Engineered like clinical infrastructure,
not a research demo.
Under the surface, Trigone is a layered system: perception feeding scene understanding, understanding feeding reasoning, and every conclusion carrying its evidence. Here is the architecture, at whatever depth you want to read it.
Six layers. One system.
Every level inspectable.
Each layer consumes the one below it and can be examined on its own. Choose how deep you want the description to go.
Perception
Finds and follows the instruments and anatomy in view, moment to moment.
Scene understanding
Connects those observations into a picture of what is happening: who is doing what, to what.
Temporal reasoning
Remembers how the procedure has unfolded, and uses that history to read the present.
Evidence & knowledge
Treats conclusions as claims that need support, and keeps the receipts.
Assessment & decision
Works out what matters right now, and what is worth the team's attention.
Presentation
Delivers it calmly, in forms a clinical team can check at a glance.
Conceptual architecture. Model internals, training specifics, and implementation detail stay private. See below for why.
The unglamorous properties
that make AI trustworthy.
None of these demo well. All of them decide whether a hospital, a surgeon, or a regulator can rely on the system.
Deterministic behavior
The same case, replayed, produces the same result. Determinism is what makes debugging honest, audits possible, and regulators comfortable.
Evidence grounding
Interpretations are claims with linked observations behind them. If the system says something, it can show you why.
Provenance everywhere
Data, knowledge, and models carry their origin, license, and history. Nothing enters the system anonymously.
Calibrated uncertainty
Confidence scores are measured against reality and corrected. A 0.8 should behave like a 0.8, and we test that it does.
Context modeling
Procedure, phase, and history are inputs to interpretation, not afterthoughts. The same observation can, and should, mean different things.
Human authority
Model outputs are proposals. What becomes trusted knowledge is decided by qualified people, under versioned governance.
Transparent architecture.
Private machinery.
We describe the system honestly at the conceptual level and keep competitive and partner-sensitive detail private. That boundary is deliberate, and this page tells you exactly where it sits.
What we publish
- The conceptual architecture, honestly described
- Our validation philosophy and its current status
- How we treat privacy, security, and governance
- What the platform is not yet ready to claim
What we keep private
- Model architectures, weights, and training specifics
- Dataset composition and curation pipelines
- Benchmark internals and evaluation protocols
- Anything a partner shares with us under agreement
Architecture is a promise. Validation is proof.
See how we test what this page describes, and what we do and do not claim today.