Introduction

Geltre is Nichols AI's flagship model and the first Orienting Model. It selects the task-relevant evidence from large, noisy state before a decision model or LLM sees it.

Geltre does not generate prose, make the downstream business decision, or invent missing facts. Its job is narrower: determine which supplied evidence deserves attention while preserving exact provenance.

Why orientation?

AI applications often ask one expensive model to locate evidence, decide what the evidence means, and reason or generate an answer. Geltre separates the first job so downstream compute can operate on a smaller, more relevant state.

What Geltre returns

{
  "evidence": [
    {"ref": "log:2", "relevance": 0.992}
  ],
  "model": "geltre-serviceops-linear-v03"
}

Every returned ref must exist in the state supplied to Geltre. Unknown references are rejected.

Developer preview

Geltre is proprietary Nichols AI model infrastructure. The hosted developer preview uses validated ServiceOps as the default route and also exposes an explicit CodeOps alpha profile under the frozen H256-R4 reliability contract. Neither profile is presented as a universal selector. Signed-in developers can create named live API keys, add prepaid Stripe credits, inspect recent usage, and revoke keys from the console.

Where it fits

• Bounded decisions: orient noisy state before JEV or another classifier chooses from a fixed action set.
• Agent and operations state: select the relevant logs, status, history, and tool output before an automation acts.
• Code verification: surface mutation and verification evidence before a VERIFY / REVISE / ACCEPT decision.
• LLM escalation: reduce the evidence passed to a slower generative model when deeper reasoning is actually required.

ServiceOps and the explicit CodeOps alpha are the measured hosted profiles today. Browser, document, multimodal, and other adapters are future/candidate applications until they have their own frozen evidence.