TRACTIAN API · Limited availability
Introducing JANUS
JANUS is TRACTIAN's reasoning vision-language model and agent harness for industrial machines. It runs inside the TRACTIAN platform today, and is now offered as an API, through OpenAI-compatible endpoints, to a limited group of partners.
General-purpose models are unreliable on machine diagnostics. They recall plausible answers from training data instead of deriving them from the physics of the specific machine, and small errors early in an analysis compound into wrong verdicts.
JANUS takes a different approach. The model is trained to reason directly over industrial data, and an agent harness checks its claims against TRACTIAN's world models before a response is returned. On our benchmark, it scored above every frontier model and both certified human specialists, with the smallest variance between cases.
One system, two parts
JANUS is both a model and the harness around it. The model is a reasoning vision-language model optimized for agentic industrial analysis. TRACTIAN owns the weights, and inference runs on private TRACTIAN infrastructure in the United States.
The harness is the agent system that grounds the model in TRACTIAN's world models, verifies its claims, and assembles the evidence chain. Model and harness are the same system that powers diagnostics in the TRACTIAN platform today; the API exposes them to your own stack.
Owned weights
JANUS is TRACTIAN's own model. No third-party model sits behind the API.
Private US inference
Requests are served from TRACTIAN-operated infrastructure in the United States.
Data sovereignty
Your data stays under your control and is not retained after processing.
Capabilities
One endpoint covers six tasks. Each response is grounded in TRACTIAN's world models.
Reliability
Diagnose faults and component-level root cause from raw vibration, current, and process signals.
Maintenance
Turn a diagnosis into a prioritized work recommendation with a clear urgency.
Time-series analysis
Anomaly detection, trend, and change-point analysis on high-frequency sensor streams.
Process analysis
Surface inefficiencies, deviations, and interactions across multi-stage processes.
Prediction
Time-to-failure and degradation forecasts grounded in TRACTIAN's world models.
Recommendation
The next action to take, with the evidence chain that justifies it.
Benchmarks
We evaluated JANUS against five frontier models and two certified human specialists on five diagnostic cases with known physical outcomes.
| JANUS compared with | Accuracy | Consistency |
|---|---|---|
| Frontier models (Fable 5, GPT 5.6) | +47% | +210% |
| Certified human specialists | +93% | +450% |
How it works
Every request is grounded in TRACTIAN's world models: physical representations of each asset that define what the machine can and cannot do. Claims are checked against them rather than recalled from training data.
A lead agent coordinates a pool of expert agents: mechanical, electrical, operational, and more. Each agent reviews the evidence in its own domain, and the lead agent cross-checks their claims before issuing the verdict.
OpenAI-compatible endpoints
Point any OpenAI SDK at our base URL and use your API key; existing tooling works without changes. Reference an asset registered in your fleet, and JANUS retrieves its signals and context server-side, returning a structured verdict with the evidence behind each claim.
POST https://api.tractian.com/janus/v1/chat/completions
{
"model": "janus-1",
"messages": [
{
"role": "user",
"content": "Diagnose asset gbx-4471."
}
],
"response_format": { "type": "json_object" }
}{
"verdict": "input_shaft_bearing_fault",
"severity": "high",
"time_to_failure_days": 14,
"confidence": 0.94,
"recommendation": "Replace within 10 days",
"evidence": [ /* full reasoning chain */ ]
}Every response includes its evidence
Responses carry the full reasoning chain, with each step checked against physics, so your engineers can audit any verdict. Here is the chain for an example case: a compressor whose energy use started creeping up.
- 1Start from the symptomEnergy per unit produced has crept up 8% in six weeks. Output, product mix, and ambient conditions are flat. Something is wasting work.
- 2Line up the signalsThe compressor's discharge temperature rises in step with the energy, and vibration analysis picks up a faint signature growing at the same rate. Three symptoms, one clock.
- 3Rule out the processAt this operating point, the world model expects the original energy draw. The process has not changed; the machine has.
- 4Localize the causeOnly one explanation fits all three signals at once: internal leakage past a worn discharge valve, returning compressed gas as heat. The vibration signature sits where the valve operates, not at the bearings or the motor.
- 5Quantify and recommendAt the current drift rate, the valve reaches failure in about five weeks. The next planned stop is in three. Recommendation: replace it then.
Calibration on healthy machines
False alarms cost teardowns and downtime. One benchmark case was a control: a healthy compressor with a benign resonance, where the correct answer was to recommend no action.
- Production ML monitoring platformfired an alert in production
- Plant engineerescalated the alert
- Mechanical PhD specialistreported a serious fault
- Reliability engineer (Cat IV)recognized the correct tone but had the wrong reasoning
- GPT 5.5diagnosed a fault and recommended intervention
- JANUSidentified the resonance as benign; recommended no action
Results on the healthy-machine control case.
Availability
TRACTIAN maintains a vast asset library: general-purpose digital twins and world models across industrial verticals. We are accepting partners and early customers to expand JANUS beyond the TRACTIAN platform.
We work with a small cohort at a time: industrial operators with large fleets, OEMs and service providers building diagnostics into their own software, and teams in verticals our world models do not cover yet.
Partners get API access, engineering support to map their fleet to our asset library, and a direct line to the team building JANUS. Applications are reviewed on a rolling basis.