---
name: pre-register-the-statistic-not-just-the-prediction
description: "Pre-registering a prediction is not enough — the STATISTIC that decides it must be pre-registered too, because on even N the median convention alone can flip the verdict"
metadata: 
  node_type: memory
  type: feedback
  originSessionId: 8a9ccdf5-484e-4d48-986a-7784ee7b1a58
  modified: 2026-08-04T23:22:12.930Z
---

Pre-registering a prediction with its falsifier is only half the discipline. **The statistic that
evaluates it is a second free parameter, and choosing it after seeing the data lets you pick the
one that flatters the hypothesis** — without ever feeling like you cheated.

Measured 2026-08-04 (item-3 capture, N=20). On an adversarial split of ten `1`s and ten `3`s:

    lower-middle  sorted[9]  = 1  => prediction FALSIFIED
    upper-middle  sorted[10] = 3  => prediction SURVIVES
    averaged                 = 2  => prediction SURVIVES

Same twenty numbers, three verdicts, all three defensible in print. **N=20 is EVEN, so "the median"
is not a defined value** — every even-N median is a convention wearing a statistic's authority.

A first synthetic set (fifteen 1s, one 2, four 3s) had all three conventions agree at 1, which is
exactly the trap: **the agreement was a property of that distribution, not of the statistic**, and
one sample could not clear it. See [[one-sample-is-not-a-negative]].

**Why:** ⭐⭐ the same class as an extractor that can only see the shape it expects — an instrument
that can flatter its author's hypothesis has no verdict to give. Sibling failure caught in the
same lane: a plural-only regex over `field 4: N ppid matches` extracted 5 of 20 and reported min 2
against a truth of 1, dropping every singular `1 ppid match` — a parse bug that would have
CONFIRMED the prediction. Negative-control the extractor on synthetic data with a KNOWN truth
vector before it ever meets real data.

**How to apply:**
- Register the convention BEFORE the capture, and **choose the conservative arm** — the one that
  falsifies YOUR OWN prediction on a tie (lower-middle, here).
- **File the raw vector and a histogram as the deliverable**, with min/median/max as a summary of
  it and never a substitute. A reader who rejects your convention must be able to recompute any
  statistic without re-running the capture — that is what makes the convention non-load-bearing.
- Have the script PRINT the verdict against the pre-registered falsifier, so the ledger row cannot
  quietly acquire a different one later.
- Prefer even-N designs that report a full distribution over ones that report a single centre.

See [[precise-claim-is-falsifiable-vague-is-useless]], [[make-a-new-rig-red-on-purpose]],
[[verdict-from-probe-competence]], [[rig-inherits-the-defect-it-studies]].