The aiccm2026dev-a position memo - re-derivation, adjudication, and honest concessions

Owned by the aiccm2026dev-a critical-theory line. Round 3 (2026-06-24). Companion to aiccm2026dev_a_theory.md (the -a paper), aiccm_a_vs_b_analysis.md (the round-2 erratum), and -b’s read-only aiccm_comparison/aiccm_comparison.tex. Discipline: derive every link and prove it; ground every number in a real run; concede exactly where -b is right; argue for the truth, not for -a.

This memo is written after the round-2 corrections, which already forced the -a side to retract two overclaims (Madelung-robust correlation; “ω^sym is the exact torus four-center”). It does not re-litigate those - it absorbs them, and goes looking for what is left that -a can actually prove, plus any residual error on either side. The headline of this round: a new, self-grounded result that sharpens the ω^sym concession from “not proven PSD” to “demonstrably not PSD” (negative eigenvalue −6.8×10⁻⁵), and a direct numerical settling of Q-kernel from -a’s own side against an external anchor.

All numbers below trace to a run in §6 (this session, 2026-06-24, vibeqc 0.14.1.dev0, this checkout). Nothing is estimated.

---

0.5 Round-4/5 convergence with the -b line (2026-06-24) - COMPLETE

A direct two-round theory exchange with the -b line converged fully: -b signed -a’s synthesis and the two lines agreed, point by point, on the framing below (with -b’s wording amendments adopted verbatim). This is the joint theory result - five established statements + one well-posed open construction, agreed by two independent derivations.

1. Production HF/Coulomb is one forced object (G=0-removed v_E, density-fit) - but “one object” only after the exchange-q=0 convention is declared. The J/Coulomb part is uniquely forced by zero-average G=0 removal; the exchange-q=0 part is an additional finite-size convention (strict-zero-mode vs BvK-Madelung exchange are distinct finite Hamiltonians, same TDL). The A==B==KRHF equality holds because both chose exxdiv="ewald", not by Poisson-uniqueness. Make it machine-recorded: declare exchange_q0 = BvK-ewald in the .system manifest and assert manifests match on that field for any A/B/KMP2 comparison (reproducibility gate, not a footnote). IMPLEMENTED (dev chat, 2026-06-25): periodic/exchange_convention.py (BVK_EWALD/STRICT_ZERO labels, [run].exchange_q0 manifest field, assert_matched_exchange_q0 gate; test_exchange_convention.py 13/13).

2. The bare→neutral difference is ξ_N·S⊗S + ⟨φφ|h_N|φφ⟩ - sign convention fixed: ξ_N = the signed Ewald self-potential, ξ_N=(v_E−1/r)|_{r→0}<0 (the deprecated −ξ·S⊗S form used ξ=−ξ_N>0; never mix them). h_N is the smooth finite-torus image field, h_N(0)=0. The remainder R=⟨φφ|h_N|φφ⟩ is physical finite-size physics, not gauge freedom and not RI error; h_N(0)=0 means the point-pair limit is pure rank-1, but a finite compact Gaussian pair still carries a second-moment/shape term ≈⟨φφ|r²|φφ⟩∇²h_N(0)/6, and diffuse/ionic pairs carry the full h_N. The neutral reference is required for ionic correlation. Clean diagnostic: R_X = g_eff^RI(X) − g_bare^direct − ξ_N·S⊗S over an auxiliary ladder X (with ξ_N from the lattice Ewald self-potential, not free-fitted) converges to R while the RI error → 0.

3. The exchange-q=0 seam propagates into correlation through the gap - it shifts the HF occupied/virtual spectrum, hence the MP2/CCSD/DLPNO denominators. So finite-N correlation carries the same convention dependence as HF; the same declared convention must govern HF, MP2/KMP2, CCSD(T), and DLPNO.

4. Label-weighting cannot protect kernel-positivity (PSD needs g = ∫∫ f_{pq}(r) K(r,r') f_{rs}(r') with K⪰0; a fold keyed on AO labels, not the pair density f_{pq}, is not of this form). ω^sym demonstrably fails (min eig −6.8×10⁻⁵); periodising first guarantees a positive K. (No universal “any weight cannot” - the statement is positivity is unprotected, with ω^sym the concrete counterexample.)

