Quick Reference

Quick Reference

TL;DR: LLMs have hard physics—attention decays in a U-shape, errors compound as $(1-p)^N$, context grows until overflow. Build around these constraints with checkpoint-every-turn, priority stacks, and cognitive offloading.

The Three Laws

1. Finite Attention → Put critical info at start/end, not middle
2. Stochastic Accumulation → Checkpoint + retry; errors grow as $e^$
3. Entropic Expansion → Compress or evict; $C(t) = O(\log t)$ not $O(t)$

Where to Find It

• U-shaped attention curve → Fig 1, Law 1
• Priority Stack architecture → Law 1 Solution Box
• Retry math turning 36% → 99.96% → Law 2
• Poisoned well / context contamination → Part III, "The Poisoned Well"
• Parameter injection pattern → Part V, code block
• Checkpoint-every-turn diagram → Fig 4
• Auth token amnesia case study → Part VI Case Study 1
• Production KPI thresholds → Part VII table

Architecture Mantras

• Temperature: 0.2 for execution, 0.7 for planning
• Context budget: $B_ = B_ - B_ - B_$
• Never drop P0: Mission + current task are sacred
• Compress, don't truncate: Log what was evicted

Quick Reference — FikAi notebook for The Physics of AI Engineering.