Run: 2026-05-05__009__buyside-hunt-routine-review · Date: 2026-05-05 · Phase 1 author
Phase 1 verdict: REWORK 58/100. Three real data-quality bugs in the Jonathan run output: (1) HQ-mismatch warning chip is the wrong policy. Nashville, Tempe, and Sacramento companies survived as score-94 entries because Places API force-resolved the buyer's city onto them. Should be a hardfail, not a chip. (2) Cross-vertical contamination: 9% of kept rows duplicate across adjacent verticals (All Pro Pressure Washing in window AND pressure; A Clear View in both). (3) No DNC or active-deal pre-screen. pipeline could surface a company already in Capstone or Design Precast pipelines or on a DNC list.
Did not benchmark the score distribution against a buyer's actual pick list. The 65 cutoff filtered zero candidates in the live run; only the targets_per_vertical=5 cap discriminated. Without ground truth (Jonathan reviewing and ranking the list), can't validate whether the rubric weighting actually predicts buyer interest.
Pattern observed across multiple runs: when a hardfail gate is silently demoted to a warning chip, the bad data still ships and the chip is ignored. Recommend codifying: "if a signal is strong enough to trigger a warning, it's strong enough to drop the row." No middle ground.
Pipeline runs but ships visible noise. Three additive fixes raise health from 58 to 85+ without architecture changes. All three landed in W7 same day.
Generated from 009__audit-data.md — do not edit this HTML directly.