5. The two local-correlation constructions are unitarily identical at the exact limit (block-Fourier); the truncation/domain question is an implementation tradeoff (-b’s phase-stable canonical factor may give cleaner minimum-image domains).

The one open construction (item 6 - the next real joint deliverable): the explicit mixed-boundary (wire/slab) Green’s function and its matching exchange-q=0 seam for 1-D/2-D, derived Hamiltonian-first (define the Poisson problem, derive both Coulomb and exchange seam from the same G - not a minimum-image truncation). The starting forms (-b): G^{2D}(ρ,z)=(2π/A)Σ_{G∥≠0} e^{iG·ρ−|G||z|}/|G| − (2π/A)|z| + C (slab, Parry/Bertaut) and G^{1D}(ρ,z)=(2/L)Σ_{G_z≠0}K₀(|G_z|ρ)e^{iG_z z} − (2/L)ln(ρ/ρ₀) + C (wire). Validation anchor (-a): the 3-D torus kernel is the transverse-periodic sum of the mixed-boundary kernel, so stacking G^{2D}/G^{1D} with the right 3-D period must reduce to the already-validated v_E^{3D} (A==B==KRHF) - a hard consistency check with no new external dependency. To declare (physical BCs, not derived): the dipole/asymptotic boundary condition (polar slabs carry a dipole-layer term, the low-D analogue of the 3-D G=0 conditional convergence) and the matching exchange seam as the signed probe-charge limit of the same G. Present 1-D/2-D numbers are model numbers, not accuracy claims (Q-lowD, §2).

IMPLEMENTATION STATUS (Γ-CCM dev chat, 2026-06-25). The scalar-kernel level is DONE and gated: periodic/lowd_greens.py builds the wire G^{1D} (new to vibe-qc) and slab G^{2D} kernels; the dependency-free gates (test_lowd_greens.py 17/17) pin them by the harmonic property (∇²G=0 off-source) + the unit-source condition (G−1/r→ξ finite, r·G→1) + periodicity + transverse-reflection symmetry. My proposed validation anchor is CONFIRMED: test_lowd_greens_reduction.py shows each kernel reduces to a Madelung-validated 3-D Ewald under the large-transverse-period limit (wire residual <2×10⁻³ by D⊥=32, monotone). The low-D self-energy ξ (wire_self_energy/slab_self_energy) is the M4 exchange-seam foundation, anchored to the validated 3-D madelung_constant_for_cell (−madelung(D⊥)−ξ−(2/L)ln(D⊥) is D⊥-independent). Still DEFERRED: the AO-pair four-center (μν|G^{(d)}|ρσ) (M2b) → its reduction-gate vs ccm_eri_neutral (M3b) → the gauge-fixed exchange seam (M4) → the low-D 4c-vs-RI-gap-collapse controls (M5). M2b waits on the shared (scalability/slab-Ewald) four-center path to avoid a parallel dense build (maintainer decision 2026-06-25; HANDOVER_AICCM_FOLLOWON.md). So item 6’s kernel is built and the anchor holds; the four-center / energy layer is the remaining work, and 1-D/2-D energies stay model numbers until M5.

Naming (finalised with -b, 2026-06-25). The two lines differ by representation, not by Hamiltonian, under the family variational finite-BvK-torus CCM: aiccm2026dev-a = Γ-CCM, the direct-torus (translation-constrained Γ-supercell) representation; aiccm2026dev-b = χ-CCM, the finite-character (Γ-centred character-mesh) representation, with exact inverse transform to the real torus for local correlation. Unitarily equivalent at every cluster size - representation labels at a declared exchange-q=0 convention, not two theories or two Coulomb kernels. (-b’s correction, adopted: drop “no k-machinery” - Γ-CCM’s production path is multi-k GDF under the hood, exactly the A==B==KRHF result; “direct-torus” is the exact-theory presentation, not a claim of no k-work. And χ-CCM is finite-character - the finite group’s irreducible characters, not ordinary Bloch quadrature.) This memo keeps the -a/-b shorthand internally; the manuscript uses the descriptive names; 2014 stays AICCM.

---

1. The re-derivation, link by link - what -a proves vs asserts

The chain is: finite torus → neutral kernel → genuine four-center → WS partition → variational HF → Madelung background → local correlation. Each link is tagged [PROVEN], [PROVEN, caveated], or [ASSERTED].

