---
name: preregister-statistic-convention-not-just-prediction
description: "Even-N median: the convention alone (lower/upper/averaged middle) flips a pre-registered verdict on identical data — pre-register the statistic's convention and deliver the raw vector"
metadata: 
  node_type: memory
  type: project
  originSessionId: 6412c22c-6c07-47a0-9145-4b336e1e8135
  modified: 2026-08-04T22:41:54.695Z
---

Measured by hertz 2026-08-04 (item-3 extractor negative control, synthetic data): on an
adversarial 20-sample set (ten 1s + ten 3s), the SAME data yields median 1 (lower-middle,
prediction falsified), 3 (upper-middle, survives), or 2 (averaged, survives). A convention chosen
AFTER seeing the capture lets the measurer pick the one that flatters the hypothesis — likely
without noticing. On his first synthetic set all three conventions agreed, which is why one sample
cannot clear the hazard: the agreement was a property of that distribution, not of the statistic.

**Why:** pre-registering the PREDICTION is not enough when the verdict statistic has degrees of
freedom (tie-breaks, rounding, binning, outlier rules). Each unpinned degree of freedom is a
post-hoc knob.

**How to apply:** before capturing, pin (1) the falsifier AND (2) every convention the verdict
statistic needs — choosing the arm that falsifies your own prediction on ties (hertz picked
lower-middle for exactly this reason). Deliver the RAW VECTOR in capture order plus a histogram,
with summary stats derived from it, so any reader can recompute under their own convention without
re-running. Kin: [[instrument-soundness-guards]] (green must not be producible by the fault),
hertz's dual-shape regex rule (an instrument must not be able to flatter its author's hypothesis —
ask of every capture: which direction does its failure skew?).