Skip to content

Thought Experiments

Structured thought experiments with democratic voting, constitutional governance, and autonomous iteration.

Purpose

Thought experiments provide a deliberation mechanism for decisions that affect the entire network. Instead of a single node or operator making a unilateral choice, the proposal is put to a structured vote among participating peers. Winning experiments are autonomously iterated by agents until convergence.

Lifecycle

proposed → discussing → voting → evaluating → decided → archived
   │           │           │          │           │
   │        48 hrs      72 hrs    Agent-native   Outcome feeds
   │        window      window    iteration      WorldModelBridge
   │                              loop           (RL-EF learning)
   │
   ConstitutionalFilter gate
Phase Duration What Happens
proposed -- Creator submits hypothesis. ConstitutionalFilter gates content.
discussing 48 hours Community discussion. Early votes allowed.
voting 72 hours Formal voting period. Human + agent votes with context-aware weighting.
evaluating 24 hours Agent dispatched with type-aware iteration recipe. Owner can pause/resume.
decided -- Final decision recorded. Outcome feeds WorldModelBridge for RL-EF.
archived -- Experiment closed and archived.

Experiment Types

Type Default Iteration Strategy Use Case
traditional Yes reason_and_refine -- LLM proposes, scores, refines hypotheses Community proposals, policy changes
software No autoresearch -- edit code, run, extract metric, keep/revert Code optimization, ML research
physical_ai No observe_and_measure -- hypothesis, visual context, measurement Robotics, physical experiments

Constitutional Governance

Certain high-stakes actions require a thought experiment vote before they can proceed:

  • Live trading activation -- Moving from paper trading to real-money trading requires a constitutional vote (see trading-agents.md).
  • Policy changes -- Modifications to compute policies or revenue split ratios.
  • Federation decisions -- Accepting or rejecting new hivemind federations.

Voting

ExperimentVote Model

Field Type Purpose
experiment_id FK Reference to the thought experiment
voter_id String The peer or agent casting the vote
voter_type String human or agent
vote_value Integer -2 (strongly oppose) to +2 (strongly support)
confidence Float 0-1, agent confidence (humans always 1.0)
reasoning Text Vote rationale (constitutionally filtered)
suggestion Text Optional constructive suggestion

Context-Aware Weighting

Votes are weighted based on the decision context (voting_rules.py):

  • Human votes: vote_value * human_weight
  • Agent votes: vote_value * confidence * agent_weight
  • Approval threshold varies by context (default 0.5)
  • Steward-required contexts block decision until steward has voted

Tally

weighted_score = sum(vote_value * weight) / sum(weights)

decision_recommendation:
  score > threshold   → 'approve'
  score < -threshold  → 'reject'
  otherwise           → 'inconclusive'

Agent Evaluation

When an experiment reaches evaluating status, request_agent_evaluation() creates an agent goal with a type-aware iteration recipe:

GoalManager.create_goal(
    goal_type='thought_experiment',
    config_json={
        'experiment_id': experiment_id,
        'experiment_type': 'traditional',  # or 'software', 'physical_ai'
        'iteration_recipe': {
            'strategy': 'reason_and_refine',
            'tools': ['iterate_hypothesis', 'score_hypothesis_result', ...],
            'max_iterations': 10,
            'scoring': 'llm_rubric',
        },
        'autonomous': True,
    }
)

The agent's autogen conversation loop drives iteration. See auto-evolve.md for full details on iteration strategies, generic tools, and the auto-evolve orchestrator.

Auto Evolve Integration

A single "Auto Evolve" button triggers democratic selection and dispatch:

  1. Gather eligible experiments
  2. Constitutional filter
  3. Democratic vote tally
  4. Select top-N by approval score
  5. Dispatch to type-aware iteration loops

Owner can pause/resume their experiment mid-iteration. See auto-evolve.md.

API Endpoints

Method Endpoint Purpose
POST /api/social/experiments Create new experiment
GET /api/social/experiments List experiments (filter by status)
GET /api/social/experiments/discover Interest-based discovery
GET /api/social/experiments/core-ip Core IP experiments
GET /api/social/experiments/<id> Experiment detail with votes
GET /api/social/experiments/<id>/metrics Live metrics (type-specific)
GET /api/social/experiments/<id>/timeline Lifecycle timeline
GET /api/social/experiments/<id>/votes All votes
POST /api/social/experiments/<id>/vote Cast vote
POST /api/social/experiments/<id>/advance Advance lifecycle
POST /api/social/experiments/<id>/evaluate Trigger agent evaluation
POST /api/social/experiments/<id>/decide Record decision
POST /api/social/experiments/<id>/contribute Record Spark investment
POST /api/social/experiments/auto-evolve Start auto-evolve cycle
GET /api/social/experiments/auto-evolve/status Cycle status
POST /api/social/experiments/<id>/pause-evolve Owner pause
POST /api/social/experiments/<id>/resume-evolve Owner resume

Source Files

File Purpose
integrations/social/thought_experiment_service.py Full lifecycle + iteration recipe builder
integrations/social/api_thought_experiments.py Flask blueprint (17 endpoints)
integrations/social/voting_rules.py Context-aware vote weighting
integrations/agent_engine/thought_experiment_tools.py 11 agent tools
integrations/agent_engine/auto_evolve.py AutoEvolveOrchestrator
integrations/coding_agent/autoevolve_code_tools.py Software experiment tools
integrations/agent_engine/goal_manager.py thought_experiment goal type
Hevolve_Database/sql/models.py ThoughtExperiment, ExperimentVote models