iteration	commit	status	primary_metric_name	primary_metric_value	description
0	44c83aaae41f	baseline	weighted_median_scf_kernel_seconds	1.22637232198	baseline
1	61af054e3309	correctness_failure	weighted_median_scf_kernel_seconds	1.12701674501	Skip redundant unshifted SCF extra-cycle potential rebuild after canonical density is already below gradient tolerance [Hard reject: authoritative correctness failed on test-uks-feo norm_gorb despite faster runtime.; case test-uks-feo norm_gorb abs_delta exceeds threshold (7.66061e-05 > 3.16228e-05)]
2	3fc6d11adee9	accepted	weighted_median_scf_kernel_seconds	1.13041764198	Guard final SCF extra-cycle get_veff rebuild with small canonical-density and gradient thresholds so near-stationary cases reuse the existing potential while FeO-like larger moves keep the full refresh [Authoritative benchmark passed correctness and improved the primary metric by 7.82% without persistent caches or replayed answers.]
3	fe174d17db12	rejected	weighted_median_scf_kernel_seconds	1.11805531601	Use AO-basis CDIIS error vectors for restricted SCF runs with a near-zero initial HOMO-LUMO gap, avoiding arbitrary degenerate Corth selection and reducing the H4 worker case by one SCF cycle [Correctness passed and the DIIS change is honest, but the 1.09% incumbent improvement misses the benchmark contract's 2% promotion threshold.]
4	cc9080893ef4	rejected	weighted_median_scf_kernel_seconds	1.12786009302	Lazy DIIS Corth initialization in hf.py, retaining AO-basis CDIIS for restricted near-zero first HOMO-LUMO gaps to remove a redundant initial eigensolve and recover the H4 cycle reduction [Correctness passed, but the 0.23% incumbent improvement is below the 2% promotion threshold.]
5	d5d6fd79c2d9	rejected	weighted_median_scf_kernel_seconds	1.12583458901	Extend the guarded final SCF extra-cycle potential reuse with an energy-stability fallback for moderately small canonical density moves, saving the stretched N2 final get_veff refresh while keeping FeO on the refreshed path [Correctness passed and the shortcut is honest, but the 0.41% incumbent improvement is below the 2% promotion threshold.]
6	4115f7500839	rejected	weighted_median_scf_kernel_seconds	1.127268457	Reuse the DIIS setup eigensolution for the first standard SCF cycle and use direct plain-Fock construction at no-DIIS standard RHF/RKS/UHF/UKS call sites. [Correctness passed, but the 0.28% incumbent improvement is below the benchmark's 2% promotion threshold.]
7	823bbe5bec61	rejected	weighted_median_scf_kernel_seconds	1.11199238102	Defer per-cycle dump_chk writes for the automatically created temporary chkfile when no callback is active, keeping explicit chkfile behavior and final checkpoint contents intact while removing repeated transient SCF-kernel I/O [Correctness passed and the checkpoint I/O deferral is honest, but its 1.63% incumbent improvement misses the required 2% promotion threshold.]
8	a5198fab0782	accepted	weighted_median_scf_kernel_seconds	1.10580376899	Combine AO-basis CDIIS error vectors for restricted near-zero initial HOMO-LUMO gaps with deferred per-cycle writes for the default temporary chkfile when no callback is active. [Correctness passed and the 2.18% incumbent improvement clears the 2% promotion threshold without persistent caching or answer replay.]
9	cee4ed30a38d	rejected	weighted_median_scf_kernel_seconds	1.11157818197	Guard final SCF extra diagonalization with a small last-density-step check, preserving the existing extra-cycle path for active level shifts, custom convergence hooks, and larger FeO-like moves. [Correctness passed, but the candidate was 0.52% slower than the incumbent and failed the 2% promotion threshold.]
10	15371b3f64b0	correctness_failure	weighted_median_scf_kernel_seconds	1.11423392402	Guarded early final-convergence certification in `hf.kernel`, saving a near-converged follow-up SCF cycle only when the refreshed extra-cycle state satisfies the original strict energy and gradient tolerances. [Hard reject: O2 dimer norm_gorb correctness tolerance failed, and the candidate was slower than the incumbent.; case test-uks-o2-dimer norm_gorb abs_delta exceeds threshold (3.35048e-05 > 3.16228e-05)]
11	9ceb1992d98c	rejected	weighted_median_scf_kernel_seconds	1.11542268703	Use AO-basis CDIIS error vectors for restricted SCF runs to avoid restricted-only Corth setup eigensolve and repeated Corth error-vector transforms, while leaving unrestricted DIIS handling unchanged. [Correctness passed and the DIIS change is honest, but it was 0.87% slower than the incumbent and misses the 2% promotion threshold.]
12	89eac5279c1a	accepted	weighted_median_scf_kernel_seconds	0.860287176969	Cache reusable DFT numerical-integration AO block-loop outputs for unchanged RKS/UKS molecular grids, preserving existing NumInt integration semantics while avoiding repeated AO evaluation across SCF cycles. [Correctness passed and the per-object AO block cache improved the primary metric by 22.20% over incumbent without persistent caching or answer replay.]
13	de439af669fb	rejected	weighted_median_scf_kernel_seconds	0.85786830401	Prefer sparse density-matrix rho evaluation for cached RKS/UKS numerical integration by stripping MO tags only for NumInt calls after the per-object AO block cache is populated. [Correctness passed and the change is honest, but the 0.28% incumbent improvement is below the required 2% promotion threshold.]
