# Remora: modular LLM orchestration

Status: running
Protocol: fixed. This document does not change while the study runs; results land in the entry, not here.

## Question

A small, cheap model splits a task and calls large models only where the work needs them. What does that routing actually save, and where does it quietly cost more than it saves?

## Method

1. A router model (small, fixed) decomposes each task into steps and assigns each step a tier: local, small remote, large remote.
2. The same task set runs three ways: routed, all-large (ceiling), all-small (floor).
3. Every step logs tokens, wall time, and whether its output survived to the final answer unchanged.

## Measures

- Cost per completed task, against the all-large ceiling.
- Quality delta, judged blind against the ceiling output.
- Overhead: routing tokens and re-work caused by a tier that was too low.

## Stopping rule

The study stops after the fixed task set has run under all three conditions three times, or when the routing overhead exceeds the savings for two consecutive runs.