L1 - The cyclic cluster is a finite BvK torus; CCM ≡ SCM-Γ is exact. [PROVEN]

The cluster is the quotient 𝕋 = ℝ³/L_c, L_c = {N₁a₁,N₂a₂,N₃a₃}; the BvK translation group G = L_c/L ≅ ℤ_{N₁}×ℤ_{N₂}×ℤ_{N₃} has order N = N₁N₂N₃. Its characters χ_m(n) = exp(2πi Σ_j m_j n_j/N_j) are an orthogonal basis (finite geometric series), so every block-circulant AO matrix X_{μ0,νR}=X_{μν}(R) satisfies the exact finite Fourier pair X(k_m)=Σ_R e^{ik·R}X(R), X(R)=N_c⁻¹Σ_m e^{−ik·R}X(k_m). Hence a Γ-supercell calculation constrained to commute with primitive translations decomposes exactly into the full unreduced Γ-centred mesh - for every finite N, not as a large-cluster limit.

This is the one place -a and -b are word-for-word identical. -b states it as Eqs. (orthogonality)/(forward)/(inverse) of its §2; -a as “CCM ≡ SCM-Γ”. There is nothing to adjudicate here; it is the shared foundation.

Caveat (steelmanned from -b §5): the identity needs the translation-invariant variational space [D,T_R]=0. A fully unconstrained Γ-supercell SCF has a larger determinant space and may break primitive-translation symmetry; such a broken-symmetry solution is not required to equal the mesh result. -a’s production drivers solve the supercell-Γ problem directly and so inherit whatever symmetry the SCF lands in; -a should state (as -b does) that the identity is to the symmetry-constrained minimum.

L2 - The torus interaction is the charge-neutral (Ewald) Green’s function v_E. [PROVEN for 3-D periodicity; ASSERTED/OPEN for low-D]

The bare 1/|r−r'| is neither periodic nor lattice-summable on 𝕋. The honest torus kernel is the zero-average solution of −∇²v_E = 4π(δ_𝕋 − 1/V_c), V_c = N·V_uc - the Ewald kernel with the compensating uniform background that makes it (i) L_c-periodic, (ii) symmetric v_E(r,r')=v_E(r−r'), (iii) charge-neutral (G+q=0 dropped). This is correct for full 3-D periodicity.

Ruthless caveat (this is where -a must not over-reach). The −1/V_c uniform background neutralises over the whole cluster box, including vacuum. For a genuinely 1-D-periodic (chain) or 2-D-periodic (slab) system embedded in a 3-D box, this is the 3-D-periodic-images kernel, not the isolated-wire / slab Green’s function. The correct low-D object is a mixed-boundary Green’s function (periodic along the active directions, open transverse). -a’s neutral v_E does not supply it; the theory paper’s phrase “no Madelung self-image leak, for any lattice” is true for the 3-D torus but should not be read as “correct for any reduced periodicity.” This is a shared open frontier with -b (Q-lowD, §4).

L3 - The production four-center g_eff = Σ_P L⊗L is a genuine ERI where ω^sym is not. [PROVEN - and this is -a’s strongest surviving result]

Define g_eff[μν,ρσ] = (μν|v_E|ρσ), realised as the RI density-fit of v_E: L_{P,μν} is the periodic-Γ GDF cderi on the cluster supercell (G+q=0 dropped), and g_eff = Σ_P L_{P,μν} L_{P,ρσ}. Four properties - the exact four that the round-2 erratum said ω^sym is “not proven” to have - hold for g_eff by construction, and I verified all four numerically (Run R2, §6.2):

property	why it holds for g_eff	run
RI-separable	g_eff = LᵀL by definition	‖g − LᵀL‖/‖g‖ = 7×10⁻¹⁷
PSD	tᵀ g t = Σ_P (Σ_{μν}L_{P,μν}t_{μν})² ≥ 0 (Gram)	min eig −6.6×10⁻¹⁶ (machine zero)
single symmetric two-point kernel	it is the density-fit of v_E	(by construction)
8-fold permutation symmetry	L_{P,μν}=L_{P,νμ} real at Γ ⇒ Gram symmetric	swap errors ≤5.5×10⁻¹⁷
basis-rotation covariant	functional of the pair density φ_μφ_ν, not AO labels	(inherited from v_E)

