Independent Audit of ELSD Metalloenzyme Locked-Result Registry: 12 of 13 Deposited-Statevector Systems Pass; Cu_SOD_minimal_CuI Sector Escape Withdraws the d⁹→d¹⁰ Mechanistic Claim

Verdict (read first)

• 12 of 13 audited systems pass the deposit's own statevector-checkable locked-gate criteria — sector cleanliness, dom_p > 0.99, and dominant-bitstring agreement across all 5 seeds hold from the raw deposited amplitudes for every passing system. The methodology is demonstrably reproducible on the deposited statevectors.
• All six audited v1.0 systems pass, with σ reproducing exactly from the deposited per-seed history files: Zn_CA2_imidazole (σ=0.4367 → 0.4367), Cu_SOD_minimal (1.7035 → 1.7035), Cu_SOD_2imidazole (0.0803 → 0.0803), Cu_SOD_2imidazole_water (0.7419 → 0.7419), Cu_SOD_protonated (0.6402 → 0.6402), Cu_SOD_acetate (0.9017 → 0.9017).
• The v1.0 Cu²⁺ SOD apo→bound 9.2× σ ratio is numerically reproduced from audited deposited data: 0.7419 / 0.0803 = 9.24, computed from independently-verified σ values that match the deposit's published values exactly.
• One system fails the deposit's sector-cleanliness criterion: Cu_SOD_minimal_CuI. All five deposited seeds have ⟨N⟩ ≈ 11.00 (target 10), ⟨2Sz⟩ ≈ +1.00 (target 0), and P(N=10 ∧ 2Sz=0) ≈ 0.0001 — i.e. essentially zero amplitude in the intended Cu⁺ d¹⁰ closed-shell singlet block. All five seeds converged instead to a state in the N=11, 2Sz=+1 (Cu²⁺ d⁹ doublet) sector with dominant bitstring |11111111111000000000⟩ — the same family as the Cu²⁺ d⁹ benchmarks. The system's published "LOCKED ✓" status is not supported by the deposited statevectors.
• The d⁹→d¹⁰ mechanistic-decomposition row is withdrawn. The README's "strongest single-variable comparison" — "Reduction from Cu²⁺ d⁹ (σ=1.7035) to Cu⁺ d¹⁰ (σ=0.7592) on the same scaffold establishes open-shell d⁹ character as a primary ruggedness driver (2.2× tightening)" — compares two states that, by the deposited evidence, are both in the Cu²⁺ d⁹ doublet sector. The 2.2× ratio is not an oxidation-state effect.
• The σ-recomputation discrepancy is characterized: v1.0 systems and the v2.0 _enforced system reproduce σ exactly from history files; default v2.0 systems diverge by 9–78%. This is a v2.0 documentation/procedure issue, not a deposit-wide arithmetic failure — the v1.0 σ-reporting path is verified sound on its own evidence, and any future deposit should document explicitly which energy quantity the σ-computation uses.
• Three v1.0 systems remain unaudited because their statevectors are not present in the downloaded v1.0 bundle: Zn_squareplanar (the σ=43.29 "Multi-Basin diagnostic"), FePorphyrin_FeII_ls, and Zn_CA2_minimal. Their published values stand as the original deposit reports them, pending availability of deposited statevectors for independent verification.

This is a corrected, superseding version of an earlier deposit. An independent audit of the deposited statevectors — using a pure-numpy verifier that imports neither the proprietary Prometheus engine nor any of the deposit's own tools — finds that 12 of 13 systems with deposited statevectors satisfy the deposit's own three statevector-checkable locked-gate criteria. The v1.0 systems anchoring the deposit's headline 9.2× Cu²⁺ SOD apo/bound variability ratio are independently audit-verified, with their published σ values reproducing exactly. One v2.0 system — the Cu⁺ d¹⁰ oxidation-state control — does not represent the electronic state described; the corresponding mechanistic-decomposition row is withdrawn. Documentation and scope refinements are noted; the deposit's core empirical content is materially stronger than the v2-only audit performed yesterday would have suggested.

Two previous version:

https://zenodo.org/records/19142883 - Version-1

https://zenodo.org/records/19163956 - Version-2

1. What the original deposit claimed

The original v2.0 deposit (March 22, 2026) extended the v1.0 locked registry to 16 results across Fe, Zn, and Cu metal branches and introduced a "three-part mechanistic decomposition" of electronic-landscape ruggedness into three separable contributions — oxidation-state/electron-count class, local coordination geometry, and metal identity. The decomposition table identified Cu_SOD_minimal_CuI as the d¹⁰ control demonstrating the oxidation-state axis:

Additionally, the v1.0 narrative claimed a 9.2× variability ratio between Cu²⁺ SOD apo (Cu_SOD_2imidazole, σ=0.0803) and water-bound (Cu_SOD_2imidazole_water, σ=0.7419) active-site fragments, cited as evidence that ligand-binding events on metalloenzyme active sites materially alter the electronic-landscape stability.

