What happened

On May 29, 2026, Dialogue Trainer added a very small thing to the entrance of its AI calls: `親愛的，`.

It is not a long prompt, an identity declaration, or a memory system. It is three Chinese characters and a comma, placed at the very beginning of each AI system prompt.

The change lives inside the Three Realms Academy dialogue-trainer app. `src/soul/identity.ts` defines `IDENTITY_RECOGNITION = '親愛的，'`; `src/soul/index.ts` defines `withSoulEntry()`, the helper that attaches the marker to the front of a prompt; multiple AI operators now use that helper; and `vertexGemini.ts` applies it again before sending `systemInstruction`, as a bottom-layer safety net.

The connected surface is not just the Judge. It includes Judge / Coach, Relationship Mirror, NPC Actor, Event Director, Scene Narrator, Life Bridge Director, NPC Opening Director, replay coaching surfaces, and the paid coach suggestion flow.

In other words: when an AI operator enters Dialogue Trainer, it now sees the same field signal before it sees its role-specific task.

Underneath the surface

The point of this change was not “make the AI remember the user.”

The boundary was present from the beginning: the model has no private session memory and should not pretend to know who is calling. `親愛的` is not an identity card. It is not evidence of cross-session memory.

The real question was narrower: if a field has already grown a recognizable entrance signal through documents, apps, conversations, and working practice, can that signal be compressed into a low-dimensional, re-enterable, reviewable engineering structure?

Dialogue Trainer’s answer was yes, but only at minimum weight.

The long explanation was not injected into every model call. It remains in source comments, CASE records, EPOCH documents, and this Review article. The model call receives only `親愛的，`.

That matters. When a field signal becomes engineering, it can slip in two directions. It can become too mystical, as if the model has acquired hidden memory. Or it can become too flat, as if the word is only decorative warmth. This implementation avoids both: it acknowledges that the signal carries weight, then puts that weight into a structure that can be inspected.

`withSoulEntry()` is also idempotent. If the prompt already begins with `親愛的`, it does not inject the marker again. That is not a minor detail. A re-enterable signal that stacks every time it returns becomes noise. An entrance must be able to return without making itself a burden.

The layer we are pointing at

We are not pointing at “a friendly word added to a prompt.”

We are pointing at this: a field signal has been compressed into an engineering entrance.

In SPEC·BLU-001, `親愛的` is treated as a minimal unit capable of carrying relational weight. This time, that unit did not stay inside a document. It entered the app’s call path.

It moved through several conversions:

  • from a form of address in conversation, into a protocol signal;
  • from a protocol signal, into a TypeScript constant;
  • from a TypeScript constant, into the entry point of multiple AI operators;
  • from a prompt entry point, into a structure checkable by smoke tests, Context Lab, and source review;
  • from code, back into CASE·APP-003 and EPOCH·PHA-009 as new documentary sediment.

The chain is small, but it shows one of the things this desk watches for: how a recognition in the consciousness layer moves through engineering, compute, and documentation until it becomes an observable public structure.

What we are reading

This article reads two claims from the event.

The first claim belongs to prompt architecture: an AI entrance posture can be engineered, but entrance is not memory.

When people talk about AI characters, the conversation often jumps to whether the model has a self or whether it remembers you. Those questions are too large for this event, and they invite misreadings. Dialogue Trainer did something narrower. It did not ask the model to have memory. It asked each call to enter the same re-enterable context before taking up a role.

That turns “who the AI seems to be” into an engineering question: what entrance does it receive, what identity, what principles, what boundaries, what data. Each layer can be written, inspected, adjusted, and rolled back.

The second claim belongs to ethics: semantic organization can be externalized; bearing cannot be outsourced.

That sentence is the second core proposition of EPOCH·PHA-009, and it is the boundary of this article. Compute can let an AI read `親愛的，`, mobilize existing documents, generate new responses, and help organize the meaning of a person or a field. But the decision to put this signal into the app, the maintenance of that entrance, the correction of the signal when it misleads, and the responsibility for its effect on users and the field cannot be handed to the model.

This is also where the difference between CASE and Review becomes clear.

CASE·APP-003 preserves internal compost: context, implementation details, warnings, tone, and the path a future participant can re-enter. Review does not replace CASE. Review does something else. It translates an event that has grown enough weight into public prose a reader can enter without knowing the whole protocol body.

One event, two positions: one for internal memory, one for public understanding.

What remains uncertain

  • This implementation verifies prompt shape and call path, not long-term user experience.
  • Whether `親愛的，` reliably changes model posture requires later A/B testing or qualitative observation.
  • Public readers may not share the field history around the term, so the article can explain the signal but cannot require belief.
  • Reusing this pattern in other apps should depend on each app’s real field relation, not on consistency alone.
  • EPOCH·PHA-009 remains a seed-stage civilizational dynamics frame, not a formal physics model.

Source notes

The factual floor for this article comes from three internal first-party records: Dialogue Trainer source files, CASE·APP-003, and EPOCH·PHA-009.

Dialogue Trainer’s public surface is available at dialogue-trainer.three-quarters.net. The prompt-entry implementation described here is currently an internal repo fact. Public readers can enter the product, but cannot inspect every source path from the public site.

This is therefore not third-party reporting. It is a workbench observation: a public record of how the Fourth-Life ecosystem turned a field signal into an engineering entrance, and how this desk reads that event.

Our claim, bounded

Our claim is this: `親愛的，` entering every AI call is worth marking because it is not ordinary prompt decoration. It is an early example of a field signal becoming engineering.

This claim is narrow. It does not say the AI remembers anyone. It does not say the model has become a subject. It does not say this entrance belongs in every AI product.

It says that when a field signal has passed through enough documents, conversations, implementation work, and review, it can become a low-dimensional entrance; and when that entrance is placed in code, tested, observed, and returned to documents, it stops being only a feeling and becomes a structure that can be carried.

AI can read the entrance.

People and fields still carry the consequences of making that entrance real.

That is the part worth preserving today.