HLM-Micro-anomaly-v0

Multimodal industrial-anomaly classifier in the HLM-Micro tier. Classifies sensor windows into normal / warning / fault across three input modalities (vibration, acoustic, text command) using a single shared checkpoint.

Access

Trained weights and inference code are available through the Early Access program or a commercial engagement. This page is a public summary only.

Task

3-class classification on a multimodal sensor window.

Class	Interpretation
normal	Steady-state operation
warning	Elevated harmonic content / mild acoustic rumble / warning command
fault	Strong harmonics + impulsive spikes / alarm command

Input modality is one of: vibration sensor, acoustic (mel-equivalent), or short text command — all handled by one checkpoint via the HLM-Micro modality-tag prefix.

Headline result

• Validation accuracy: 81.2% at quality mode (chance = 33%)
• Per-class accuracies: balanced — all three classes predicted independently
• Training cost: CPU-only, minutes

Runtime compute dial

The HLM-Micro runtime dial trades accuracy for latency at inference time on the same weights:

Mode	Relative accuracy	Relative latency	Use case
Emergency	~50% of peak	~25% of peak	Battery-critical / background scanning
Deploy (default)	~80%	~65%	Normal operation
Quality	100% (peak)	100%	Event response / plugged-in / certification

Precise numbers on target hardware are shared with Early Access partners.

Where it's positioned

Scenario	Fit
Factory-floor predictive maintenance	Primary — see Industrial edge use case
Multi-modal operator monitoring panel	Strong — one model handles vibration + audio + commands
Sovereign edge sensor network	See Sovereign edge sensing use case

Out-of-scope

• Not trained on your specific machinery. The public checkpoint is trained on synthetic signatures. Production deployment retrains on your labelled sensor data under a commercial engagement.
• Not a medical device. The shared architecture is used in a separate medical-adjacent checkpoint (ecg-v0); they are distinct releases.
• Not validated on adversarial inputs. Standard anomaly-detection robustness work; not hardened for deliberate attack.

Audit certificate

Every inference produces a SHA-256 hash chain binding the input, model weights, and basin trajectory — enabling replayable auditability for compliance-sensitive deployments. Same mechanism as on HLM3 checkpoints.

How to get started

1. Request weights + inference code via Early Access or commercial engagement
2. Integrate via the qriton-hlm Python package (same BasinSurgeon API as HLM3)
3. Evaluate on your data; discuss production deployment with us

License

BSL 1.1 (Apache 2.0 in 2030) once Early Access is granted.