STARXION

Trust every computation beyond Earth.

STARXION is an experimental AI-compute integrity layer for detecting silent numerical corruption during neural-network inference before a corrupted result is accepted.

The current proof of concept wraps Transformer-style linear operations with lightweight algebraic checks, injects synthetic silent faults, rejects corrupted outputs, and recomputes the affected result.

Current benchmark

• PyTorch Transformer-style block: MultiheadAttention + LayerNorm + GELU FFN
• 1,000 / 1,000 injected single-logit faults detected
• 500 / 500 injected double-logit faults detected
• 500 / 500 injected internal FFN faults detected
• 1,000 / 1,000 corrupted output cases recovered after rejection
• 0 false positives across 300 clean checks

The raw result from the reference run is included in starxion_results.json.

Read first

• One-page overview
• Technical note
• Benchmark report

Run it

Windows

Double-click:

RUN_STARXION.bat

Command line

python -m pip install -r requirements.txt
python starxion.py

The benchmark writes:

• starxion_results.json
• starxion_report.html

How it works

For a protected linear operation, STARXION computes two independent algebraic checksum relations alongside the normal inference path. A mismatch marks the output as corrupted before acceptance. The current demo then simulates peer recovery by recomputing the affected operation locally.

Why this exists

AI inference deployed in fault-prone compute environments needs a way to distinguish a plausible-looking model result from a numerically corrupted computation. STARXION explores a software layer for that problem with low-cost verification and selective recomputation.

Scope and limitations

This repository is an engineering proof of concept.

• Faults are software-injected, not radiation-induced.
• Peer recomputation is currently simulated on the same machine.
• The project has not been tested under a radiation beam or on spacecraft hardware.
• It is not flight-qualified and not a production LLM runtime.
• The reported 100% rates apply only to the injected fault campaign in this repository; they are not a general reliability guarantee.

Next validation targets

1. Pretrained Transformer weights rather than the tiny locally trained model.
2. Real multi-device peer recomputation.
3. GPU-level fault injection and larger fault distributions.
4. Radiation/accelerator testing with hardware telemetry.
5. Measurement of end-to-end latency, energy and bandwidth overhead.

---

STARXION — silent-compute integrity for AI systems.