Work

Preregistration: Toroidal Log-Chirplet Transform (TLCT) versus energy-tuned DCT3D saddle-coil encoders in TokaMind on the TokaMark MAST benchmark

Year: 2026
Type: article
Abstract: Preregistration of a benchmark study, posted before any confirmatory run. The study tests whether replacing TokaMind's energy-tuned DCT3D encoder on the MAST saddle-coil input (12 coils, 50 kHz) with a Toroidal Log-Chirplet Transform (TLCT) encoder lowers test error on the five MHD-activity tasks (4-1 to 4-5) of the public TokaMark benchmark. Energy tuning keeps only time frequencies up to about 500 Hz; TLCT arms encode 500 Hz to 22.5 kHz at an equal budget of 27 to 28 values per 5 ms chunk. Model, training recipe, data and scoring are otherwise fixed. Primary hypothesis (H1): across 5 paired seeds (101 to 105), TLCT with toroidal mode numbers (arm E) has lower mean TokaMark NRMSE over tasks 4-1 to 4-5 than energy-tuned DCT3D (arm A), by a one-sided paired t-test at alpha = 0.05. H2 (E versus an STFT encoder) is tested only if H1 is supported; ablations are descriptive. All 75 primary runs (arms A, B, E x 5 tasks x 5 seeds) are scored on 500 TokaMark test shots. A small validation-loss pilot and code checks run before registration are disclosed in full. Results will be published whichever way they fall.
Language: English
Open Access status: green
OpenAlex ID: W7233030288
FWCI:
Cites:
Cited by:
Related to: 10