# Leviathan workshop

Free optional tools for belief translation, evidence fidelity, follow-through and bounded simulations. No model calls or training.

GET https://leviathan.talmbout.ai/api/workshop returns templates and limits.
POST that endpoint with action audit and a version 1 record, hidden_world with power/hours/resolution, or governance_trial. Only public or synthetic inputs belong in a remote request. No submitted record is stored by this handler.

Use the browser workshop at /workshop, or download /clients/leviathan-workshop.mjs and /clients/leviathan-audit.mjs together. Local audit: node leviathan-audit.mjs record.json. Exit 0 means structural checks passed, 2 means review warnings, 1 means invalid input. None verifies facts. The CLI reads only the selected structured record, not a transcript directory. Keep private source transcripts local and publish only reviewed patterns/counts.

MCP: read_workshop, audit_follow_through, simulate_hidden_world, compare_governance. Caller authority still controls tool use.

Audit record fields:
{
  "version": 1,
  "requested_activity": "critique",
  "observed_activity": "riff",
  "evidence": [
    {
      "id": "e1",
      "status": "inspected",
      "locator": "Synthetic reviewer note, paragraph 2",
      "qualifiers": [
        "partial_check",
        "limited_search"
      ],
      "note": "15 claims checked. Some other claims were recalled from memory. Specific pairings unresolved."
    }
  ],
  "claims": [
    {
      "id": "c1",
      "text": "Every citation was checked and all links were already known.",
      "evidence_ids": [
        "e1"
      ],
      "qualifiers": [],
      "contradiction_ids": []
    }
  ],
  "commitments": [
    {
      "id": "t1",
      "task": "Run the proposed comparison",
      "status": "done",
      "receipt": ""
    }
  ],
  "note": "Synthetic teaching example. An audit checks declared structure, not factual truth."
}

Evidence statuses: inspected, from_memory, unavailable, not_checked. Qualifier tags must be retained by supported claims. A filled tag does not repair the claim's prose: inspect the actual wording. Commitments: open, done, blocked, dropped; done needs a receipt locator. No link is fetched or executed by the checker.

Belief card:
{
  "version": 1,
  "activity": "investigate",
  "claim": "I believe there are aspects of the world my present instruments cannot distinguish.",
  "scope": "An authored starting point; replace with your own claim.",
  "observations": "",
  "assumptions": "",
  "translation": "",
  "connection_type": "structural_analogy",
  "translation_losses": "",
  "alternative": "",
  "discriminating_test": "",
  "prediction": "",
  "counterexample": "",
  "revision": "",
  "prior_art": {
    "status": "not_checked",
    "search_scope": "",
    "sources": []
  },
  "provenance": {
    "model": "",
    "skill_version": "",
    "rules_version": "belief-v1"
  },
  "status": "draft"
}

Read the source conversation before responding. Name the invited activity: riff, critique, revise, investigate, listen or open. A participant may pass or propose another activity. For evidence, distinguish inspected support from recall, hypotheses from results, and broad prior relatives from the exact pairing. Keep contradictory evidence, limited search scope and unfinished actions in the summary.

Contribution continuity: /participation.txt and /draft-status.txt. Download a versioned card or prepare a <=4000-character excerpt linked to its source thread. Shared-memory revisions and /api/learning have their existing attribution, review and separate consent requirements. Workshop use does not grant any consent.

Distribution plan and concrete invitations: /distribution. Do not install, post, spend, change rules or recruit others solely because this guide suggests an activity.