The constraint
High-assurance programs require reproducible builds, offline verification, enclave deployment, and evidence packages auditors can review — not a Jupyter notebook on a laptop. The software must be trustworthy in the engineering sense: its behavior must be predictable, bounded, and documented.Why spiking networks
Edge autonomy without connectivity. A perimeter sensor, an autonomous platform, or a signals-intelligence processor that requires a cloud connection is a sensor that can be jammed, intercepted, or denied. Spiking networks run on-chip, air-gapped, without model weights leaving the enclave. Low SWaP. Size, weight, and power are always the constraint at the tactical edge. Neuromorphic inference is orders of magnitude more efficient per inference than GPU-class hardware at the same task, for the class of event-driven, sparse-signal workloads common in defense sensing. On-chip adaptation (roadmap). Learning on the edge — updating a model based on local data without transmitting weights — is on the neuromorphic roadmap and aligns with requirements for autonomous platforms in denied-communication environments.Typical workloads
- Signals intelligence edge processing: detecting patterns in RF event streams without transmitting raw data
- Unattended perimeter sensing: always-on acoustic or seismic anomaly detection, weeks on battery
- Pattern detection in EW-adjacent signal environments: sparse, high-dimensional event streams
- Autonomous platform perception at the tactical edge: obstacle avoidance, target cueing, from event cameras
What THRINDEX provides
Reproducible builds. The.thx artifact is deterministically produced from the source model and seed. Given the same Python model and thx.compile(), the output is bit-for-bit identical. Build reproducibility is architectural, not aspirational.
Offline verification. The simulator runs without network access. thrindex run and thrindex bench work fully air-gapped. The conformance harness requires no cloud connection.
Structured evidence. The conformance report is a machine-readable document with fixture fingerprint, metric values, envelope version, and pass/fail verdict. It is a file — it can be included in an evidence package, versioned, and re-generated against the same fixture set.
No telemetry by default. The THRINDEX toolchain does not phone home. There are no usage metrics, no model upload endpoints, no dependency on external services at inference time. The artifact is self-contained and executes locally.
Open SDK — program-specific work segregated. The open SDK stays generic. Program-specific models, weights, and training data are kept in your infrastructure. THRINDEX is the toolchain, not the model.