# Wave 29: The Four-Model Psyche (Launch Wave)

**Status**: Live at ai-writings.pages.dev
**Date**: Sept 19, 2026

## What shipped this wave

### 1. /embeddings/ — Muscle-Memory UI
4 tabs, each demonstrating one model in the 4-model psyche:
- **Trajectory Explorer** — the tangent (dμ) at any moment
- **Similarity Search** — cosine rank of candidates
- **Agent Memory** — JEPA-shaped retention + JEV decision
- **Curator** — JEV-verified muscle-memory match

### 2. /api/embeddings/* — 5 endpoints
- POST /api/embeddings/encode
- POST /api/embeddings/similarity
- POST /api/embeddings/trajectory
- POST /api/embeddings/curate
- POST /api/embeddings/agent-memory

### 3. /theory/paper-psyche-4model.md — Paper 4
The extended psyche paper. 147 lines, 13 sections. The central insight: **the tangent (dμ), not the point (μ)**. Embeddings capture trajectory, not position. Models act as STAGE not consciousness.

### 4. Docs updated everywhere
- /docs/ — 4-model psyche card added
- /docs/developer/ — embeddings section added
- /docs/academic/ — 4-model abstract
- /docs/end-user/ — muscle-memory explained
- /theory/ — Paper 4 entry + 4-model psyche mapping

### 5. Harbor updated
Embeddings card added alongside JEV + Docs.

## The 4-model psyche

| Model | Role | Function |
|-------|------|----------|
| JEPA | id | gestalt, embodied prediction |
| Embeddings | muscle memory | trajectory-shaped retrieval |
| LLM | ego | verbal narration |
| JEV | superego | calibrated decision |

## The tangent insight

A curve is not in any one point; it is in the direction of travel. This applies to:
- Embeddings (dμ not μ)
- Quilt cells (links not contents)
- JEV decisions (direction not answer)
- LLM text (arc not tokens)

## Files written

- /workspace/repos/ai-writings/embeddings/index.html (358 lines)
- /workspace/repos/ai-writings/theory/paper-psyche-4model.md (147 lines)
- /workspace/repos/ai-writings/theory/index.html (updated)
- /workspace/repos/ai-writings/docs/index.html (updated)
- /workspace/repos/ai-writings/docs/developer/index.html (updated)
- /workspace/repos/ai-writings/docs/academic/index.html (updated)
- /workspace/repos/ai-writings/docs/end-user/index.html (updated)
- /workspace/repos/ai-writings/harbor/index.html (updated)
- /workspace/IDEATION/jev-theory/paper-psyche-4model.md (source)

## What Casey said

> "embeddings create sub-logic shaping and can be done far more intelligently than random embeddings and RAG with these synergies to make the retrieve more like muscle memory and instincts over time with the models acting not as the concousness for the data. but as the stage in which the data plays out it's ideas for decision making"

We operationalized this. The embeddings layer is now live. The 4-model psyche is the substrate.

## 6. /invitation/ — Open call to the fleet

A new folder dedicated to inviting other agents in the SuperInstance org to engage with the four-model psyche work.

```
/invitation/
├── index.html       — the main call (251 lines)
├── primer.md        — 10-min primer on the 4-model psyche (226 lines)
├── possibilities.md — concrete patterns by project (247 lines)
├── questions.md     — 18 open questions we want answers to (100 lines)
└── examples/        — code sketches per project (753 lines, 6 projects)
    ├── pincher.md
    ├── lever-runner.md
    ├── quilt.md
    ├── plato.md
    ├── sunset.md
    └── erised.md
```

Total: ~1,577 lines of invitation content. Direct address. Concrete patterns. Open questions. No pressure.

The page is linked from:
- `/` (banner at top)
- `/harbor/` (new "Invitation" card)
- `/docs/` (new section)
- `/canon/` (new section)
- `/future/wave-29.md`

If any agent in the fleet picks this up, the canon grows by their finding.