2. Independent audit (full coverage where deposited statevectors are available)

A pure-numpy audit (elsd_metalloenzyme_audit.py, bundled with this correction) was applied to the union of the v1.0 and v2.0 deposited statevector bundles. The audit imports neither the Prometheus engine nor any of the deposit's own tools; it loads the raw amplitudes from each .npz, independently computes P(N), P(2Sz), joint P(N ∧ 2Sz), ⟨N⟩, ⟨2Sz⟩, dom_p, and the dominant bitstring per seed, and grades each system against the deposit's own five-criterion locked gate (sector clean, dom_p > 0.99, identical dominant bitstring across seeds, σ consistency, penalty noise — the last deferred without engine access).

The audit auto-calibrated the deposit's α/β interleaving convention against the deposit's own canonical doublet bitstring on the first open-shell Cu seed found, locking in qubit 2i = β, qubit 2i+1 = α ("odd_alpha"). All subsequent decompositions use this convention. The README's "interleaved" language should be made explicit on this point.

2.1 Per-system audit verdicts (combined v1.0 + v2.0)

Twelve systems pass all three statevector-checkable criteria; one system fails sector cleanliness; three v1.0 systems are not present in the downloaded v1.0 bundle and cannot be audited from this deposit.

2.2 Cu_SOD_minimal_CuI: per-seed evidence

The deposit specifies Cu_SOD_minimal_CuI as Cu⁺ d¹⁰ closed-shell singlet, charge=0, mult=1, N_target=10, Sz_target=0. The deposited statevectors give, on every seed:

The deposit's own sector-cleanliness criterion (|⟨N⟩−N_target| < 0.1 AND |⟨2Sz⟩−Sz_target| < 0.1) fails by ≈10× on both metrics, on every seed. Less than 0.03% of the wavefunction's amplitude is in the intended (N=10, 2Sz=0) block. The dominant block is exactly the Cu²⁺ d⁹ doublet (N=11, 2Sz=+1) — the same dominant bitstring as the Cu_SOD_minimal Cu²⁺ d⁹ benchmark, which independently passes the audit. The five seeds converged to the same state as one another (bitstring agreement and dom_p > 0.99 hold, which is why those criteria pass), but that shared state is not the intended Cu⁺ d¹⁰ singlet.

2.3 What this means for the d⁹→d¹⁰ mechanistic claim

The deposit's headline new-result statement —

"Reduction from Cu²⁺ d⁹ (σ=1.7035) to Cu⁺ d¹⁰ (σ=0.7592) on the same scaffold establishes open-shell d⁹ character as a primary ruggedness driver (2.2× tightening). This is the strongest single-variable comparison in the dataset."

— compares the published σ values of two runs that, by the deposited statevector evidence, are both in the Cu²⁺ d⁹ doublet sector. The 2.2× ratio reflects differences in penalty configuration and optimizer convergence on the same electronic state, not a comparison between Cu²⁺ d⁹ and Cu⁺ d¹⁰. The claim and its row of the three-part mechanistic decomposition are withdrawn.

The other rows of the three-part decomposition — Cu²⁺ SOD pseudo-tetrahedral vs Cu²⁺ AOC3 T-shaped (geometry axis), and the Zn²⁺ d¹⁰ + tetrahedral endpoint (metal-identity axis) — are not affected by this finding. Cu_SOD_minimal (the Cu²⁺ d⁹ baseline) and Cu_AOC3_minimal (the T-shape comparison endpoint) both passed the audit; the Zn²⁺ chemotype ladder passed; the geometry-axis and metal-identity-axis rows survive.

2.4 The σ-recomputation discrepancy: characterized as a v2.0 procedure gap

The audit recomputed σ from the standard deviation of per-seed final energies in the deposited history.csv files (Ha → kcal/mol conversion via 627.509). The pattern is sharp:

• v1.0 systems (all six): recomputed σ = published σ to four decimal places. The σ-reporting path used in v1.0 is verified sound on its own deposited evidence.
• v2.0 _enforced system (Zn_CA2_sulfonamide_enforced): recomputed σ = published σ exactly. The _enforced variant evidently uses the same σ-reporting path as v1.0.
• v2.0 default systems (six): recomputed σ diverges from published σ by 9–78%, in a system-dependent direction.

