Raw DNC Comparison Is Localization Not a Physical Realization
Entry 542 leaves comparison with raw algebraic DNC geometry as the apparent foundational frontier. Entries 425--429 already determine the nature of that comparison locally. Propagating their invariant through the even-arity induction shows that a raw-scheme equivalence preserving the physical class is impossible.
At every physical radial chart, the logarithmically saturated PC stalk is
The raw DNC relation (u=Xt) implies that inverting (u) also inverts (X) and (t). Hence the raw radial stalk is
and the canonical comparison is localization at (X).
The retained boundary module
is nonzero, but
This module is not incidental: it is the occurrence/Cartier support carrying the closed residue of the framed connector. Therefore the raw comparison is flat but nonconservative and cannot be a derived equivalence. A failure on one chart cannot be repaired by global descent.
Every even-arity physical Cut chart contains this same radial local model. The checker verifies the witness on all physical charts through (n=14), with counts
The general count at (n=2m) is (m(m-2)), so the obstruction propagates to every nontrivial even arity.
Consequently
The correct algebraic home is the fs/Kato logarithmic model of Entry 426 or the trivial-inertia sector of its Artin-cone presentation from Entries 427--428. The raw DNC remains available only as the generic nonconservative localization obtained after forgetting occurrence support. Its annihilation of the physical class is the required ordinary-forgetting behavior, not a defect awaiting repair.
Thus the phrase "raw algebraic realization" must be narrowed. What can still be sought is a global logarithmic algebraic stack realizing the Kato diagram and its mixed-variance transform. What cannot be sought is an equivalence to the ordinary raw DNC structure sheaf while retaining the connector.
The executable audit is
research/voevodsky/check_general_raw_dnc_comparison_no_go.py.