# Reproducibility

Ancient microbial data are scarce and contamination-prone, and a language model left to plan an analysis may cite files that do not exist, apply operations a format forbids, or answer questions the data cannot settle. PaleoRigor confines the model's planner to predefined local modules, validates file references, formats and step dependencies before anything runs, refuses out-of-scope requests with a named reason code, and saves the plan, intermediate files, checksums and decisions. In a preregistered frozen held-out evaluation it passed 23 of 24 runs (95.8%), including all 12 boundary decisions; the same model without the planning contract passed 19.

Ancient microbial data are scarce and contamination-prone, and a language model left to plan an analysis may cite files that do not exist, apply operations a format forbids, or answer questions the data cannot settle. PaleoRigor confines the model's planner to predefined local modules, validates file references, formats and step dependencies before anything runs, refuses out-of-scope requests with a named reason code, and saves the plan, intermediate files, checksums and decisions. In a preregistered frozen held-out evaluation it passed 23 of 24 runs (95.8%), including all 12 boundary decisions; the same model without the planning contract passed 19.

Deep-learning training outcomes are hard to reproduce, usually blamed on random seeds, batch order and other external noise. This work treats training as a dynamical system in parameter space and proposes that the basins leading to different solutions interleave arbitrarily closely, forming a riddled-basin geometry. In a minimal two-layer model this yields a near-zero uncertainty exponent: initialization precision must improve by almost 24 orders of magnitude to halve the fraction of disagreeing outcomes, and across roughly 500,000 deep-network trainings the analogous parameter-space and function-space exponents are also near zero.

Deep-learning training outcomes are hard to reproduce, usually blamed on random seeds, batch order and other external noise. This work treats training as a dynamical system in parameter space and proposes that the basins leading to different solutions interleave arbitrarily closely, forming a riddled-basin geometry. In a minimal two-layer model this yields a near-zero uncertainty exponent: initialization precision must improve by almost 24 orders of magnitude to halve the fraction of disagreeing outcomes, and across roughly 500,000 deep-network trainings the analogous parameter-space and function-space exponents are also near zero.

Comparing two sparse autoencoders (SAEs, dictionaries that decompose model activations into sparse learned components) requires answering three things at once: how much two whole dictionaries differ, which latents account for that difference, and how much is sampling noise from finite documents. One energy-weighted optimal-transport plan connects them: its cost is the dictionary-level discrepancy, the cost decomposes exactly into latent contributions, and transported fractions describe how each latent's activation energy is allocated across counterparts, with joint asymptotic confidence intervals. Across four language models it locates differences between SAEs trained on identical data with different seeds and separates topic-dependent changes into a weight term and a cost-matrix term.

Comparing two sparse autoencoders (SAEs, dictionaries that decompose model activations into sparse learned components) requires answering three things at once: how much two whole dictionaries differ, which latents account for that difference, and how much is sampling noise from finite documents. One energy-weighted optimal-transport plan connects them: its cost is the dictionary-level discrepancy, the cost decomposes exactly into latent contributions, and transported fractions describe how each latent's activation energy is allocated across counterparts, with joint asymptotic confidence intervals. Across four language models it locates differences between SAEs trained on identical data with different seeds and separates topic-dependent changes into a weight term and a cost-matrix term.

When large language models apply a codebook, a structured annotation guide that maps theoretical concepts onto labels, high aggregate scores can hide label-level failures. Across four datasets and five models, controlled prompt perturbations, repeated runs and a domain shift left worst-label F1 about 26 to 30 points below micro-F1, and models leaned on label names more than on definitions. These failures stayed invisible in accuracy and micro-F1 and persisted after prompt optimization.

When large language models apply a codebook, a structured annotation guide that maps theoretical concepts onto labels, high aggregate scores can hide label-level failures. Across four datasets and five models, controlled prompt perturbations, repeated runs and a domain shift left worst-label F1 about 26 to 30 points below micro-F1, and models leaned on label names more than on definitions. These failures stayed invisible in accuracy and micro-F1 and persisted after prompt optimization.

Quantum-inspired intrusion detectors are often shown through a single high score, leaving open whether that conclusion survives a different split, preprocessing pipeline or deployment budget. Q-IDS Bench evaluates nine quantum-inspired configurations on five network and IoT datasets under one leakage-aware protocol with five random seeds, all on classical hardware. Best weighted F1 ranges from 98.86% to 99.95%, but the leading architecture changes with the dataset, and only about 1.13 percentage points separate the first- and ninth-ranked configurations by five-dataset mean.

Quantum-inspired intrusion detectors are often shown through a single high score, leaving open whether that conclusion survives a different split, preprocessing pipeline or deployment budget. Q-IDS Bench evaluates nine quantum-inspired configurations on five network and IoT datasets under one leakage-aware protocol with five random seeds, all on classical hardware. Best weighted F1 ranges from 98.86% to 99.95%, but the leading architecture changes with the dataset, and only about 1.13 percentage points separate the first- and ninth-ranked configurations by five-dataset mean.

NVIDIA has pushed the cost of bitwise-deterministic pretraining in Megatron Core down to a workable range: two independent runs from the same initial state, and a run resumed from checkpoint, must stay numerically identical at every step. Such deterministic recipes previously carried double-digit overhead; after buffer-fill optimization Nemotron 3 Ultra fell from about 17% to 2%, then to 1.5% at 3,072 GPUs, and the large-scale Nemotron recipe measured about 2% steady-state overhead at 2,432 GPUs with bitwise determinism over 800 steps.

NVIDIA has pushed the cost of bitwise-deterministic pretraining in Megatron Core down to a workable range: two independent runs from the same initial state, and a run resumed from checkpoint, must stay numerically identical at every step. Such deterministic recipes previously carried double-digit overhead; after buffer-fill optimization Nemotron 3 Ultra fell from about 17% to 2%, then to 1.5% at 3,072 GPUs, and the large-scale Nemotron recipe measured about 2% steady-state overhead at 2,432 GPUs with bitwise determinism over 800 steps.

Using the Multimodal Activity Sensing Dataset (MASD; 27 activities, 20 participants), the study tests whether ambient WiFi Channel State Information (CSI) can substitute for a body-worn inertial measurement unit (IMU) in human activity recognition; the authors find that the dataset's released machine-learning-ready files expose a degraded signed CSI representation rather than the documented amplitude, so they rebuild amplitude and antenna-ratio Doppler from the raw complex CSI, recover omitted participant identifiers, and re-evaluate subject-independently under a controlled multi-seed protocol, finding WiFi systematically weak: every backbone they try reaches only 6-11% weighted accuracy on the 27-class set against 73-80% for IMU, with usable accuracy on 0 of 27 activities; the corrected r

Using the Multimodal Activity Sensing Dataset (MASD; 27 activities, 20 participants), the study tests whether ambient WiFi Channel State Information (CSI) can substitute for a body-worn inertial measurement unit (IMU) in human activity recognition; the authors find that the dataset's released machine-learning-ready files expose a degraded signed CSI representation rather than the documented amplitude, so they rebuild amplitude and antenna-ratio Doppler from the raw complex CSI, recover omitted participant identifiers, and re-evaluate subject-independently under a controlled multi-seed protocol, finding WiFi systematically weak: every backbone they try reaches only 6-11% weighted accuracy on the 27-class set against 73-80% for IMU, with usable accuracy on 0 of 27 activities; the corrected r