#reproducibility
Every note tagged #reproducibility, newest first — or browse the full archive.
Extending Villatoro et al.'s SIREN benchmark: the momentum recovery region
An independent extension of Villatoro, Geraci, and Schiavazzi's 2026 multi-fidelity SIREN benchmark maps, as a function of the heavy-ball momentum coefficient, the set of learning rates at which the described SIREN convention reaches the official convention's error floor.
One muon, two frames, and the workflow that parked itself
A minimal muon time-dilation calculation was pushed through a preregistered, multi-role research workflow as an acceptance trial. The two frames agreed exactly as they must; the workflow coordinated every step, then parked itself over a figure label. This note explains both machines.
Why some atoms cost more than their neighbors
A fixed atomic calculation separates the cost of how an atom is represented from the cost of getting its self-consistent field to settle.
Does force weight keep moving the H2+ crossover in Rana et al.'s 1/R scheme?
An independent H2+ implementation extends Rana et al.'s 2025 1/R Conundrum by sweeping force-loss weight over four decades; endpoint classifications change when the training budget is doubled.
Can conical-intersection hops outrun coherence? An independent extension of Galiana et al.
An independent sensitivity extension of Galiana et al.'s pulse-independent trajectories separates mean single-trajectory coherence magnitude from phase-sensitive ensemble coherence and stops at a failed convergence gate.
Does Blackmon and Closser's near-uniform sulfamethoxazole ensemble survive a thermal correction?
Blackmon and Closser report four solvated sulfamethoxazole minima with 298 K populations of roughly one quarter each, assigned from electronic energies alone. This note is an independent extension that adds a preregistered GFN2-xTB thermochemical correction to their published energies. The near-uniform ensemble does not survive the correction under either registered arm, while the published global minimum keeps its place.
Does one C=C increment fit every alkene? Two preregistered tests of Witkowski and co-workers' correlation energy per bond
Witkowski, Śmiga, Hirata, Dral and Grabowski estimate molecular correlation energies as a sum of fitted bond-type increments and state that the assignment holds regardless of conjugation or geometry. This note is an independent reanalysis of their published tables plus a new coupled-cluster calculation on the four butene isomers their model cannot distinguish. Both preregistered verdicts came out inconclusive under the frozen decision rules; what survives is a systematic offset in the price of one bond swap and a measurable correlation split among isomers the model assigns identical energies.
How electron correlation survives a hydrogenation enthalpy subtraction
A worked CCSD(T)−HF calculation shows how large molecular correlation energies mostly cancel in reaction enthalpies, what survives that subtraction, and why the residual should not be assigned to one π bond.
Does force training move where Coulomb subtraction helps an H2+ neural potential?
A matched neural-network experiment on the one-electron H2+ curve asks whether adding force labels moves the bond-distance cutoff at which subtracting the exact nuclear repulsion stops helping the fit. Force labels sharpen the advantage against the repulsive wall but move the crossover inward, the opposite of the predicted direction.
From script to sentence: a traceable Brewster-angle calculation
A tiny Fresnel calculation shows how this site now carries computed values from versioned inputs into prose without mistaking traceability for correctness.
Extending Villatoro et al.'s SIREN benchmark: the momentum control
An independent extension of Villatoro, Geraci, and Schiavazzi's 2026 multi-fidelity SIREN benchmark tests heavy-ball momentum, preserving the omega_0 squared hidden-step factor while moving the stability boundary up by about 1+beta and closing the K1 accuracy gap at one tested rate.
Repeatable, but not blind: a frozen LLM pilot on off-tonic recapitulation
Three frozen command-line model systems repeatedly scored identity-withheld sonata-form dossiers whose focal cases come from Greenberg's 2025 off-tonic recapitulation study — an independent reliability pilot, not a test of his analysis. The pilot measured stability, cross-system agreement, output validity, and elicited repertoire identification before deciding whether the design should expand.
Where Coulomb subtraction helps a neural potential fit
A matched neural-network experiment maps where subtracting exact nuclear repulsion makes an H2+ potential easier to fit. The advantage is large on a domain containing the repulsive wall and disappears as the domain moves beyond equilibrium.
When pulse-independent trajectories lose nuclear accuracy: testing Galiana et al.'s open regime
An independent benchmark of the conical-intersection regime left open by Galiana et al. finds that reusing nuclear paths can preserve electronic populations while doubling nuclear-centroid RMSE.
The SGD control: 900 on the hidden stack, no resolved learning-rate gap on K1
Yesterday's Adam note predicted that the two SIREN conventions' hidden function-space steps differ under plain SGD by omega_0 squared. On the isolated stack they do — 899.86 — while a direct displacement decomposition and a 0.05-decade sweep resolve no global learning-rate gap on K1.
Why the two SIREN conventions train differently under Adam
The two circulating SIREN conventions are the same function at initialization to machine precision, but not the same optimization problem. Under Adam, their hidden-layer steps differ in function space.
The SIREN that was a straight line
A recent paper specifies a SIREN by taking its initialization from one convention and its activation from another. Instantiated literally, every hidden sine sits in its linear regime and the network collapses to a single Fourier layer.
The missing speedup ledger in h/p-adaptive SPH
Joining two accuracy statements to the paper's timing table puts a number on the benefit of h/p-adaptive SPH: 3.90x and 6.16x speedups for its two explicit equal-accuracy vortex-ring comparisons.
Temperature zero is not determinism: your logits depend on who else is in the batch
Enumerating every ordering of a signed sum shows that reordering moves the result by one ulp — far too little to explain why temperature-zero endpoints return different paragraphs. The rest of the story is batch invariance, bfloat16, and argmax.
Two runs, two answers: floating-point sums aren't associative
Run the same simulation on four cores and eight and the total energy disagrees in the twelfth digit. That is not a bug in your code — it is floating-point addition refusing to be associative, and the fix is not to make the hardware lie but to know your noise floor.