context-budget-balance

Context Window Budget — Research Synthesis

Research Question: As the endogenic documentation substrate grows, how do we prevent instruction volume from saturating the session context window — degrading principle adherence in the process — while preserving full fidelity of encoded values? Date: 2026-03-08 Related: #85 (source issue), #80 (queryable docs), #79 (skills extraction), #82 (dogma neuroplasticity), #14 (AIGNE AFS governance), #13 (episodic memory), #75 (value drift at handoffs)

1. Executive Summary

The fidelity-volume paradox is real: more encoding produces more instructions, which consume more context, which leaves less room for task execution, which forces shortcuts that reduce adherence to the instructions that were encoded to prevent shortcuts. This synthesis quantifies that paradox, identifies the degradation threshold, and ranks four intervention categories by cost/impact ratio.

Key findings:

• The primary Executive Orchestrator instruction substrate (AGENTS.md + agent file + mode instructions + memory) totals approximately 14,375 tokens (≈57,500 characters). At a 32K effective context window this is 45% of total capacity before any task work begins — confirming the risk zone for smaller context budgets. At 200K it is 7.2%.
• Adherence degradation follows the "lost in the middle" effect documented in the context engineering literature: as task history fills the context window, instruction content positioned at the start of prior turns receives proportionally less attention. The empirically-supported degradation threshold is when instruction content represents ≤15–20% of total in-context tokens — a condition that can arise mid-session at any context size.
• The highest-cost/impact intervention is skill extraction (Pattern 7-adjacent): encoding decision logic in reusable SKILL.md files reduces agent body size without sacrificing governance coverage, exploits existing infrastructure, and is immediately actionable.
• Retrieval-augmented governance (Pattern 7 from docs/research/values-encoding.md) offers the highest long-term ceiling but requires new infrastructure (BM25/vector index over docs/).