing raw-source query. The captured JSON error is in agent-session.txt.

  All four explain.sql files ran successfully through EXPLAIN (SEMANTIC).
  The pipeline signature-discovery query also ran successfully.

Expected answers
  Retail: Brooklyn 19 / 36 / 255.00; Philadelphia 506 / 798 / 6100.00;
    totals 525 / 834 / 6355.00; definition revisions 6730.75 / 6355.00 / -375.75.
  Ossie: logical-field sales 84 and refunds 7.
  Measurements: A Jan 1 rainfall 5, mean 10; A Jan 2 rainfall 7, weighted mean 40;
    B Jan 1 rainfall 11, mean 30. The two durations are one and two hours.
  Pipelines: direct 5 / 10, named direct 5 / 10, model-owned pipeline 3 / 6.

Recorded evidence, not rerun by this website task
  The homepage's detailed expanded SQL/JSON capture comes from the earlier native
  Jaffle walkthrough on the matching engine. Internal generated names are not identities.
  GitHub typed-scope values are from the independently authored integration fixture:
    tests/crossfield/overloads.test in vgi-github. The current website work does not
    rerun GitHub HTTP tests or migrate its entire generator to typed overloads.
  The full native Crossfield suite's earlier 396/396 pass is separate engine evidence,
    not a fresh website-task full-suite run.
  Current Open-Meteo actual-worker validation remains blocked by VGI 2.1.0 versus
    worker 2.2.0. Earlier worker passes do not qualify that current pair.

Boundaries
  Bundles contain source data/models, not executable binaries or public releases.
  A passed fixture is not numerical acceptance of a new business model or arbitrary
  provider compatibility. Native validator authoring/binding checks do not establish
  numerical correctness. No claim of production approval or live API values is made.