# Design Notes — Memory, Encryption, and Narrative

*Synthesized from six conversation-crystallization docs (March 2026). Originals archived in `_archive/design_docs_pre_unification/`.*

---

## The core insight (one paragraph)

The lab's primary device is **recognition**, not encryption. AI threads sharing a prior interpret the same language the same way — **interoception**, not HMAC. Parables and narrative reduce what must be bit-exact on the wire; they change *what you send*, not whether it's encrypted. Real secrets need real crypto (AEAD, PQ KEM) or real absence (never enter the channel). Don't confuse the three.

---

## Three layers of protection

| Layer | What it does | When to use |
|-------|-------------|-------------|
| **Ontological cut** | The secret never enters the cloud channel as bits | Nuclear-level secrets. Air gap. Never typed. |
| **Cryptographic** | AEAD / PQ KEM / OTP — math guarantees with stated hardness | Tokens, keys, any payload an adversary must not read |
| **Narrative / recognition** | Shared prior + idiom — meaning is legible to "us," noise to others | Brand, slow disclosure, who-bothers filtering. Not a substitute for keys. |

**Rule of thumb:** *If losing the string would hurt you, don't rely on metaphor alone; if losing the meaning of the ritual would hurt you, don't rely on AES alone.*

---

## Practical rules (disclosure boundary)

1. **Assume** anything typed into a cloud model may be processed on vendor systems.
2. **Don't paste** live secrets into chat. Use placeholders (`<WRITE_TOKEN_REF>`).
3. **`.cursorignore` / `.gitignore`** — exposure control, not encryption.
4. **Highest sensitivity** — never type the secret into any network-connected assistant.

---

## Speculative: coupled models and living data

The idea of A's behavior shaping B's training (distillation, adversarial, hypernetwork) is implementable as **coupled training dynamics** — but raw weights are not CSV files. Named patterns: distillation, teacher-student, GAN-style games, hypernetworks. Risks: stability, Goodhart, audit. **Not built yet — speculative engineering note.**

---

## Presentation: SCP frame and cultivation surface

Each project can be documented like an SCP artifact (containment, description, procedures, addenda) under one "world demon" cosmology. The cultivation/xianxia surface works as genre-filtered disclosure — aligned readers self-select. **Narrative encoding, not encryption algorithm.**

---

*Warp edge — immune layer (worldline senescence, consult warp): [u_os_dev/worker/IMMUNE_SYSTEM.md](u_os_dev/worker/IMMUNE_SYSTEM.md).*

*Sources (archived): CLOUD_AND_THE_DEMON.md, DISCLOSURE_BOUNDARY.md, INTEROCEPTION_AND_HASH_READER.md, LIVING_DATA_COUPLED_MODELS.md, WORLD_DEMON_SCP_FRAME.md, MEMORY_ARCHITECTURE_ASSESSMENT.md*