AUDIT METHODOLOGY • SCIENTIFIC INTEGRITY

Research Methodology & Scientific Freeze

The Q-Psi research program adheres to strict pre-registration discipline, cryptographic evidence sealing, independent statistical recomputation, and transparent preservation of negative comparator results.

FREEZE COMMIT: be74ad05187e148e2fc52309f4d7f57be3784157

RECOMPUTATION SUITE: 14/14 PASSED (0 FAILED)

1. Pre-Registration & Two-Phase Commit Protocol

To prevent hypothesis drift, cherry-picking, or post-hoc parameter adjustments, every physical QPU experiment in the Q-Psi program followed a strict two-phase commit protocol:

Pre-QPU Freeze Commit: Test definitions, circuit generation algorithms, hypothesis statements, analysis scripts, and seed controls were committed to Git prior to submitting jobs to IBM Quantum.
Physical QPU Execution: Jobs were submitted to ibm_marrakesh (IBM Heron r2, 156 programmable qubits) using pre-registered parameters.
Post-QPU Ingestion Commit: Raw measurement counts and execution manifests were ingested without modifying hypotheses, seeds, or evaluation functions.

LINEAR COMMIT ANCESTRY

origin/main → main → research/quantum-data-foundry-coding-v1 → research/quantum-data-foundry-coding-v2-120qsec → audit/qpsi-master-evidence-freeze-2026-08-18

All commit timestamps chronologically precede QPU execution ingestion timestamps. Zero post-hoc parameter tuning was detected.

2. Independent Deterministic Recomputation

Every physical experiment was evaluated through independent, deterministic Python recomputation scripts directly from raw counts. The consolidated test suite executes all 14 campaign verifications:

$ python research/audit_2026_08_18/recompute/recompute_all_and_verify.py

[PASS] recompute_fractional_gate_lowering.py (6 cases verified)

[PASS] recompute_calibration_aware_layout.py (6 cases vs Qiskit O3)

[PASS] recompute_mcm_qubit_reuse.py (6 streaming cases, median 78.1% reduction)

[PASS] recompute_directed_execution.py (4 cases under bootstrap)

[PASS] recompute_mqt_cross_compiler.py (18 compilation + 6 physical cases)

[PASS] recompute_mqt_sensitivity.py (leave-one-out sensitivity verified)

[PASS] recompute_qwalk_investigation.py (NISQ decoherence floor verified)

[PASS] recompute_dynamic_routing.py (18 circuits across 3 distances)

[PASS] recompute_data_foundry_v1.py (14 records, 100% provenance)

[PASS] recompute_data_foundry_v2.py (93 records, 127.58s runtime, 100% provenance)

[PASS] recompute_grover_advantage.py (9 instances, 1.19x–2.43x query reduction)

[PASS] recompute_dynamic_bv.py (N=4..16 scaling, slope 0.1532 vs 0.6963, t=-30.65)

[PASS] recompute_restricted_simon.py (12 cases across k=2,3 and N=8..28)

[PASS] recompute_stage6f.py (8 optimization instances verified)

CONSOLIDATED VERDICT: 14/14 PASSED, 0 DISCREPANCIES

3. Statistical Resolution: Finite-Sample Jensen's Bias

During the audit, certain marginal physical fidelity point estimates (F_obs) appeared slightly above the reported 95% bootstrap confidence interval [CI_low, CI_high]. An adversarial mathematical review resolved this mechanism:

Mathematical Mechanism: Concavity of Square-Root in Bhattacharyya Estimator

The Bhattacharyya fidelity estimator is defined as F(p_hat, p_ideal) = (∑_x √(p_hat(x) · p_ideal(x)))². Because the square-root function g(u) = √u is strictly concave for u > 0, by Jensen's Inequality:

E*[√p_hat*(x)] < √E*[p_hat*(x)] = √p_hat_obs(x)

In high-dimensional support distributions (e.g. 64–256 states), the cumulative variance shifts the unadjusted percentile bootstrap distribution downward by 0.005 to 0.030.

Invariance of Difference Estimators (Δ): Because the downward shift affects both comparison arms nearly identically (Bias(F_A*) ≈ Bias(F_B*)), the bias subtracts out completely:

E*[Δ*] = E*[F_A*] - E*[F_B*] ≈ F_A - F_B = Δ_obs

Scientific Impact: ZERO. All hypothesis tests, non-inferiority margins, and case classifications were governed strictly by Δ, which remains mathematically sound and unbiased.

4. Negative Result Preservation Policy

Q-Psi explicitly preserves and publishes negative comparator results:

No Rescue Runs: When Calibration-Aware Layout failed to outperform Qiskit O3 best-of-8 seeds (mean delta $-0.0106$), no post-hoc parameter tweaks were conducted to overturn the result.
Inconclusive Boundaries: Studies with high noise floors (Restricted Simon secret recovery failures at $N > 16$, long-range dynamic routing crossover failure at 13 hops) are retained as inconclusive rather than interpolated.
Incomplete Protocol Disclosure: The unexecuted multi-backend campaign is explicitly labeled INCOMPLETE_NOT_FROZEN rather than claimed.

RESEARCH PROGRAM FREEZE MANIFESTO

NO NEW QPU JOBS: The 14 audited physical campaigns constitute the final frozen research record.
IMMUTABILITY: All SHA-256 evidence seals are cryptographically locked.
CLAIM CONSERVATISM: All external reporting strictly reflects the allowed wording and explicit qualifications.

View Verified Evidence Registry Back to Research Overview