A funding-rate signal that got stronger on more data, and still died

The original read left this as "not ready, re-evaluate in April". The re-evaluation trigger fired, was flagged overdue twice, and then nobody ran it for 116 days. When it finally ran, on 4.7 times the data, the headline effect came back stronger rather than weaker. It still died. Every construction that removes the overlap between measurement windows loses all significance, the on-default versus off-default split the original spec itself prescribed is flat, and on data collected after the original window the effect points the wrong way.

Re-run on 4,244 hourly observations (4.7x the original 904), 2026-02-16 to 2026-08-12. The pre-committed primary statistic (6-hour horizon, non-overlapping sample, funding above the venue default) came in at −81.55 bps across 28 observations, t=−1.33, p=0.1947, failing its own p<0.05 bar. What replicated: the overlapping quintile spread strengthened to +38.64 bps (t=4.03, p=0.0001) against the original +36.0 bps (t=1.67, p=0.096). Three separate killers. One, every de-correlated construction fails: 6h p=0.1947, 3h p=0.1384, 1h p=0.6486. Two, the binary on-default versus off-default split the original document prescribed is flat at every horizon, p at or above 0.30. Three, out of sample after 2026-03-27 the high-funding mean is +16.85 bps (t=0.26, p=0.80) against a predicted negative. The fragility was visible from the start: funding sat pinned at the venue default for 71.5% of hours (71.5% replicating the original 72.0%), so the quintiles were really a default-versus-not split in disguise. Caveat shipped with killer three: window sub-samples re-anchor the sampling stride, so they are not a clean partition of the pooled sample. Disclosed non-coverage: the three-way combination test was not re-run despite enough data now existing, because rescuing a signal after its primary statistic has failed is banned by our own rules.