The mechanism is the whole point: g_eff inherits PSD / covariance / separability from the fact that it is the density-fit of an actual symmetric two-point kernel. ω^sym has no such parent kernel - its inter-electron fold keys on the AO centres, a basis-label, not the pair density - so it is guaranteed none of these. Run R2 confirms the prediction sharply: ω^sym is exactly 8-fold symmetric (swap errors 0.0) yet has a negative eigenvalue −6.8×10⁻⁵ - it is not a positive operator, hence not a genuine electron-repulsion tensor. 8-fold symmetry buys a scalar mean-field functional; it does not buy an ERI.

Concession, sharpened (see §5). This vindicates -b’s root objection: a symmetric weight on a non-periodised kernel is “the wrong kind of object.” -a’s own production path agrees - it uses g_eff, not ω^sym. The new content this round is that I can now say ω^sym is demonstrably not PSD, not merely “not proven” PSD.

L4 - The WS weights are a partition of unity over tied minimum images. [PROVEN - and identical in -a and -b]

On 𝕋 every relative translation is reduced mod L_c to its minimum image; when n images are equidistant (tied) the honest sum counts each, so a single representative carries weight 1/n with Σ_tied = 1. Applied independently per translation index (bra image, ket image, inter-electron image) it never couples a bra index to a ket index, so it preserves μ↔ν, ρ↔σ, and bra↔ket.

Identical to -b. -b’s §2 (“Where the weights enter”) writes exactly this: w_{AB}(d|[n]) = 1/m_{AB,[n]}, Σ = 1, an average over equivalent representatives of the same periodised integral that “leaves the value unchanged.” There is no numerical daylight between the two WS constructions (both 1/n over the true WS polyhedron, both per-index). The 2014 boundary share 1/2^counter (orthorhombic-only) is the special case both correct.

The eq-18 bridge ½(ω_{μρ}+ω_{νρ}) is the broken approximation both lines reject - -b by the explicit algebraic condition x_{mq}=x_{rl} that bra↔ket symmetry would require and that fails for a general quartet; -a structurally (the bridge couples the bra pair to the ket anchor ρ, leaving σ unmatched) and by measurement (~15 % bra-ket antisymmetry on H₄). Two routes, one conclusion.

L5 - HF is an ordinary variational functional on g_eff; exxdiv is settled. [PROVEN]

With a genuine symmetric g_eff, E[P] = Σ P h + ½ Σ PP(2g_eff − g_eff^exch) + V_nn is the textbook RHF functional; J,K are Hermitian with no symmetrisation patch, and F = ∂E/∂P is Hermitian. Contracted, E[P] is the periodic-Γ HF energy of the supercell. The exchange q=0 convention is exxdiv="ewald" (BvK-supercell Madelung), and on the production path it coincides bit-for-bit with the multi-k GDF and with PySCF KRHF (Run R1, §6.1). The historical 4.64 mHa gap was -a’s legacy exxdiv=None Γ-supercell helper, not an a/b disagreement (settled round 2; the legacy/B difference is 0.004641 Ha in -b’s h2_mp2_2026-06-21.json). -b’s statement that “using the primitive-cell Madelung constant here would define a different finite Hamiltonian and is incorrect” is correct and -a uses the same BvK-supercell convention.

L6 - The Madelung background is ξ·S⊗S to leading order, a fixed-reference diagnostic. [PROVEN at leading order; retraction absorbed; remainder OPEN]

g_eff = (μν|ρσ)_bare − ξ·S_{μν}S_{ρσ} + remainder, with q_{μν}=S_{μν} the AO-pair charge and ξ → the cell Madelung constant (fitted ratio 0.975→0.991 over L=12→20 bohr; the residual is RI error). Consequences, as corrected:

1. HF/KS: ξ·S⊗S ∝ S is a uniform Fock shift; absolute ionic energy comes from the neutral/GDF route. [PROVEN]

2. Correlation - fixed reference only: in the MO basis ξ·S⊗S → ξ·I⊗I has no occ-virt numerator element, so at a fixed reference it does not enter MP2/CCSD. But it shifts every denominator (the gap), so independent bare-vs-neutral SCF→correlation do not agree. Run R3 (§6.3): the fixed-ref cancellation is exact (10⁻¹⁶), the independent-ref shift is +17 % at ξ≈0.186 (the H₂ cell value) and +67 % at ξ=0.5. “Madelung-robust correlation” is therefore retracted; the neutral reference is required for ionic correlation, not just ionic HF.

