Counterexample CE-DS-INTERVAL-SEED-UNSTABLE-001

What it falsifies

The exact efficient-score interval labeling - the DP optimum of the full-sample efficient score, the documented initializer for profiled exchange and the finite analogue of the DS15 degenerate attainer J* - is itself one-point exchange-stable for the in-bin profiled-Ds objective. FALSE: on this exactly centered N=8 sample from the centered06 grid law, relocating row 7 to cell 1 raises the profiled value by 0.447 exactly. The improving move works by growing the binned nuisance block 27-fold (0.0042 -> 0.111): near the interval seed the projection tax is the dominant term and the profiled objective climbs by buying back nuisance information, the finite mechanism behind DS15 Proposition 6's steering.

CE-DS-INTERVAL-SEED-UNSTABLE-001: labels before, 8 atoms in 3 cells-2-101-2-101score 1score 2
Labels before
CE-DS-INTERVAL-SEED-UNSTABLE-001: labels after / optimum, 8 atoms in 3 cells-2-101-2-101score 1score 2
Labels after / optimum
  • Cell 0
  • Cell 1
  • Cell 2
Atoms of CE-DS-INTERVAL-SEED-UNSTABLE-001
AtomWeightScore 1Score 2Cell beforeCell after
11/8-1613/40961517/2048000
21/84875/4096-3003/2048011
31/82563/409630109/2048000
41/8-9421/4096-44131/2048000
51/8-1657/4096-15751/2048022
61/81571/4096-5379/2048022
71/8-1949/409616717/2048000
81/85631/409619921/2048021

Objective

The criterion value at each labelling, exactly as the fixture certifies it.

LabellingObjective
Labels before4644025/112623616
Labels after724740241/1483132192

Consequences

  • the documented profiled initializer is generally not a terminal state: seeding inside the DS15 degeneracy does not mean returning it, so the optimizer's output is a genuinely different object from the projected rule (study blocker 'selection effects')

  • the finite projection tax at the interval labeling is Theta(1) and one-point moves reduce it, the mechanism DS15's achievability construction exploits at scale

How it was verified

Method: exact_formula.

exact rational arithmetic; gain verified by from-scratch recomputation of both binned information matrices; regression test tests/test_research_claims.py::test_ds16_efficient_score_interval_seed_is_not_exchange_stable

Claims that cite it

Recorded 30 August 2026. The fixture as JSON.