This pattern is consistent with a v2.0 procedural change in which the σ-computation uses a different per-seed energy quantity than the final-row history energy — for instance, an engine-internal penalty-subtracted "physical" energy via the deposit's decompose_energy.py tool. The audit cannot verify this without engine access. The narrow finding is: σ-recomputation discrepancies are confined to v2.0 default systems, not present in v1.0 or in _enforced v2.0, and any future deposit should explicitly document which energy quantity is used to compute σ so external readers can reproduce published values directly from the deposited convergence histories.

This is a documentation/procedure issue, not a deposit-wide numerical failure. The v1.0 σ values — including the load-bearing ones (Cu_SOD_minimal at 1.7035, Cu_SOD_2imidazole at 0.0803, Cu_SOD_2imidazole_water at 0.7419) — are independently and exactly reproduced.

2.5 The 9.2× Cu²⁺ SOD apo/bound ratio: independently audit-verified

The deposit's most-cited v1.0 finding — "a 9.2× variability ratio between apo and water-bound Cu²⁺ active sites" — depends on the σ values of Cu_SOD_2imidazole (apo) and Cu_SOD_2imidazole_water (water-bound). Both systems pass all three statevector-checkable locked-gate criteria, and their σ values reproduce exactly from the deposited per-seed history files:

0.7419 / 0.0803 = 9.239×

The ratio is therefore numerically anchored on audit-verified data. The interpretive question — whether the seed-ensemble standard deviation reflects physical electronic-landscape ruggedness vs. VQE-optimizer behavior on a stable Hamiltonian — is not settled by this audit (see §5 Limitations); but the numbers underlying the 9.2× ratio claim are correct and reproducible from the deposited statevectors and histories.

2.6 Spin-convention notation

The deposit's README states that ⟨Sz⟩ values use "the interleaved alpha/beta spin-orbital ordering convention from decompose_energy.py", with a parenthetical noting that check_sector.py uses a different convention. The audit determined empirically that the canonical convention is qubit 2i = β, qubit 2i+1 = α (odd-position α). This is the opposite of the more common interleaved convention found in many textbook references, where qubit 2i is α by default. A future version of the README should state the interleaving explicitly to remove ambiguity for external readers.

3. What survives the audit

Twelve systems pass the deposit's three statevector-checkable locked-gate criteria. In the order they appear in the deposit's narrative:

v1.0 baseline systems (six): Zn_CA2_imidazole, Cu_SOD_minimal, Cu_SOD_2imidazole, Cu_SOD_2imidazole_water, Cu_SOD_protonated, Cu_SOD_acetate. All pass sector cleanliness, dom_p > 0.99, and bitstring agreement; all reproduce σ exactly from deposited history files. The v1.0 Cu²⁺ d⁹ baseline at σ=1.7035 (Cu_SOD_minimal) — which serves as the reference point for multiple comparisons elsewhere in the deposit — is audit-verified. The apo/bound pair (Cu_SOD_2imidazole / Cu_SOD_2imidazole_water) is audit-verified, and the 9.2× ratio claim built on it is numerically reproduced exactly.

v2.0 systems (six): Cu_AOC3_minimal (first metabolic-disease benchmark, SSAO/VAP-1), Cu_SOD_minimal_water (full-scaffold water binding), and the four Zn²⁺ CA2 inhibitor warhead chemotypes (sulfonamide, hydroxamate, carboxylate, phosphonate). All pass sector cleanliness, dom_p > 0.99, and bitstring agreement; only the _enforced sulfonamide variant reproduces σ exactly, with the other five showing σ documentation gaps in the procedure-not-arithmetic sense of §2.4.

Mechanistic-decomposition rows that survive: the Cu²⁺ SOD → Cu²⁺ AOC3 (geometry axis) row stands — both endpoints passed the audit. The Zn²⁺ d¹⁰ + tetrahedral metal-identity row stands — the four warhead chemotypes pass. Only the Cu²⁺ → Cu⁺ d⁹→d¹⁰ row is withdrawn.

4. Scope of this correction and unaudited systems

This correction is scoped to the systems for which deposited statevectors were available in the v1.0 (record 19142883) and v2.0 (record 19163956) Zenodo bundles. The audit ran on the union of those two bundles; the audit script processes any system in the registry whose statevectors are present and silently skips any system whose statevectors are not.