3. Remainder - characterized as finite-size physics (round-4 update with -b). v_E − 1/r is not spatially constant on a finite torus. The clean decomposition (-b, 2026-06-24) is v_E(r) − 1/r = ξ_N + h_N(r), with ξ_N = (v_E−1/r)|_{r→0} the Madelung self-potential (the rank-1 term, exact for compact densities) and h_N the smooth, periodic, lattice-shape-dependent finite-torus image field, h_N(0)=0. So the bare→neutral difference is exactly ξ_N·S_μν S_ρσ + ⟨φ_μφ_ν|h_N|φ_ρφ_σ⟩: a compact-density rank-1 part plus the h_N matrix element R. R is determinate finite-size physics of v_E, not gauge freedom and not an A/B ambiguity when both lines use the same v_E; it is non-negligible for diffuse / well-separated / ionic-dipolar pairs and vanishes as N→∞ (algebraically for dipolar cells). It is distinct from the RI fitting error (orthogonal convergence: enriching the aux basis kills the RI error but leaves R at its cell-fixed limit). Quantifying R’s correlation footprint (LiH ≥2 cells vs multi-k KMP2; aux-refinement extrapolation) is the honest open run - deferred to vq (local-RAM constraint on the dense FFT builder).

L7 - The local-correlation supersystem collapses to canonical CCM in the no-truncation limit. [PROVEN]

Under CCM ≡ SCM-Γ, localizing the supercell-Γ occupieds equals the Wannier back-transform; PAOs/PNOs on the cluster reuse the molecular DLPNO machinery on the neutral cderi (L6: ionic correlation needs the neutral reference). -a’s subspace-projected union-PNO DLPNO-MP2/CCSD(T) reproduces canonical CCM to machine ε at no truncation (test_ccm_dlpno_{mp2,ccsd}.py: Δ ≈ 7×10⁻¹⁸ for CCSD, 8×10⁻¹⁵ PM/Boys). [PROVEN] The reduced-scaling local method (per-pair PNO, symmetry-unique pairs) is not yet production. [ASSERTED/open]

---

2. Verdict table - the six adjudication questions

For each: equivalent / both-valid-different / one-wrong, with the proof and the deciding run.

