Skip to content

Context Windows

The context window is the fixed-size buffer that defines the total amount of text a large language model can process in a single invocation. Current models range from 128K to over 1M tokens. For reference, 128K tokens corresponds to approximately 200-300 pages of text — roughly equivalent to a complete PWR Emergency Operating Procedure set.

Hard Boundary

The context window is an absolute boundary. Everything outside the window does not exist to the model. There is no partial awareness, no degraded access — information beyond the window is simply absent. This interacts directly with statelessness: since the model has no memory between calls, the context window is the entirety of what the model knows at any given moment.

Usable vs. Raw Capacity

The usable capacity is significantly less than the raw token limit. System prompts, conversation history, tool definitions, and retrieved documents all consume context space. A 128K window may have only 80-90K tokens available for actual operational content after infrastructure overhead.

Context Degradation

Two empirical findings constrain effective use:

Lost in the middle (Liu et al., 2024): Models systematically underuse information in the middle of long contexts, preferring the beginning and end. This means the ordering of information matters for accuracy — identical content produces different results depending on position within the context.

Context rot (Hong et al., 2025): Every model degrades at every length increment. Performance is not flat up to the window limit and then drops off — it degrades progressively as context grows. Du et al. (2025) measured 13.9% to 85% degradation with length even when retrieval was perfect. This is not a retrieval problem; it is an attention problem inherent to the transformer-architecture.

Over-Retrieval Problem

A counterintuitive finding: retrieving more relevant documents can degrade output quality. Adding marginally relevant content consumes context space, introduces noise, and triggers context rot. For retrieval-augmented-generation systems, retrieval precision matters more than recall. It is better to include three highly relevant paragraphs than thirty somewhat relevant pages.

Nuclear Implications

For nuclear applications, context window limitations mean:

  • Long procedures may need to be chunked and managed across multiple invocations via context-management
  • Multi-agent systems with shared context fill windows rapidly (8 agents generating 300 tokens every 45 seconds fills 128K in approximately 38 minutes)
  • Critical safety information must be positioned carefully within the context, not buried in the middle
  • The system must track what is and is not in each agent's current context to understand what it can and cannot reason about — directly linking to situation-awareness