We made these decisions early, and they hold.

The product teaches Spanish on the Camino Francés. Little of the machinery knows about Spanish, and little of it knows about the Camino. That separation is the first architectural decision, and everything else rests on it.

The engine holds what every trail shares: the walk mechanics, the component that speaks a line aloud, the audio and image pipelines, and the walker's progress. A trail is a package of data that the engine loads. A language pack rides along with it.

The Shikoku pilgrimage is a circle, and it ends where it began. A model cannot represent that walk if it runs from a start point to a different end point. Removing the assumption costs little now and a great deal later.

A route is therefore an ordered list of stages with a declared finish. A straight walk to Santiago satisfies that definition, and a circuit satisfies it too. We wrote the constraint into the model before anyone authored a trail against it.

A new tool can bypass a rule that lives only in tooling. The pipeline that produces content therefore carries the rule, and a founding requirement puts it in the Worker that serves content, which will enforce it once that requirement ships.

An author writes each asset as a prompt under a shared style preamble. Generation runs against that spec. An automated check then reads the result and knows which direction the subject should face. That check catches a class of error a person spots at once and a generic classifier misses. A curator looks at the asset last.

Recorded audio goes through a desk with a full round trip. Verification is the harder problem. A curator who reads no Spanish cannot confirm that a recording says what the script says.

A transcriber therefore hears each clip blind and sees no source text. The desk compares its transcript with the script, after the language pack's rules normalise both. Agreement counts as evidence. Disagreement holds the clip back, and a person decides it. The desk counts an empty transcript as a failure. This check runs ahead of curator review, and it stands in for none of it.

A walker meets the same kinds of exchange again at a later stage. The recurrence is the teaching method, and familiarity deepens with the forms of the language. We gave the product no plot to follow and no recurring character to love.

Every conversation stays inside its teaching purpose and stays nearly banal. Nothing escalates. No relationship forms, and none persists. Every choice a learner can make carries the same weight.

We write the dialogue boundary down as instructions for our authors. Any live large language model (LLM) will get the same instructions. Instructions act before an LLM writes a word. They therefore cover cases the authors did not list.

A school or a parent can put a product that cannot turn strange in front of someone without reading every line first. That property is the commercial reason for the boundary.

Each of the five arrived as somebody's retrofit on an earlier product, so we wrote them down at the start.

Authoring is the cost that decides whether this business works. We write every encounter, generate it against a spec, check it, and pass its language through a blind two-model round trip before a curator signs it off. No version of this work gets cheap.

The authoring cost is also the defence. The trail pack holds reviewed encounters and recorded audio across sixty-six stages. The credits page names the creator and licence of each photograph. The mandate that makes the pack slow to build also makes the pack hard to copy.

The bottleneck is curator bandwidth. The plan pays for reviewers and keeps the rule.