Superpose

Validation and evidence

Show the work.

Technical reports, selected measurements, and explicit boundaries for the capabilities presented in the collection.

Method

Validation starts from the claim a component is meant to support. It checks mathematical fidelity, library delegation, determinism, distribution realization, robustness boundaries, and empirical behavior. Findings and fixes remain visible in the technical record.

Communications probe confusion matrix across the 66-class catalog. A diagnostic of generated class structure, not downstream benchmark performance.

A decisive negative result

The controlled Counter-UAS transfer test trained on synthetic FHSS, OcuSync-surrogate, and noise spectrograms, then tested on real captures. Accuracy was 0.33—chance for the three labels—and real-drone recall was approximately zero: essentially every real drone window was classified as noise. The tested synthetic baseline therefore supports spec-faithful synthetic pretraining and benchmarking, not direct field detection.

Counter-UAS train-synthetic → test-real confusion matrix. Transfer failed at 0.33 chance accuracy; real-drone recall was approximately zero.

This result is a floor for the tested clean-synthetic and weak-learner setup, not proof that transfer can never improve. Capture augmentation, hardware impairments, real drone-free spectrum, and a stronger baseline remain required work.

Technical validation reports

Communications v1 technical validation report (PDF)

Counter-UAS v1 technical validation report (PDF)

What we do not claim

  • The hosted reports are technical validation reports, not peer-reviewed publications.
  • Generated-to-generated separation is not real-world model performance.
  • The Counter-UAS capability provides a field-ready detector; its controlled transfer test failed at chance accuracy.
  • “Validated capability” does not mean the public corpus is downloadable.

Provenance

These curated reports and figures were copied from the authoritative RF repository at commit 10571a12da2113a474a8f1006bb064ef235bfab9. Inspect the source tree.