Iteration 14 — First FRAGILE↔CONDITIONAL harvest 5 HARVESTED · 72 TOTAL MRP LEAD RESOLVED · ARXIV OUTAGE RECORDED

In plain words: first dedicated haul for the boundary that separates fragile features (work in a couple of regimes, luck) from conditional ones (reliably work in named regimes). Five instruments entered. Two measure how consistently a feature keeps getting selected — including the correlation-aware adjustment that fixes the exact bias that helped doom an earlier boneyarded method (unadjusted measures punish a model for swapping between two near-identical features, which is not real instability). One compares feature rankings across walk-forward steps with extra weight at the top, answering the house incumbent's design bar explicitly: it complements the exact flip-counter, it doesn't clone it. One reads WHERE in the distribution dependence lives — center versus tails — naming a failure mode the global measures are blind to. And yesterday's web-search lead paid off: Minimum Regime Performance resolved to a real 2025 Journal of Portfolio Management paper via the citation registry, entering as the worst-regime aggregation angle. Also on the record: the arXiv channel was down the whole firing (503s), so three candidates wait rather than enter unverified — the no-hearsay rule doesn't take outages as excuses. The branch also absorbed a new main-side audit (feature-realness) via the now-routine rebase.