Hypothesis-testing

The descriptive first-pass battery (E0–E10, no modelling, data in hand) worked backwards from the management message — coverage gate, realization waterfall, realized-bps distribution, fee-code discount ladder, never-asked share, practical-maximum frontier, raw driver slices, within-MG gap, waived-pool concentration, and the return-on-AUM × custody-rate pricing quadrant. Each carries question, data, chart, grain, caveat, and report home; doubles as a build tracker. The base table is anchored on the near-complete account universe with the full-waiver pool flagged; eligible AUM is the authoritative backend figure, not re-derived.

The transformation layer: build_base_table(...) turns the raw ingested frames into the joined, account-master-anchored base table — one row per account in the near-complete universe, the charging outcome left-joined, the full-waiver pool flagged (accounts absent from the charging file), management type classified, and join keys normalised. The reusable substrate every downstream analysis builds on; behaviour pinned by a synthetic-fixture test suite.

Runnable Python (matplotlib), analysis only: consumes the base table from the transform layer and renders the coverage panel, realization waterfall, realized-bps distribution, fee-code ladder, charged-at-all share, raw driver small-multiples, within-MG candidate pool, waived-pool Lorenz, and the return-on-AUM × custody quadrant. Descriptive; figures are generated corp-side and not committed.