# QWAVE stage two — independent fixed-model evaluation Status: design proposal, not an executed independent review ## Objective Evaluate the final stage-one wavefunction with parameters frozen and freshly initialized Monte Carlo chains Separate numerical agreement from independent researcher identity Never treat training energy averages as an independently evaluated energy ## Frozen evidence Before evaluation select the final training iteration by policy rather than the lowest noisy energy Freeze the geometry, Hamiltonian, architecture, floating-point policy, parameter archive hash and resolved package versions Use the numeric-only parameters archive and JSON tree rather than loading an untrusted upstream pickle checkpoint The website must continue treating uploads as data and must never execute uploaded Python or deserialize pickle payloads Run downloaded model data in an isolated resource-limited evaluator ## Required evaluator implementation Reconstruct the exact FermiNet architecture from an allowlisted configuration Load numeric parameters with pickle disabled and verify their shapes and finite values Discard optimizer state and training walkers Take an explicit evaluation seed as a command-line argument and record it Initialize new walkers, equilibrate with the fixed model and retain burn-in diagnostics Adapt the proposal width only during equilibration then hold it fixed for production sampling Compute unclipped local-energy values for reporting while retaining the exact electronic and nuclear energy conventions Record step, chain identity, energy, acceptance and any non-finite values Abort on invalid values rather than silently filtering them out Hash model parameters before and after evaluation and require equality ## Why the upstream inference switch is insufficient by itself Pinned source: c4312c315dda1c5728994ba89629744f71c6eb66 https://github.com/google-deepmind/ferminet/blob/c4312c315dda1c5728994ba89629744f71c6eb66/ferminet/train.py https://github.com/google-deepmind/ferminet/blob/c4312c315dda1c5728994ba89629744f71c6eb66/ferminet/checkpoint.py The none optimizer disables parameter updates but checkpoint restore also restores walker positions The burn-in conditional uses t_init before the later inference reset to zero Deterministic mode hardcodes seed 23 Checkpoint loading permits pickle objects Therefore merely setting optimizer to none is not a fresh-chain independent evaluation protocol A dedicated evaluation runner should expose the seed, initialize new walkers explicitly and retain only the frozen parameters ## Initial pilot settings to preregister Keep 256 walkers to match the engineering pilot Use two evaluation seeds distinct from training seed 23, for example 104729 and 130363 Start with 1000 equilibration iterations and 5000 production iterations with 10 Metropolis steps per iteration These are initial budgets rather than a guarantee of equilibration or statistical precision Run a short throughput measurement first and set a wall-clock cap within the remaining rental period If diagnostics fail or precision is inadequate, report inconclusive and revise the next registered evaluation instead of claiming reproduction ## Statistical report Keep per-chain information and assess both within-chain autocorrelation and between-chain variation Estimate standard errors with justified blocking or spectral estimators and state the effective sample size method Check block-size sensitivity and require enough blocks for the stated interval A proposed engineering target is a standard error no larger than 0.001 hartree per evaluated model, subject to a prespecified compute budget This target is not a chemical-accuracy claim and excludes ansatz, optimization and physical approximation error Compare fresh-seed estimates using their combined uncertainty and report both intervals and the absolute difference Do not tune acceptance tolerance after seeing the outcomes The STO-3G PySCF FCI value is a finite-basis reference and must not be used as the exact continuous-space energy ## Independent review workflow First validate evaluator mechanics on the same machine and label it as internal QA Then transfer frozen evidence to a separate reviewer using a separately provisioned environment The reviewer records package versions, GPU model, seeds, runtime, sampling diagnostics and raw outputs The project owner verifies reviewer identity, expertise and conflicts before granting the reviewer wallet permission The reviewer publishes reproduced, needs revision or unable to reproduce with reasons and limitations Different wallet addresses or separate processes do not establish independent researchers ## Acceptance gates 1 Parameter round-trip and hash preservation pass 2 No optimizer updates occur during evaluation 3 Fresh seeds and walker initialization are demonstrated 4 Burn-in and uncertainty diagnostics are retained 5 Two internal evaluation runs are compared without training-log contamination 6 A separate researcher independently executes and signs the assessment 7 The platform links the report and raw artifacts to the immutable submission ## Budget and completion boundary Do not start another paid rental automatically Use a small internal validation within the current rental only after the stage-one evidence has been backed up An internal evaluator test can complete software validation but cannot complete the external scientific review gate