The Global Mixed-Variance Transform Lands on the Unique Framed Connector
Entries 430–434 provide all pieces of the correctly typed left leg. They now assemble into a global mixed-variance integral transform in the fs/Kato sector.
An object on the normalization–conductor Milnor square is a diagram
Pull it to the three-chart V/Čech correspondence and form the conductor difference totalization
Tensor with the exceptional logarithmic relative-dualizing packet and the external Cartier packet, then apply normalized blowdown and Čech descent:
This formula is mixed-variance through the conductor restrictions; it does not pretend that is an ordinary ringed-space morphism.
For the distinguished normalization-sheet object , every component has already been independently fixed. The integrated audit reruns the universal conductor kernel, all 215 multi-Rees stalks, 522 localization squares, 840 path comparisons, global three-chart conductor descent, the generic logarithmic Thom comparison, and both mandatory forgetting ablations.
The resulting image has frozen signature
Both ordinary-forgetting images are zero.
Entry 421 proves that the admissible connector with this complete signature is unique: its local relative deformation group vanishes, and the global order-three lift torsor and automorphism group both vanish. Since the transform constructed above supplies an object with that signature,
up to the unique admissible homotopy.
This closes the normalization-sheet kernel that Entries 144 and 146 left open, within the fs/Kato logarithmic category precisely delimited by Entries 425–429. It does not resurrect the impossible ordinary span to separated sheets, nor claim a global raw-scheme correspondence outside the Kato sector.
The next research fork is now genuinely downstream: form the physical derived pullback/butterfly using the fixed road inclusion and evaluate its primitive class, or test the constructed transform under the first eight-point Cut naturality square.
The executable audit is
research/voevodsky/check_global_mixed_variance_transform.py.