Skip to content

THOX.ai Research · Whitepaper, Third Edition v3.0 · 2026-11-04

Preprint, not peer reviewed

First Device Models, Trained In-House

The first THOX device models, built on in-house training rails on Kickstarter launch day: a voice-command pilot, a five-model edge wave, and every gate verdict reported with its verified-recall figure beside it.

Sadie Weaving

Preparer

Tommy Xaypanya

Accountable human author (corresponding author)

ORCID 0009-0000-4425-9565

THOX.ai LLC · Austin, Texas

DOI 10.5281/zenodo.23969824

Abstract

On October 27, 2026 — Kickstarter launch day — THOX built its in-house training rails and produced its first device models with them: a voice-command LoRA pilot on LFM3-1.7B (held-out: 1/14 function names, 0/14 full exact matches — pipeline evidence, not a production adapter) and a five-model edge wave gated by the 49-question GATE49 factuality gate. Every gate verdict in Waves A and B is FAIL under the gate’s zero-hard-failure policy; the paper reports verified recall (of 12) as the primary figure beside each total. Wave B (982-row corpus) took ThoxEdge-0.5B from 21/49 to 47/49 (verified 11/12) against a constant-abstainer floor of 25/49; ThoxEdge-0.8B’s Wave B artifact was never validly measured and was held. An eleven-class failure taxonomy, a model-card rule, and a full provenance register accompany the results. This Third Edition is a Process v1.0 reissue of the Second Edition (hardened v2.1, 2026-10-30) in the THOX Research Paper Template v1.0, content unchanged except the recorded sweep and edition changes.

In one line

First Device Models is the record of THOX’s first in-house training program: the rails, the pilot, the edge wave, and the failures documented beside the results.

47/49

ThoxEdge-0.5B after Wave B, up from 21/49, verified recall 11/12

Measured

25/49

Constant-abstainer floor on GATE49

Measured

5

Models in the edge fast-track wave

Measured

11

Failure classes in the taxonomy the program produced

Calculated

Download the paper (PDF)Zenodo record Read the paper (HTML)

Contents

  1. 1. Executive summary
  2. 2. Background: four devices and a missing model layer
  3. 3. The rails: a repeatable training and verification system
  4. 4. Program 1: the voice command pilot
  5. 5. Program 2: the edge fast track wave
  6. 6. Failure taxonomy: how the work was actually done
  7. 7. The model card rule
  8. 8. Finishing pipeline: Wave A completion record
  9. 9. Limitations and honest boundaries
  10. 10. Evidence and provenance register
  11. 11. Wave B: the corpus expansion and second gate
  12. 12. Forward: the Wave C flagship program
  13. Appendix A. Per-run configuration and provenance tables
  14. Appendix B. The ThoxAir voice bridge bench record
  15. Appendix C. Glossary
  16. Document control

Where the paper lives

Cite this paper

@misc{weaving_sadie_2026_23969824,
  author       = {Weaving, Sadie and
                  Xaypanya, Tommy},
  title        = {{First Device Models, Trained In-House}},
  month        = nov,
  year         = 2026,
  publisher    = {Zenodo},
  version      = {v3.0},
  doi          = {10.5281/zenodo.23969824},
  url          = {https://doi.org/10.5281/zenodo.23969824}
}

Licensed CC BY 4.0, matching the Zenodo record.

Your AI. Your Data. Your Rules.™