ThoxMesh-Head-C6
Mesh Event and Slot Predictor — a THOX model trained from scratch: a 9,848-byte int8 TFLite classifier that triages mesh events on a node's microcontroller.
9,848-byte int8 TFLite model · Event accuracy 0.9979 (synthetic test set) · Action slots below gate — triage use only
Your AI. Your Data. Your Rules.™
Status: EVENT HEAD PASSES — ACTION SLOTS BELOW GATE. Event accuracy 0.9979 and both event-class F1 scores pass their gates; command, target, and argument slot accuracies do not (0.9608 / 0.7843 / 0.4519). Do not ship the action path on these weights; triage use only. Training data is synthetic — field behaviour is unmeasured. The full gate table is under Status in the card.
Summary
ThoxMesh-Head-C6 is a THOX model trained from scratch: a 9,848-byte int8 TFLite classifier that runs on a THOX mesh node's microcontroller and predicts, from a 16-feature radio/sensor event vector, the event class plus action slots (target, command, argument). It is the triage head that decides which mesh events are worth waking a larger model for. It emits a structured action line assembled from enumerated slots, so its output is always well-formed; well-formed is not the same as correct, and the measured results below say exactly where it stands.
Highlights
- Trained by THOX from scratch: no external base model — a 9,848-byte int8 TFLite classifier for a 16-feature radio/sensor event vector.
- Event head passes its gates: event accuracy 0.9979, motion F1 0.9903, sensor_alert F1 0.9862 on the synthetic test set.
- Action slots fail their gates and are labeled as such on the card: the remaining error is a task-definition question, recorded rather than tuned around.
Specifications
| Base model | None — trained by THOX from scratch |
| Model size | 9,848 bytes (int8 TFLite) |
| Format | TFLite / LiteRT-Micro (plus C array export, model_data.cc) |
| Task | Tabular classification: event class + action slots from a 16-feature vector |
| License | Apache-2.0. Trained by THOX.ai LLC; the model and additions in this repo are covered under Apache-2.0. |
| This repository contains | The int8 TFLite model, C export, and meta.json (feature and class orders — the machine-readable source of truth) |
Files
| File | Size | SHA-256 |
|---|---|---|
thoxmesh_head_c6_int8.tflite |
9,848 B | 983ce6f767cfc3fada4d8138869e828b5cbb51d636a04cd0a2116da9c827973b |
model_data.cc / meta.json / .gitattributes / model_data.h |
— | Supporting files (see repo) |
Status: read before integrating
| Gate | Threshold | This build | |
|---|---|---|---|
| Event accuracy (int8) | ≥ 0.98 | 0.9979 | Pass |
motion F1 |
≥ 0.97 | 0.9903 | Pass |
sensor_alert F1 |
≥ 0.97 | 0.9862 | Pass |
| Command-slot accuracy | ≥ 0.97 | 0.9608 | Fail |
| Target-slot accuracy | ≥ 0.95 | 0.7843 | Fail |
| Argument-bucket accuracy | ≥ 0.90 | 0.4519 | Fail |
The event head is production-quality. The three action slots are not: do not ship the action path on these weights. The remaining target/argument error is a task-definition question (the input vector does not carry the information those labels need), not a tuning gap, and it is recorded here rather than tuned around.
Run it
This is an int8 TFLite / LiteRT-Micro classifier, not a language model; it is flashed with the firmware toolchain and is not served through llama.cpp or Ollama.
hf download Thox-ai/ThoxMesh-Head-C6 --local-dir ./ThoxMesh-Head-C6
The repo's meta.json is the machine-readable source of truth for the feature and class orders; generate firmware headers from it, never hand-code the orders. Firmware must clamp predicted arguments regardless of prediction, and the head never executes an action itself: it emits a line that the host validates against policy and the live roster.
Platforms
| Platform / device | Run path |
|---|---|
| LiteRT-Micro / TFLite Micro (mesh node microcontroller) | Flashed with the firmware toolchain (see Run it) |
| Hugging Face | Download the model and meta.json from this repo |
Intended use
- Mesh event triage on the node: classify radio/sensor events and decide wake-ups.
- Out of scope: direct actuation on these weights (see Status), and medical, legal, financial, safety-critical, or other high-stakes decisions without human review.
Training data and procedure
Trained by THOX from scratch as part of the THOX small-model research line. The corpus and procedure are as described in this card at release; no external base model was used.
Evaluation
Measured results are the gate table above. Training data is synthetic: generated from published cluster and attribute identifiers and realistic radio ranges, not captured from hardware. The perfect wake-up separation in testing reflects that generator, not field performance; a labelled capture from a populated mesh is what would make these numbers field claims.
Limitations
- Action-slot accuracy is below gate on target and argument prediction; triage use only.
- Synthetic-data training: field behaviour is unmeasured.
- Behaviour reflects the generated data distribution; unusual real-world event mixes may classify worse.
License
Apache-2.0. Trained by THOX.ai LLC; the model and additions in this repo are covered under Apache-2.0.
THOX.ai — Your AI. Your Data. Your Rules.™
Provenance
Published by THOX.ai as Thox-ai/ThoxMesh-Head-C6. THOX from-scratch research line; no upstream base model. This card was conformed to THOX Model Card Standard v1.0 on 2026-10-31 (frontmatter, tags, header, files, and section structure); the technical content is as published at release.
Research
THOX MeshStack: Capacity-Priced Pipeline Sharding, Grant-Gated KV Transfer and Lease-Fenced Device Roles for a Private Household Mesh
THOX.ai Research technical report v1.1.0, 4 November 2026. Craig Ross and Tommy Xaypanya, THOX.ai LLC. DOI: https://doi.org/10.5281/zenodo.23969070 — Paper page: https://thox.ai/research/papers/meshstack — PDF: https://thox.ai/papers/meshstack-1.1.0.pdf
Citation
@misc{thoxai2026thoxmeshheadc6,
title={ThoxMesh-Head-C6},
author={THOX.ai},
year={2026},
url={https://huggingface.co/Thox-ai/ThoxMesh-Head-C6}
}