ActionBasis synthetic measurement fixture

Unzip all files into one directory. Node.js 18+ is sufficient; no installation,
network request or external dataset is required. Run:

  node replay.mjs

receipt.json contains the exported period totals and native estimator readouts
for the existing 80-seed measurement test: 40 synthetic units per arm, 13 pre
weeks and 13 post weeks. Positive +10%, negative -15% and zero shifts are
known by construction. checks.mjs independently replays the point estimate and
95% normal-approximation interval from the per-unit totals. It also checks the
reported suite counts. replay.mjs demonstrates refusal when controls are absent.

Source hashes identify the fixture and native estimator used for this export.
They establish custody, not scientific adequacy. The replay does not regenerate
the original noise, weekly panel, or pre-trend checks; it recomputes the published
point estimates and intervals from the exported period totals. This is not a holdout on real
company data, a customer result, a revenue forecast, a calibrated demand model,
or a test of the complete company simulator/decision ranking. The fixture's
independent noise and common design do not cover real-world confounding,
selection, spillovers, missing records, seasonal misspecification or all forms
of dependence. The intervals miss truth in two negative runs; one null run is
statistically significant. Those failures are retained, not removed.

Money below zero/above zero is not a rollout decision: operating costs, controls,
pretrends, field design and an actual business readout still need qualification.
No costs, ROI, payback, customer identities or live grants are supplied here.
