Integrating with
your stack.

Three tables, one question

The working core of every sheet: claims assert a relation with a grade ceiling; evidence units bind claims to studies with extraction detail; studies carry design, model, duration, RoB. They join on stable hashed keys — claim → evidence → study — and every join resolves inside the release.

# pandas — claims with their evidence depth
import pandas as pd

claims   = pd.read_csv("claims.csv")
evidence = pd.read_csv("evidence.csv")
studies  = pd.read_csv("studies.csv")

# which graded claims carry the most bound evidence?
depth = (evidence.groupby("claim_id").size()
         .rename("bound_units").reset_index())
claims.merge(depth, on="claim_id") \
      .sort_values("bound_units", ascending=False) \
      [["claim_id", "grade", "direction", "bound_units"]].head()

-- SQL — benefit claims where opposition was bound
SELECT c.claim_id, c.relation, c.grade,
       COUNT(o.record_id) AS opposing_records
FROM claims c
JOIN oppositions o ON o.claim_id = c.claim_id
                  AND o.verdict = 'BOUND'
WHERE c.direction = 'BENEFIT'
GROUP BY c.claim_id, c.relation, c.grade
ORDER BY opposing_records DESC;

# Databricks — the turmeric share, zero download
# catalog: mfi_turmeric  (Delta Sharing recipient)
spark.sql("SELECT * FROM mfi_turmeric.claims WHERE grade='HIGH'").show()

Keys you can rely on

Record IDs are content-derived and stable across rebuilds — CLM-…, STU-… identifiers don't drift between releases, so your queries survive version deltas. The version_deltas ledger records exactly which rows changed between releases, so you can diff rather than re-scan.

The tables that matter most