Horizon turns a question into a structured research artifact: direct findings, visible uncertainty, and evidence you can inspect without the sentimental filler.
The interface is part of the model
Most research assistants inherit the rhythm of consumer chat: acknowledgment, reassurance, a recap of the prompt, and only then an answer. That rhythm can be pleasant, but it also makes it harder to distinguish evidence from presentation.
Horizon treats the response as a research artifact rather than a conversation turn. Findings, supporting evidence, uncertainty, and next actions are separate objects. The goal is not to sound detached; it is to keep the useful parts inspectable.
A modular research pass
A Horizon request moves through four explicit stages: framing the question, retrieving or accepting source material, extracting claims, and composing a bounded synthesis. Each stage can fail independently, so each stage needs its own visible status and evaluation.
The modular structure also lets North use smaller specialized models where a general model would be wasteful. A compact extractor can collect candidate claims while a stronger reasoning pass resolves contradictions. Cost is assigned to the stage that actually needs capability.
What the system must admit
No interface removes bias. Retrieval can over-represent popular sources, prompts can hide assumptions, and a fluent conclusion can exceed the evidence beneath it. Horizon is designed to attach those limits to the output instead of burying them in a generic disclaimer.
The operating contract is simple: direct language, traceable claims when sources exist, no invented citations, and a clear distinction between measured results and internal targets. That contract matters more than conversational charm.