Philosophy

Control Through Architecture.

Most AI systems delegate control to the model. TRION delegates control to the architecture.

routing_frame → operation_contract → all layers

Decided Once. Valid Everywhere.

The RoutingFrame is built once. The OperationContract set once. No downstream module re-reads the raw text to answer the same question again. What the pipeline decided stands — for Orchestrator, Thinking, Task Loop, and Output simultaneously.

Classifier: 0 · Orchestrator: 0 · Task Loop: 0 · Thinking: 0–1 · Verifier: 1 · Output: 1

Three LLM Calls. Maximum.

More model is not the solution. The architecture defines where the model may think — and where deterministic code is the better choice. Tool selection, routing, and execution happen without an LLM call. A skip-Thinking config exists for confidence ≥ 0.90 to reduce the count to two, but it isn't wired into the pipeline yet — Thinking runs on every request today.

ohne operation_contract → eligible_tools = []

No Fallback to Raw Text.

The layer that derived tool selection from user text was not refactored. It was deleted. If no OperationContract is present, tool selection returns empty. Guessing is not a path.

done = required_evidence ⊆ observed_evidence · sonst: blocked

Errors Are Called Errors.

Missing arguments block before execution. Tool success alone does not complete a task. Unknown completion criteria are not treated as satisfied. The system says "I don't know" instead of guessing.