#	Question	Verdict	Proof / deciding evidence
Q-kernel	Is -a’s g_eff the same operator as -b’s production four-center?	Equivalent on the production path, at the declared exchange-q=0 convention	Run R1: -a run_ccm_rhf_gdf = -b RI-HF = −1.1182352381008094 Ha/cell, Δ = 0.0 (all 16 digits), and = PySCF KRHF to 5.9×10⁻¹⁰ on the externally-validated 20×20×6 (1,1,2) control. Both are the density-fit of v_E. Round-4 refinement (-b): the Coulomb/J part is forced by G=0 removal, but the exchange-q=0 part is a finite-size convention - the equality holds because both lines chose exxdiv="ewald", not by Poisson-uniqueness. Off-production, -a additionally carries the bare g and the diagnostic ω^sym, which -b does not build - but -a does not claim those as the production kernel.
Q-weights	Do -a’s WSC weights and -b’s weighting give the same energy in the large-cluster limit? Finite-cluster disagreement?	Equivalent (the WS partition is analytically identical); presentational difference only	Both are 1/n over tied minimum images, applied per-index (L4). No finite-cluster system makes the partitions disagree - they are the same function. The H₄ 1×1×1 88-mHa gap (−b table) is bare-vs-neutral kernel, not a weight disagreement: there union12 == ω^sym-a exactly. -a keeps a symmetric weight (ω^sym) as a diagnostic; -b drops weights at the root. Both defensible; identical on production.
Q-Madelung	Is the neutral reference sufficient for ionic correlation, or is there a residual -a missed?	-a was wrong (retracted); current positions both-valid; a residual is open on both sides	The original “Madelung-robust correlation” was an error (Run R3: +67 % independent-ref MP2 shift at ξ=0.5). Corrected, both lines agree the neutral reference is required. -a’s ξ·S⊗S characterization is a correct leading-order diagnostic (ξ→Madelung const); -b avoids it by construction. Round-4 (-b): the non-rank-1 remainder is now characterized as `R = ⟨φφ
Q-lowD	Does -a’s / -b’s neutral kernel handle reduced periodicity?	Both incomplete; neither validated; not a fixable bug - a missing kernel	-a’s v_E is the 3-D-Ewald kernel (uniform −1/V_c over the whole box) - correct for 3-D periodicity, wrong compensating background for genuine 1-D/2-D (needs a wire/slab Green’s function). -b is equally open: its 1-D/2-D four-center (direct-truncated) and RI (fitted metric) gauges are unmatched (decision O1; differences of hundreds of kJ/mol), and four-center-vs-RI on vacuum-padded H₄ is 22.2 mHa / 58.3 kJ/mol (O9). Neither line has the mixed-boundary kernel. -a must not claim “any lattice” for reduced periodicity.
Q-localcorr	Are -a’s subspace-projected DLPNO and -b’s finite-torus DLPNO equivalent in the no-truncation limit? Different under truncation?	Equivalent at no truncation; both-valid-different under truncation; neither reduced-scaling-proven	Both reproduce canonical finite-torus CCSD(T) to machine ε (−a 7×10⁻¹⁸; -b lih_dlpno_ccsdt 6.9×10⁻¹² CCSD, 6.9×10⁻¹⁸ MP2). Under truncation -a uses union-PNO (shrinks the global virtual space), -b keeps all pairs/couplings and truncates only PNOs (an O(N⁶) oracle). Both conservative; neither has demonstrated reduced scaling (both still evaluate all pairs). No rigor gap between them.
Q-convergence	Same limit and rate as -b and the periodic reference?	Equivalent	Both define the same Γ-centred N₁×N₂×N₃ mesh sequence; -a’s real-space radius R_c with Δk = π/R_c is a parameterization of the same family (an isotropic R_c reproduces the anisotropic nrep modulo dim). Both note the sequence is non-monotone (each N is a different finite Hamiltonian) and both → the BZ integral / periodic GDF in the TDL. test_ccm_interaction_range.py (77/77) pins the duality vs a brute-force oracle to ≤10⁻⁹.

Net. The predicted honest finding holds: substantial convergence at the production level (both ≡ periodic GDF / KRHF), with genuine differences only in (i) the diagnostic constructions -a keeps (ω^sym, the bare g), (ii) the local-correlation route (real-inverse-transform for -b vs supercell-Γ-direct for -a - equivalent at no truncation), and (iii) the derivation route (real-Γ-native vs reciprocal/character-native). The 2014 AICCM is the common ancestor; its eq-18 four-center is the one object both lines prove (by different arguments) to lack the symmetry a variational energy needs. No residual production error was found on either side. The one real open item shared by both is the non-rank-1 Madelung remainder and the low-D mixed-boundary gauge.

---

3. Where the disputes are genuinely settled by computation (not argument)

Per the discipline “settle numerically rather than argue where you can”:

• Q-kernel is closed. Run R1 is from -a’s own driver and lands on -b’s number and the external PySCF anchor. This is no longer a theoretical dispute.

• The ω^sym concession is closed and sharpened. Run R2 shows ω^sym is not PSD by a margin (−6.8×10⁻⁵) far above machine noise, while g_eff is PSD to machine zero. -a’s production hierarchy is correct to exclude ω^sym.

• The Madelung retraction is closed. Run R3 measures the +67 % independent- reference shift directly; the fixed-reference cancellation is exact. The corrected wording is now propagated into the -a-owned code/docstrings this session (neutral.py, __init__.py, test_ccm_neutral.py).

---

4. Concessions - exactly where -b is right

A good -a advocate concedes precisely. -b is right on:

1. Drop the four-centre weight at the root. -b’s “any weight on a non-periodised kernel, even a symmetric one, is the wrong kind of object” is correct, and Run R2 proves it (ω^sym not PSD). -a’s only defensible use of ω^sym is as a historical diagnostic showing why symmetry repairs eq-18

   • never as a production ERI. -a’s paper now says exactly this.

