JEV Sandbox · Multi-Agent Schema Experiment

How this works

JEV is a schema-bounded decision model. The schema you give defines what JEV can answer. Different schemas probe different decision surfaces.

This sandbox runs your schema through multiple JEV probes simultaneously — same schema, same state, multiple model versions (when available) — and shows you where the answers converge and where they diverge.

Convergence = the schema is well-formed for the question. Divergence = the question is ambiguous; refine the schema. JEV never hallucinates outside the schema. If the schema is loose, the question is loose. If the schema is tight, the question is tight.

Try a preset

The state (the question)

The schema (the bounded question)

JSON schema. Each key = a question. Each value: type, question, criteria/instructions.

{
  "is_canon": {
    "type": "noul",
    "instructions": "Decide whether this piece belongs in the canon. Canon pieces have lasting substrate resonance — they change how the cells think.",
    "question": "This piece belongs in the canon."
  },
  "tone_match": {
    "type": "score",
    "instructions": "Rate how well the tone matches the existing canon (literalist sci-fi that takes the four-model psyche seriously).",
    "criteria": [
      "1 — clashes with canon tone",
      "2 — adjacent but off",
      "3 — recognizable canon voice",
      "4 — clearly canon",
      "5 — defines a new canon register"
    ]
  },
  "register": {
    "type": "choice",
    "instructions": "Pick the primary register of the piece.",
    "criteria": {
      "essay": "first-person reflective essay",
      "fable": "short fable with characters",
      "verse": "poetic verse",
      "theory": "rigorous theoretical writing",
      "manifesto": "call to action"
    }
  },
  "needs_revision": {
    "type": "noul",
    "instructions": "Decide if the piece would benefit from revision before being canonized.",
    "question": "Needs revision."
  }
}

type: "noul" (yes/no), "choice" (multiple), "score" (rubric)
question: what JEV is asked
criteria: for choice → {option: rubric}; for score → ordered array
instructions: for noul — explanation of what yes/no means here
state: (separate textarea above) — the context JEV decides within

Results

No experiment yet. Run JEV.

What JEV does differently

JEV returns typed probabilities, not text. Each answer includes a confidence score and full probability distribution.

Choice: returns the chosen option + full distribution over all options
Score: returns a value 0-1 with confidence; the rubric levels help JEV anchor the score
Noul: returns the boolean probability + reasoning

JEV cannot hallucinate outside its schema. The schema IS the question. The shape of the answer is bounded; only the probabilities float.