Guff AI ta ho niProject Turtle

M101 / model folder

How did the model produce four scores?

The frozen numerical output, its component diagnostics, and the precise boundary between reproducibility and proof admission.

Frozen output

The model shortlist, before legal proof admission

H4-UNVALIDATED score nodes
France-SpainModel massBranchStatus
0-19.20%Spain clean-sheet winH4-UNVALIDATED
1-110.14%Low drawH4-UNVALIDATED
1-215.63%Spain by one, BTTSH4-UNVALIDATED
2-28.22%High drawH4-UNVALIDATED

The four cells sum to 43.2%. This percentage is conditional model mass, not a claim that the final score must land in the set.

Comparable opposition

Conditional realm diagnostics

France realm n_eff 6.853; Spain realm n_eff 9.070. Both pass the isolated numerical run after widening, but neither becomes a legal proof realm until its rows are admitted.

Sensitivity

What 31/32 means

All four nodes survived together in 31 of 32 model-only perturbations. It measures internal shortlist retention; it does not test source validity, proof completeness, or post-kickoff calibration.