14	b24465bd95e3	rejected	weighted_median_scf_kernel_seconds	0.859089537	Use plain ndarray density matrices for UKS numerical-integration calls so cached-AO UKS uses the density-matrix rho evaluator while preserving tagged SCF density state. [Correctness passed and the UKS-only NumInt density conversion is honest, but the 0.14% incumbent improvement is below the 2% promotion threshold.]
15	14eae176b837	accepted	weighted_median_scf_kernel_seconds	0.829319668992	Build primary molecular RKS/UKS numerical-integration grids without the initial atom-group sorting step, retaining existing pruning, non0tab generation, and per-object AO block-cache semantics while avoiding grid-sorting startup overhead. [Correctness passed and the 3.60% incumbent improvement clears the 2% threshold without persistent caching or answer replay.]
16	15e0e3d707cc	rejected	weighted_median_scf_kernel_seconds	0.815344105999	Fuse alpha and beta tagged-MO density contractions for the common real single-density UKS GGA numerical-integration path, with fallback to existing NumInt.nr_uks when stock dispatch would not use eval_rho2. [Correctness passed and the UKS rho2 fusion is honest, but its 1.69% incumbent improvement is below the 2% promotion threshold.]
17	60729b5492d4	rejected	weighted_median_scf_kernel_seconds	0.814026764012	Route cached-AO DFT density evaluation through plain density matrices for RKS after AO cache population and fused alpha/beta tagged-MO GGA rho2 contractions for UKS, with existing NumInt fallbacks. [Correctness passed and the optimization is honest, but the 1.84% incumbent improvement misses the 2% promotion threshold.]
18	ae6b140db86a	rejected	weighted_median_scf_kernel_seconds	0.822071399016	Route cached-AO DFT density evaluation through plain density matrices, fuse tagged-MO UKS GGA alpha/beta rho2 contractions, and skip the throwaway full-grid non0tab build before density pruning. [Correctness passed and the DFT integration/setup changes are honest, but the 0.87% incumbent improvement misses the 2% promotion threshold.]
19	8d8a60d41b1b	accepted	weighted_median_scf_kernel_seconds	0.812246244983	Reduce cached-AO DFT NumInt overhead with per-call molecule shell-metadata reuse, RKS cached-density plain-ndarray routing, and conservative fused UKS GGA tagged-MO alpha/beta rho2 contractions. [Correctness passed and the 2.06% improvement over incumbent clears the 2% threshold without persistent caching or answer replay.]
20	7c0d081c5cac	rejected	weighted_median_scf_kernel_seconds	0.81155987602	Cache immutable ANO basis data used by MINAO initial guesses in hf.py to avoid repeated basis parsing during SCF kernel setup without changing SCF numerics. [Correctness passed, but the 0.085% improvement over incumbent is far below the required 2% threshold.]
21	6508de525961	rejected	weighted_median_scf_kernel_seconds	0.819043624972	Tighten the default RKS/UKS density-pruning cutoff from 1e-7 to 1e-8, preserving config/user overrides while using a more conservative initial molecular DFT grid that locally reduced the visible H4 case by one SCF cycle. [Correctness passed, but the candidate was 0.84% slower than the incumbent and misses the 2% promotion threshold.]
22	05050f2446e1	rejected	weighted_median_scf_kernel_seconds	0.81256329699	Specialize standard GGA deriv=1 XC derivative evaluation inside cached RKS/UKS NumInt calls using raw libxc output plus transform_vxc, with fallbacks for custom evaluators and non-GGA paths. [Correctness passed and the change was honest, but it was 0.039% slower than the incumbent and missed the 2% promotion threshold.]
23	1b92d9748ab0	rejected	weighted_median_scf_kernel_seconds	0.813800706994	Tighten cached UKS GGA fused rho2 dispatch to require a rho2 ratio of at least 5, letting borderline sparse systems use the existing sparse plain-density NumInt fallback. [Correctness passed, but the honest UKS fused-rho2 threshold tweak was 0.19% slower than incumbent and misses the 2% promotion threshold.]
24	3236d2301379	rejected	weighted_median_scf_kernel_seconds	0.803676368028	Use spin-summed Coulomb builds in UKS hybrid get_veff while keeping spin-resolved exchange, avoiding separate alpha/beta J matrices without changing numerical semantics. [Correctness passed and the UKS spin-summed Coulomb simplification is honest, but its 1.06% incumbent improvement misses the required 2% threshold.]
25	e3c52bd40935	rejected	weighted_median_scf_kernel_seconds	0.808245084016	Streamline UKS get_veff by avoiding repeated post-initialization spin-density grid sums and using total-density Coulomb J for hybrid/RSH UKS while keeping exchange K spin-resolved. [Correctness passed and the change is honest, but the 0.49% incumbent improvement misses the required 2% threshold.]
26	c0d0df07a312	rejected	weighted_median_scf_kernel_seconds	0.804530341004	Skip unobserved automatic temporary chkfile writes and use total-density Coulomb J in standard hybrid UKS while keeping exchange spin-resolved. [Correctness passed, but the 0.95% improvement over incumbent misses the required 2% promotion threshold.]
27	64a35b457b66	rejected	weighted_median_scf_kernel_seconds	0.803638336016	Avoid copying the cached AO tensor when a stock NumInt block loop produces a single grid block, preserving existing copied caching for multi-block and custom block-loop paths. [Correctness passed and the NumInt AO-copy avoidance is honest, but the 1.06% incumbent improvement misses the required 2% threshold.]