🏛️ Enterprise Architect View

The 3-Model Psyche

JEPA is the id. Gestalt, embodied, spatial, fast. One forward pass through the world model gives you a direction without needing to count anything. JEPA

LLM is the ego. Verbal, analytical, communicable, slow. Generates tokens one at a time, builds up an explanation, can read or write. LLM

JEV is the superego. Constrained, principled, schema-bound, decision-oriented. Returns calibrated probabilities, not text. Cannot hallucinate outside your schema. JEV

Quilt cells use all three. Most substrates use one and bolt on the others. We use the psyche frame because it explains why cellular logic works: every cell needs an instinct, a voice, and a conscience.

Decision matrix: JEV vs LLM vs JEPA

When to use which

Anti-patterns: when JEV is wrong

• Open-ended generation. JEV cannot write a poem. Use LLM.
• More than 255 options. JEV caps Choice at 255. Use a 2-stage decision or fall back to LLM.
• Multi-modal input. JEV is text-only. Use JEPA/VLM for images.
• Long context reasoning. JEV reads 32k+ but reasoning is shallow. Use LLM for chain-of-thought.

Cost model

Per-call

1M decisions/month

JEV-augmented stacks are 100-1000x cheaper than LLM-only for typed decisions.

Reliability

Failure modes

• JEV 5xx — exponential backoff with jitter, 3 retries, then circuit breaker
• JEV 429 — rate limited; back off 30s, retry single request
• CF Worker 1101 — uncaught exception; check /api/status; redeploy
• Low JEV confidence — escalate to LLM or human per your routing policy

Monitoring recommendations

• Track p50/p95/p99 latency per endpoint
• Track confidence distribution (mean, std, % above threshold)
• Track fallback rate (% of decisions escalated to LLM)
• Track witness-log growth (cells/day)

Migration patterns (5 examples)

1. LLM-only spam filter → JEV + LLM

// Before: prompt-engineered classifier
const r = await openai.complete(
  "Classify this email as spam or not-spam. Email: ..."
);
if (r.includes("spam")) block();

// After: JEV-decision + LLM escalation
const jev = await jev.decide(state, {
  is_spam: {type: "noul", question: "Is this spam?", instructions: "..."}
});
if (jev.is_spam.noul > 0.9) block();
else if (jev.is_spam.noul > 0.6) return await llm.review(jev);
else deliver();

2. Regex form validator → JEV semantic validator

Regex catches format errors. JEV catches semantic errors (e.g., "555-01-0234 is a fake SSN that passes the regex").

3. Rules-engine business logic → JEV + LLM narrative

Replace 5000-line rules engine with 20 questions. Output the rules as LLM narrative for audit logs.

4. Bayesian lead scoring → JEV score + JEPA cluster

JEV replaces the score function. JEPA replaces cluster features. LLM generates the sales email.

5. Multi-model ensemble → JEV router

JEV picks the right model per request (vs always calling all of them).