data(qforge): CaH6 V_NL-bound bare gate re-run — λ=20.04 (357.9% off QE, gate stays HELD)#2963
Open
dancinlife wants to merge 4 commits into
Open
data(qforge): CaH6 V_NL-bound bare gate re-run — λ=20.04 (357.9% off QE, gate stays HELD)#2963dancinlife wants to merge 4 commits into
dancinlife wants to merge 4 commits into
Conversation
…root
The Kleinman-Bylander nonlocal block qforge_vnl_block builds the site-
INDEPENDENT form factor Σ_ij β̃_i(k+G) D_ij β̃_j(k+G') · [angular]. It was
added to the PW Hamiltonian with NO per-site phase, i.e. as a single atom
at the origin. For any cell with atoms off the origin (CaH6 = Ca@origin +
6 H at general sites) the H projectors were applied without their
Σ_k e^{-i(G_a-G_b)·τ_k} structure factor — the same S(ΔG) that already
phases V_loc. The KB eigenstates |ψ_k⟩ were therefore slightly wrong,
which is the V_NL-incompleteness the screened-vertex + SSCHA campaigns
both named as the CaH6 λ=4.13647 (5.47% off QE 4.376) root.
Fix: multiply the V_NL block by Re[S_s(ΔG)] = sg[idx*2] (identical to the
V_loc term) in BOTH qforge_assemble_h (single-species) and
qforge_assemble_h_multi (per-species sg_s). Single atom at origin → S=1
(prior anchors recovered exactly). DFPT/Sternheimer/screening/el-ph reuse
the assembled H closure, so the linear-response operator inherits the fix
(d4-generic, no per-species branch).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
cah6_realcell_compose_xval_vnl.hexa — clones the realcell compose driver but stages the REAL KB projectors (ca/h betas·ls·dij, nprojs=[6,2]) into the SCF instead of nprojs=[0,0]. The stock realcell driver ran V_NL OFF (local-only PP) — so the structure-factor fix only bites once projectors are actually on. This variant exercises the now-phased V_NL through the full SCF→compose→λ pipeline for the CaH6 gate A/B. eig probe: repoint DECK to the local CaH6_NC deck (was a foreign-host path). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… SCF iters=32) Bare full-basis vertex gate (cah6_fullbz_xval, npw_cap=0 n=645, nq=4, screened=0) re-run on the V_NL structure-factor-fixed (bound) SCF. SCF converged iters=32 e_band=-65.1002 Ha. λ=20.0374 vs QE 4.376 (rel-ε=357.9%) — the bound eigenstates move λ AWAY from QE (4.137→20.04), NOT toward. Hybrid xval selftest still PASS (rel-ε=1.6524e-7). nq2 cross-check pending. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… byte-identical to nq4) nq=2 reproduces nq=4 EXACTLY (bound SCF iters=32 e_band=-65.1002 Ha, λ=20.0374, 357.892% off QE 4.376) — λ=20 is the converged-eigenstate result, NOT a q-mesh artifact. Finalized VERDICT.md + nq2 log. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Re-runs the bare full-basis vertex gate (
cah6_fullbz_xval.hexa, npw_cap=0, n=645) on the V_NL-structure-factor-fixed (bound) SCF from PR #2959 (qforge-vnl-structfac), to test whether the V_NL root fix closes the CaH6 λ accuracy gap.Result (d6 VERBATIM — NOT tuned)
The V_NL-bound SCF moves the bare CaH6 λ DRAMATICALLY AWAY from QE, not toward it. This is an OUTCOME-3 honest finding:
Regression-clean
qforge_cah6_qe_xval_test PASS— hybrid L3 α²F assembler vs QE λ_BZ rel-ε=1.6524e-07 (the 1.65e-7 anchor preserved; independent of SCF eigenstates → confirms the V_NL fix does not regress the production HYBRID path).Verdict
Migration gate stays HELD. The HYBRID route (QE |g|² → QForge L3, rel-ε 1.65e-7) remains the only QE-grade production accuracy path.
Verdict:
.verdicts/qforge-cah6-vnl-bound-rerun/(VERDICT.md + nq4/nq2 logs + selftest log).🤖 Generated with Claude Code