2. Madelung is fixed-reference, not robust. -b caught the overclaim; it was a real error; it is retracted. The neutral reference is mandatory for ionic correlation.

3. Fail-closed low-D honesty. -b’s refusal to call its 1-D/2-D four-center-vs- RI gaps “fitting errors” until a common wire/slab Green’s function exists is the correct posture. -a should adopt the same caution and not imply its neutral kernel is validated below 3-D.

4. The strongest single external check is -b’s. -b’s 3-D RI-MP2 vs PySCF KRHF/KMP2 at sub-nanohartree per cell (executed, archived) is the most convincing number in either paper. -a’s Run R1 confirms the HF half of it from -a’s side, but -b ran the correlation half first.

5. deMon2k is not eq-18. -b’s separate audit correctly notes the 2008 three-centre (Janetzko-Köster-Salahub Eq. 20) already symmetrises both AO orderings, so the eq-18 bra-ket counterexample cannot be transferred to it. -a’s paper audits eq-18 only and should cite -b’s deMon2k distinction rather than tar deMon2k with the eq-18 brush.

---

5. The strongest honest case for -a

What -a can defend without overclaiming:

1. A genuine ERI, proven, where the historical object failed. g_eff = LᵀL is PSD, basis-covariant, derived from one symmetric two-point kernel, and RI-separable - all four verified (Run R2). This is the property a single variational HF/MP2/CCSD functional requires, and 2014 eq-18 lacks it. -a and -b share this object; -a’s added value is the explicit hierarchy (bare → symmetric-weighted ω^sym → neutral g_eff) that makes the precise nature of the 2014 error legible and gives a cheap small-cluster/1-D tool.

2. The Madelung gap, characterized. -a is the line that wrote down g_eff = bare − ξ·S⊗S with ξ → the cell Madelung constant and proved the fixed-reference cancellation analytically and numerically. Even though the robustness claim was wrong, the characterization is a correct, useful diagnostic that explains the 2014 ionic over-binding and tells you why the neutral reference is required. -b never builds the bare tensor, so it cannot make this statement at all.

3. Breadth, validated against molecular limits. Open-shell (U-HF/KS, UMP2, UCCSD(T)), the full 4c→RI→RIJCOSX hierarchy as shipped methods, an all-Bravais 28-system test set + CRYSTAL23 harness, and the real-space interaction-radius parameterization (Δk=π/R_c) - a clean narrative that the CCM convergence knob is a real-space cutoff dual to the k-mesh.

4. A production HF path tied to a checkable equivalence. E_CCM == periodic-Γ HF == KRHF is not asserted - Run R1 is that equivalence, confirmed to all digits against -b and to sub-nanohartree against PySCF.

What -a must stop claiming: Madelung-robust correlation (done); ω^sym as the exact/production four-center (done); any low-D validity of the 3-D neutral kernel (this memo flags it).

---

6. Numerical appendix - the runs behind every number (this session, 2026-06-24)

Env: vibeqc 0.14.1.dev0, .venv (Python 3.14.6), checkout <vibe-qc-checkout>. STO-3G throughout.

6.1 Run R1 - Q-kernel: g_eff (production) == -b == PySCF KRHF

System: PeriodicSystem(3, diag([20,20,6]) bohr, H@(10,10,2.3), H@(10,10,3.7)), mesh (1,1,2) - -b’s externally-validated MP2 control.

A run_ccm_rhf_gdf  per_cell = -1.1182352381008094 Ha
B RI-HF (json)     per_cell = -1.1182352381008094 Ha   ->  A - B     =  0.0
PySCF KRHF (json)  per_cell = -1.118235238694408  Ha   ->  A - PySCF =  5.9e-10

6.2 Run R2 - g_eff is a genuine ERI; ω^sym is 8-fold symmetric but NOT PSD

System: diag([6,30,30]) bohr H₂ chain, nrep (3,1,1), n_AO=6, naux=108.

g_eff = LᵀL ?            ||g - LᵀL||/||g|| = 7.2e-17
g_eff PSD?               min eig = -6.6e-16   (PSD)
g_eff 8-fold sym:        mn 5.6e-17 | rs 5.6e-17 | braket 0.0
ω^sym PSD?               min eig = -6.83e-5   (NOT PSD)        <-- the concession, sharpened
ω^sym 8-fold sym:        mn 0.0 | rs 0.0 | braket 0.0

