Counterexample CE-D-UNMERGED-DUPLICATES-001

What it falsifies

One-point exchange stability implies strict self-consistent nearest-centroid assignment even when coincident positive-weight score atoms may be assigned to different cells.

This is the fixture behind D-UNMERGED-DUPLICATES-FAIL: a one-dimensional example with three singleton cells and two rows equidistant from their competing centres, which is vacuously exchange-stable while strict nearest-centre assignment and deterministic label reproduction both fail. The solver refuses a terminal state whose own Mahalanobis rule would relabel a row by more than the gain tolerance, citing this fixture; the same code appears if a compile attempt hits the same condition.

CE-D-UNMERGED-DUPLICATES-001: labels before, 3 atoms in 3 cells-1-0.50000.5001score 1
Labels before
  • Cell 0
  • Cell 1
  • Cell 2
Atoms of CE-D-UNMERGED-DUPLICATES-001
AtomWeightScore 1Cell before
11/410
21/411
31/2-12

Objective

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

LabellingObjective
Labels beforelog(1)=0

How it was verified

Method: exact_formula.

All three cells are singletons, so the labeling is vacuously one-point exchange stable under the nonempty-cell rule. Information is positive definite in d=1, but split duplicate atoms prevent strict assignment and deterministic score-only reproduction. Merging the two score-1 rows removes the degeneracy.

Claims that cite it

What the library refuses

Trigger

The terminal rule would relabel a training row by more than the gain tolerance

Reason

split duplicate atoms can be vacuously stable while defeating label reproduction

Remedy

Merge or consistently label coincident score rows; inspect geometry.maximum_violation_gain

Recorded 26 August 2026. The fixture as JSON.