Case study 03 · Station 04 · Trust

The deterministic gate

The most important design decision in our journey-design engine was refusing to let the LLM design the journey.

Eight-step flow: steps one to seven — function, lifecycle stage, touchpoint, audience, context, competencies, validation — then a human-approval gate, then step eight, journey assembled. The LLM contributes rationale prose only, after the gate. 1 2 3 4 5 6 7 8 function lifecycle stage touchpoint audience context competencies validation journey assembled human approves competencies LLM: rationale prose only
Steps 1–7 are deterministic; the navy diamond is the human gate at step 7; the LLM enters only after the gate, to write rationale prose.

01 ·The problem.

The seduction: the obvious 2025-era build was a prompt: “design me a learning journey for X.” Impressive in demos, unshippable in enterprise: non-reproducible outputs, no audit trail, and failure modes that surface in front of a client’s CHRO.

02 ·Discovery and the decision.

I inverted the architecture. I specified the engine as an 8-step deterministic wizard — function, lifecycle stage, touchpoint, audience, context, competencies, validation, assembly. The decision was to keep the model out of the design step entirely and let it in only after the decisions were made.

03 ·The build.

Every step is rule-driven against structured catalog data. The output is reproducible: same inputs, same journey, every time. New academies extend the system as data, not code. Architecture Decision Records document each tradeoff, and the test suite covers the deterministic core completely — which is only possible because it IS deterministic.

04 ·Trust.

Step 7 is a hard stop: the human reviews and approves the proposed competency set before any journey is generated. Not a soft confirmation — a gate the system cannot pass alone. The LLM enters only afterwards, to write the rationale narrative explaining the assembled journey. Language is what it’s reliably good at; judgment stays human.

05 ·Launch and learn.

Why this is a product decision, not an engineering one: enterprise buyers audit. Reproducibility, explainability, and a human sign-off in the loop are what turn “AI-generated” from a liability disclosure into a selling point.

In enterprise AI, knowing where NOT to use the model is the senior skill.