ModelMetis / Other directions explored

These studies tested more diagrams, fewer diagrams and an adaptive class memory. Their measurements point the current work toward representative references for every condition and operating regime.

Adaptive Memory Raised Coverage While Fixed Memory Stayed More Accurate

Aggregated class cards replaced individual references, and the adaptive arm added reviewed examples during a 72-acquisition Ottawa stream. The added coverage came mostly from additional wrong labels, so the fixed arm remains the more accurate one. The 32 reserved acquisitions stay closed for a later evaluation.

Outcome on 72 stream acquisitionsFixed memoryAdaptive memory
Correct automatic recognition1618
Wrong automatic assignment1141
Review required4513

Class-memory results, protocol and scope. The published page summarizes the aggregate counts.

Removing Three Diagram Types Recovered One Case And Exposed Repeat Variability

Four recordings, each tested with its class present and excluded, using the 26-diagram input. These eight primary decisions are development cases; the repeat evidence is reported separately.

Observed outcomeComplete inputThree views removed
Correct primary decisions5/86/8
Correct recognition, class present1/42/4
False rejection, class present3/42/4
Correct rejection, class excluded4/44/4
D03 known-class pair of identical requestsOther / OtherOther / wrong C06

The removed views are band power, autocorrelation and dominant-frequency tracking, including their explicit numerical evidence. The model studied the 23 retained diagram types; the pruning comparison displays four essential views.

Complete crossed-study evidence

46 per-request outcomes / Factorial, regression and repeat comparisons / Selected recording provenance

All 32 primary factorial decisions on D03/D07 answer Other across the eight removal combinations. Removing spectral kurtosis alone leaves D07 unchanged. Every request completed without transport errors.

Spectra Separate Jin Families Across Directions; Ottawa Families Overlap Across Profiles

Full spectral representations group the observed Jin acquisitions by condition, while Ottawa families overlap across operating profiles. These forced-label checks use acquisition groups from the published datasets.

Jin within-family variability and correct labels with a whole direction withheld
Jin: 24 windows grouped into 12 directional acquisitions.
Ottawa within-family variability and correct labels with a whole profile withheld
Ottawa: 24 acquisitions across three profile/load combinations.
Family-by-family separation
Jin within-family versus nearest-family distance ratios
Ottawa within-family versus nearest-family distance ratios

Full numeric results and predictions / Diagram attribution audit

Every Decision Can Be Checked

Publisher labels and model answers are shown side by side for review. Blind listening comparisons and fresh recordings are the next validation steps.

Audio attribution and modifications / Archive hashes / Publication and reproduction policy