data-quality-gatelisted

# Data quality gate A gate that warns is not a gate. **This one blocks.** The economics are one-sided: catching a bad extract at the boundary costs minutes; catching it after it has flowed into a reconciliation, a metric, a board deck, and an external commitment costs days and a credibility hit. Every check below runs *before* anything downstream reads the data. Load `tie-out` — the gate emits the same PASS/FAIL discipline, with thresholds declared in advance. --- ## The seven checks Run in this order. Order matters: a failed row count makes every later check meaningless, so stop at the first `BLOCK`. ### 1. Row count within expected bounds ``` expected: 14,000 - 16,000 (prior 3 periods: 14,882 / 15,104 / 14,655) actual: 14,882 PASS ``` Derive bounds from history, not from a guess. **A row count of zero is always a BLOCK**, never an empty period — an empty result is far more often a broken query, a permissions failure, or a wrong date filter than a genuinely quiet month. ### 2. Control total agrees to source ``` source (NetSuite AR aging report): 12,441,520.11 extract sum(amount_usd): 12,441,520.11 PASS (exact) ``` This is the check that matters most, because it catches the failures the others cannot see — truncation, a fan-out join doubling amounts, a partial load. **Exact match required.** A control total tolerance is not a tolerance, it is a decision to ship unknown differences.