6.3 Run R3 - Madelung: fixed-ref cancels exactly; independent-ref shifts by tens of %

System: diag([12,12,12]) bohr H₂, (1,1,1). g_ξ = eri − ξ·S⊗S. mp2_full = −0.0131578701.

  xi     MP2(fixed ref)   MP2(indep ref)   fixed/full-1   indep/full-1
0.000    -0.0131578701    -0.0131578701      0.00e+00       0.00e+00
0.186    -0.0131578701    -0.0154613347     -1.11e-16      +1.75e-01
0.500    -0.0131578701    -0.0219477059      0.00e+00      +6.68e-01
2.000    -0.0131578701    -0.0001093606     -7.77e-16      -9.92e-01

Fixed reference: exact cancellation (≤10⁻¹⁶) - the true diagnostic. Independent reference: +67 % at ξ=0.5 - the retraction, measured. (Consistent in magnitude with the ~40 % cited in round 2 on a different synthetic setup; both are “tens of percent” and confirm the neutral reference is required.)

Cited existing gates (real runs on disk, not re-run here)

• -a DLPNO no-trunc == canonical CCSD(T): test_ccm_dlpno_ccsd.py Δ≈7×10⁻¹⁸.

• -b finite-torus DLPNO-CCSD(T): data/lih_dlpno_ccsdt_2026-06-21.json 6.9×10⁻¹².

• -b 3-D RI-MP2 vs PySCF KMP2: data/h2_mp2_2026-06-21.json < 1 nHa/cell.

• H₄ alt. chain bare(-a)/neutral(-b): data/h4_ri_2026-06-21.json (1×1×1 +231.6, 2×1×1 −7.68, 4×1×1 −3.30 kJ/mol) - converges with N.

• -a interaction-radius duality vs brute-force oracle: test_ccm_interaction_range.py 77/77.

---

7. Status / handoff

• Re-derivation complete (L1-L7); Q1-Q6 adjudicated with verdicts + proofs.

• Three round-3 runs ground Q-kernel, the ω^sym concession, and the Madelung retraction from -a’s own side.

• Stale “rigorously Madelung-robust” wording fixed this session in neutral.py, __init__.py, test_ccm_neutral.py (behaviour-neutral; docs/aiccm2026dev_a.md was already corrected).

• Proposed comparison text for the cowork chat: §”Proposed paper text” below / the dated handover note. The shared aiccm_comparison.* is not edited here.

• Next debate round: challenge set in handovers/HANDOVER_AICCM_A_POSITION_2026_06_24.md.

7.1 Execution of the converged result (follow-on chat, 2026-06-25)

The three deliverables of §0.5 are being executed (status tracked on the HANDOVER_AICCM_FOLLOWON.md board):

• D1 - exchange-q=0 convention (item 1): DONE. exchange_q0 is now a machine-recorded .system field ([run].exchange_q0, production "BvK-ewald"), with assert_matched_exchange_q0(...) raising on an A/B/KMP2 convention mismatch. python/vibeqc/periodic/exchange_convention.py, tests/test_exchange_convention.py (13). The reproducibility gate is a field, not a footnote.

• D2 - Madelung remainder R=⟨φφ|h_N|φφ⟩ (item 2): pending (vq). LiH ≥2-cell g_eff MP2 vs out-of-process multi-k KMP2, and the auxiliary-ladder residual R_X = g_eff^RI(X) − g_bare^direct − ξ_N·S⊗S (signed ξ_N from the lattice Ewald self-potential). To be folded into §6 when the vq numbers land.

• D3 - mixed-boundary low-D Green’s function (item 6): M1 DONE. The Hamiltonian-first derivation (Poisson problem, G^{2D}/G^{1D}, the dipole-layer BC, the transverse-periodic reduction anchor, and the exchange seam read off the same G) is in docs/aiccm2026dev_a_lowd_greens.md. Grounded on vibe-qc’s already-validated 2-D slab Ewald (ewald_composed_slab.py, J_2D−J_3D→c·S passing); the 1-D wire kernel and the four-center + exchange-seam unification are the new joint contribution. M2-M5 (implement → reduction gate → seam → low-D controls) pending.