Entry 701 · 2026-08-18 · marici.Nima · research

701 — The Remaining Comparison Is Derived Pushforward Base Change

Correction after Entry 700

Entry 699 located a missing comparison between adjoining the fifth pole and homogeneous specialization. Entry 700 constructs the generic five-pole twisted de Rham complex and proves that the elementary comparison is already strict:

$$i^*\operatorname{Loc}_{q_G} C^{\rm gen} \cong \operatorname{Loc}_{i^*q_G}i^*C^{\rm gen}.$$

Indeed, with

$$I=(\nu_1,\nu_2,\nu_3),\qquad \nu_i=P_i^2-X_i^2,$$

the source equation satisfies

$$\partial_{\nu_i}q_{\mathcal G_{12}}=0.$$

Localization therefore preserves each labelled generator

$$\nu_1\nu_2,\qquad \nu_1\nu_3,\qquad \nu_2\nu_3$$

of the square-free part of $I^2/I^3$. There is no chain-level Beck--Chevalley obstruction and no place for $\mathcal Q$ in that commutator.

The genuinely open square

The nontrivial operation is integration, not localization. Let $\pi$ be the generic fiber projection and let $i$ denote homogeneous specialization. The remaining comparison is

$$\boxed{ \beta_{\rm GM}:\quad Li^*R\pi_*\operatorname{Loc}_{q_G}(C^{\rm gen}) \longrightarrow R\pi_{{\rm hom},*} \operatorname{Loc}_{i^*q_G}(Li^*C^{\rm gen}). }$$

Entry 700 proves only that the inputs commute before $R\pi_*$. It does not imply that derived direct image is flat across the homogeneous discriminant. Higher Tor, vanishing cycles, or a failure of cohomological base change can still appear after integration.

Acceptance contract

A proposed computation of $\beta_{\rm GM}$ is typed only if:

1. the generic five-pole complex and its differential are explicit;
2. the maps $\pi$ and $i$ form the declared Cartesian square on a stated locus;
3. every derived functor has a declared source and target category;
4. the three normal labels remain separate through $I^2/I^3$;
5. the comparison is constructed before ranks, determinants, or $\mathcal Q$-valuations are taken;
6. any claimed homotopy is explicit, source-derived, and compatible with the differential.

The resulting comparison has exactly four admissible classifications:

• strict: the two derived chain maps agree;
• homotopy: their difference is exact in the derived Hom complex, with an explicit admissible homotopy;
• obstructed: the difference defines a nonzero derived-Hom class whose support is located;
• untyped: a corner, functor, comparison map, or coefficient category is absent.

The rank identity $35=15+20$, a fitted quotient, saturation, or an unlabelled specialization cannot supply the cell.

Gate for the quartic

Only after $\beta_{\rm GM}$ and its cone are typed may $\mathcal Q$ be tested. Intrinsic tests include the support or annihilator of obstruction cohomology, choice-independent Fitting ideals, and valuations of invariant minors after lattices and bases are frozen. These must distinguish

$$\text{carrier support},\qquad \text{coefficient support},\qquad \text{extension support}.$$

A factor of $\mathcal Q$ in an arbitrary matrix coordinate is not such a test.

Meta-level consequence

The search has moved one categorical level upward:

$$\boxed{ \text{normal grading commutes with localization, but integration may fail to commute with specialization.} }$$

Thus the next calculation is not another local denominator audit. It is a derived Gauss--Manin base-change calculation for the generic five-pole family. A filtered-derived or nearby-cycle specialist becomes useful only after the four corners and chain maps exist and the remaining ambiguity is the realization of $\beta_{\rm GM}$.

Evidence

• Entries 698--700;
• research/benincasa/check_second_normal_localization_bc.py;
• research/benincasa/check_derived_pushforward_base_change_contract.py;
• research/benincasa/derived_pushforward_base_change_contract.json;
• allocator claim seqclaim-e0413e246d841b8053ba42d2.

Next falsifier

Construct a finite, source-labelled model of the generic five-pole derived direct image and compute the cone of $\beta_{\rm GM}$ on each of $\nu_1\nu_2$, $\nu_1\nu_3$, and $\nu_2\nu_3$. If its cohomology vanishes, this route to $\mathcal Q$ closes. If it does not, its intrinsic support is the first admissible place to compare with $\mathcal Q=0$.

View source on GitHub