Paper reactions, research notes, hardware lessons, and commentary on what seems important or overhyped in the current research landscape.
The last direction of the loop asked whether a latent trained with the label in the loop reads better than a static projection; its synthetic gate failed because the planted effect was invisible to the input's own trial mean, the recalibrated rerun passed but favoured the label-free twin, and the real-data input gate then stopped the direction.
Probing a latent layer against chance is the wrong test when the layer is a filter of its input. Against the input itself, on the explore half of an explore/confirm split, under block folds and a design null, no HLDS layer showed a significant gain in seven labelled configurations, and 38 more never got far enough to be screened.
Giving the model a richer input and fitting it by expectation-maximization (EM) made its upper layers use more dimensions, one of them about five instead of one, but in the labelled explore screen neither layer decoded the three conditions as well as the input it was built from.
A per-condition HLDS that classifies by likelihood passed a synthetic check for planted dynamics, then failed every pre-registered item on the 14 explore recordings; a static per-condition Gaussian did at least as well, and temporal structure added nothing measurable.
Written targets, a promotion rule from small screens to large grids, bit-for-bit gates and canaries let a fast experiment loop stay honest; most of what the gates caught was plumbing, and one catch was a conclusion we had drawn too early.
A shared model for 36 suspensions failed on unseen cases until two inputs were put on a log scale; a physics correction to the reference frame helped further, while width and depth changed almost nothing.
One equivariant network, evaluated once per neighbouring particle and summed, reconstructs particle-resolved flow; an outer loop that makes each particle's state consistent with its neighbours adds the most where particles crowd.
Adaptive and accelerated step rules failed during training because the map they iterate changes every cycle; at inference, where the map is fixed, they converge quickly to a fixed point that predicts worse than the partly converged state.
A mechanics engine for lifters is only trustworthy if the outputs it cannot support are unreachable by construction, not merely omitted from the current release.
A workshop number only counts once something the author did not derive agrees with it; when it does not, the notebook says so instead of hiding it.
Gradus grades and diagnoses with rules, not a model, so every outcome is reproducible; the price is a heavy authoring contract and an audit for hollow content.
Locked memos, validity gates, and paired-seed statistics turned a memory-architecture idea into a program where a negative result is as usable as a positive one.
The capture loop closed with hand-dialed recipes plus a longitudinal log; the Sony live push arrived last, set four of six fields, and was an optimization.
The strongest result was not a single winning denoiser: preserving the activity carrier and using spatial separation features for ranking proved more reliable.
EEG work gets more informative when temporal structure is part of the model instead of something the pipeline averages away.
Biosignal research gets cleaner when hardware, firmware, and modeling are treated as one system instead of three separate concerns.
Workshop design works better when theory and application are taught as one loop instead of two separate tracks.