WITHHELD PREDICTION HASHES
==========================

On 2026-08-22, eighteen predictions were registered in v1. Sixteen were published;
two were withheld. The SHA-256 of each withheld prediction's exact text is below, so
the count verifies without disclosing the content.

  D2  sha256 bddfe9698eda4d3674a7267aba4da6ce1657b23f04bdcb9602a3383f365822fa
  E4  sha256 ea4e874b6fe95a8fb8d55886b2a478f5058a32a701d5ad1b0a231c19669d1132

WHY EACH WAS WITHHELD

  D2 - predicts whether a named individual candidate actively contests his race.
       That is commentary about a person, not a test of a method, and this project
       does not publish commentary about candidates. It will be scored privately and
       its outcome published in October beside this hash.

  E4 - was conditional on obtaining access to a government file not under our control,
       so it could not be fairly scored either way. In v2 it is retired rather than
       carried forward, because a prediction that cannot be scored is not a prediction.

HOW THE COUNTS RECONCILE

  v1 (2026-08-22)   18 registered = 16 published + 2 withheld
  v2 (2026-08-22)   19 scored predictions + 1 commitment (F2)

  v2 is not v1 plus two. It is a rebuild after four independent reviews found a
  logical paradox in F2 and four falsifiers looser than the claims they scored:

     A1 split into A1a and A1b so the incumbents-only clause can cause a miss   +1
     F2 became a commitment; F2' added as the scored calibration claim           0
     G1 and G2 added on military service and legal record                       +2
     D2 remains withheld; E4 retired as unscoreable                             -2
     ------------------------------------------------------------------------------
     18 + 1 + 0 + 2 - 2 = 19

  v3 (2026-08-23)   25 scored predictions + 1 commitment (F2)

  v3 is a correction, not an expansion. On 2026-08-23 the A-block headline figure
  (pooled Pearson r = 0.82, incumbents-only r = 0.70) was found not to reproduce from
  the frozen artifacts; the reproduced values are 0.576 and 0.311. Four predictions had
  been priced against the wrong number. The FACT was corrected; the BETS were not.

     v2 scored set carried forward unchanged, defects and all                   19
     A1a, A1b, A3, A4 left frozen and scored as misses                     (of the 19)
     A1a', A1b', A5, A3', A4' registered against the reproduced baseline         +5
     F2'' registered as the honest reprice of F2' (60% -> 93%); both scored       +1
     ------------------------------------------------------------------------------
     19 + 5 + 1 = 25          mean stated confidence 73.80%

  D2 remains withheld and remains scored privately against the hash above. E4 stays
  retired. Neither is affected by the A-block defect.

VERIFYING

  sha256sum -c MANIFEST.sha256

  Every file in MANIFEST.sha256 is published unmodified. Any file whose digest does
  not match has been altered - which is the point of publishing the digests.