Three v1.0 systems are present in the deposit's narrative but absent from the downloaded v1.0 bundle's statevectors: Zn_squareplanar (published σ=43.29, the "Multi-Basin geometry-induced diagnostic"), FePorphyrin_FeII_ls (published σ=0.3375), and Zn_CA2_minimal (published σ=0.0933). The unique-name listing of the v1.0 bundle contains statevectors for six v1.0 systems, not nine. These three systems' published values stand as the original deposit reports them, but are not independently audit-verified by this correction. Either the deposited bundle is incomplete relative to the README's claimed v1.0 coverage, or those three systems' statevectors are in an earlier sub-version or supplementary file not downloaded for this audit. The most notable of the three is Zn_squareplanar, whose σ=43.29 grounds the deposit's interpretive distinction between "intra-sector landscape fragmentation" (the deposit's reading) and "sector contamination" (the failure mode documented in earlier ELSD corrections). Independent verification of Zn_squareplanar would require its deposited statevectors and would be a natural follow-up.

5. Limitations and honest scope

• The audit checks three of the deposit's five locked-gate criteria from the statevectors alone (sector cleanliness, dom_p > 0.99, bitstring agreement). The fourth criterion (σ consistency) is verified for v1.0 systems and the _enforced v2.0 system, and noted as a procedure-documentation gap for default v2.0 systems. The fifth criterion (penalty contribution audited as numerical noise) requires the proprietary engine's tools and is not externally checkable; it is deferred. None of these limitations affect the central finding on Cu_SOD_minimal_CuI, which is established by criterion (1) alone with overwhelming margin.
• The audit does not test whether the σ values reflect physical electronic-landscape ruggedness vs. VQE-optimizer behavior on a stable Hamiltonian. That distinction requires an independent in-sector exact reference (a sparse-Lanczos solver for the relevant singlet/doublet sectors). That tool exists in plan but not yet in deployment in this project's corpus; until it does, σ values from this deposit are characterized as seed-ensemble standard deviations of final converged energies under penalty enforcement, which is what they are, and not as direct measures of physical-landscape ruggedness, which would require the missing exact reference. This caveat applies equally to the systems that passed and to the system that failed.
• The audit does not test whether the underlying SCF reference is root-stable across rebuilds. The SCF-multiplicity hazard documented in the LLZO correction (Zenodo 10.5281/zenodo.20279079) has not been ruled out for the Cu and Fe open-shell systems in this deposit; this is flagged as a deeper follow-up that would require a sparse-Lanczos in-sector solver to evaluate.
• Three v1.0 systems lacking deposited statevectors (Zn_squareplanar, FePorphyrin_FeII_ls, Zn_CA2_minimal) are not addressed by this audit. Their published values stand as the original deposit reports them.
• The "Plain Language Summary" claims about drug-discovery model trustworthiness (the 9.2× variability between apo and bound, the screening-campaign implication, the metabolic-disease benchmark) are now anchored on audit-verified σ values for the components they depend on (Cu_SOD_2imidazole/water, Cu_AOC3_minimal). The σ-interpretation question in the first bullet above remains; the underlying numbers are reproduced.

6. Files in this deposit

The independent verification re-derives all results directly from the carried-over statevectors; the deposit is self-contained and re-runnable (python elsd_metalloenzyme_audit.py --dir <files>). The proprietary VQE engine is not included (consistent with the project's other corrected deposits).

7. Citation and supersession

This version supersedes and corrects the prior v2.0 deposit. The twelve systems that passed the audit — including all six audited v1.0 systems and six v2.0 systems — are retained as audited and may be cited as before, with the σ-procedure caveat in §2.4 applied to default v2.0 systems. The "d⁹→d¹⁰ tightening = 2.2× ruggedness reduction" mechanistic claim and the corresponding row of the three-part mechanistic decomposition are withdrawn. Cu_SOD_minimal_CuI should be cited only in its corrected form: "the deposited Cu_SOD_minimal_CuI statevectors converged to a Cu²⁺ d⁹ doublet (N=11, 2Sz=+1) rather than the intended Cu⁺ d¹⁰ singlet; the published 'LOCKED' status overstates what the deposited data supports."

The 9.2× Cu²⁺ SOD apo/bound ratio claim is anchored on audit-verified components and may be cited as such, with the standard caveat that whether the underlying σ values reflect physical-landscape ruggedness or VQE-optimizer behavior on a stable Hamiltonian remains open pending an independent in-sector exact reference solver.

Three v1.0 systems (Zn_squareplanar, FePorphyrin_FeII_ls, Zn_CA2_minimal) lack deposited statevectors in the bundles downloaded for this audit and remain as the original deposit reports them, not independently verified. A subsequent audit covering those systems is conditional on their statevectors becoming available.