Library
The Christensen Engine Has No Twin — But It Has Relatives in Surprising Places · 6 of 12
The Christensen Engine Has No Twin — But It Has Relatives in Surprising Places
technology HIGH

Financial and Legal Domain Case Studies

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

Key Principle

Origin Financial and Neota Logic demonstrate the dual-system architecture in finance and legal compliance — domains where auditability is non-negotiable. Origin proves the pattern scales commercially (SEC-regulated, $500M ARR) with a heterogeneous LLM ensemble and 138-check compliance gateway. Neota Logic proves that declarative rules beat procedural decision trees at scale, and that the knowledge elicitation bottleneck — not technology — is the true constraint.

Why This Matters

Financial and legal domains require full traceability from input to conclusion. Every determination must be auditable. These case studies show how to build assessment systems where the LLM handles natural language while deterministic logic handles every decision that could face regulatory scrutiny.

Good Examples

Origin Financial (Composite 8/10, $500M ARR, 330% YoY growth):

  • Heterogeneous LLM ensemble: different models optimized per task (data extraction, conversation, analysis, compliance checking). (pp. 3-4, chunk 003)
  • EVAL Compliance Gateway: 138 automated checks organized as a categorized pipeline — numerical accuracy → factual consistency → suitability → disclosure compliance → privacy validation. Each stage returns pass/fail/warning independently. (p. 4, chunk 003)
  • Token Budget Allocation: 15% system prompt, 25% state context, 20% relevant history, 15% tool definitions, 10% user query, 15% output buffer. Never append entire conversation history. (p. 4, chunk 003)
  • Structured Context Persistence: "Memory isn't just token history; it's structured context persisted across sessions." User decisions stored as typed fields in state objects, not raw conversation turns. (p. 4, chunk 003)
  • Deterministic code execution for all mathematical calculations — cent-level precision. (p. 4, chunk 003)

Neota Logic (Composite 7/10, 15 years in production):

  • Declarative vs. Procedural: In procedural mode you build the flowchart. In declarative mode you state rules and the inference engine determines optimal question order automatically. Beyond ~100 rules, procedural becomes "impossibly fragile to maintain." (p. 2, chunk 004)
  • The 85/15 Rule: 85% of value from deterministic workflow design, 15% from AI augmentation. The LLM serves as a data extraction tool feeding deterministic rules — never the decision-maker. (p. 2, chunk 004)
  • Glass Box Pattern: Full traceability from input to conclusion. LLM/probabilistic layer extracts variables; deterministic layer runs immutable logic against them. Autodocs feature auto-generates human-readable documentation of the entire application's logic. (p. 2, chunk 004)
  • Smart Prompt Libraries: Prompts treated as first-class configuration objects with version history, dynamic variable injection at runtime. (p. 2, chunk 004)

Counterpoints

  • The Knowledge Elicitation Bottleneck: "The knowledge elicitation problem is consistently harder than the technology." (p. 2, chunk 004) Teams invest engineering effort to solve what is fundamentally a knowledge-capture problem. The system works technically but encodes wrong, incomplete, or non-expert rules.

  • Scaling Threshold: Procedural state machines are manageable at 15 states and 22 transitions. Beyond ~100 rules across 200+ variables, they fragment into unmaintainable spaghetti. Newcomers instinctively build procedural — the approach that feels natural at small scale collapses at large scale. (p. 2, chunk 004)

  • Compliance Validation Latency: Origin's 138 checks per output create latency concerns. Categorized pipelines allow stages to run in parallel to offset this. (p. 4, chunk 003)

Key Quotes

"memory isn't just token history; it's structured context persisted across sessions." (p. 2, chunk 001) — Origin Financial

"the knowledge elicitation problem is consistently harder than the technology" (p. 2, chunk 004) — Neota Logic

Rules of Thumb

  • Organize compliance checks as categorized pipelines, not monolithic validators.
  • Use deterministic code for all mathematical calculations — never trust LLMs with math.
  • Treat prompts as first-class configuration objects with version history.
  • Plan for the declarative transition: procedural works initially but breaks at scale.
  • Use LLMs to draft initial rule sets from reference material, then validate with domain experts.
  • Generate human-readable documentation of logic automatically for audit readiness.

Related References