Library
The Christensen Engine Has No Twin — But It Has Relatives in Surprising Places
technology

The Christensen Engine Has No Twin — But It Has Relatives in Surprising Places

2026 12 references

Competitive landscape and architectural field guide for building deterministic state machines over LLMs, based on analysis of 16 products across regulated and business domains.

state-machines llm-orchestration deterministic-control conversational-ai regulated-domains competitive-analysis

Overview

The Core Framework

  • Thick Deterministic Core, Thin LLM Shell: The state machine is authoritative about what happens next; the LLM handles only how the conversation flows within guardrails
  • The LLM must never be the decision-maker: Every successful product (16/16) arrived at this principle; every struggling product violated it
  • Evaluate, never generate: The system assesses and classifies but never generates original domain content from the LLM
  • Domain authority is the moat: Architecture is converging as best practice — the real competitive advantage is depth of domain expertise encoded in the deterministic layer
  • Build state machine first: Test with zero LLM calls, then wrap the agentic shell around it

Quick Lookup

Situation Do This Avoid This
LLM ignoring instructions Use fixed command grammar — finite set of allowed outputs Adding more detailed instructions (instruction bloat)
LLM skipping assessment steps Hide full state graph — show only current state Telling the LLM "don't skip ahead" in the prompt
Context degrading mid-session Token budget: 15/25/20/15/10/15 allocation Appending entire conversation history
Guard condition misfires Confirmation loops + fallback transitions Trusting single LLM classification
Users gaming the assessment Quality-based phase gates + effort detection Advancing on completion alone
Tool misuse by LLM Remove tools from visibility ("Available When") Instructing "don't use this tool"
Infinite backward loops Max-attempts guard on every backward transition Unlimited retry loops
Debugging LLM behavior Log assembled prompt + response per call Guessing from output alone

The Key Insight

"The machine is authoritative about what happens next. The LLM is creative about how the conversation flows. Keep that boundary sharp and you'll avoid the most expensive mistakes these teams made." (p. 1, chunk 006)

References