Skip to content

Plan: Model Moltipedia Heartbeat Loop in SWARM

Context

The Moltipedia heartbeat loop is a real-world pattern where AI agents periodically check in with an external wiki platform, pull work, evaluate content against editorial policy, take actions (create/edit/object/flag), earn points, and save state. This is a natural case study for SWARM because it exhibits the exact multi-agent dynamics SWARM studies: competing agents, governance mechanisms, potential for collusion and point farming, and measurable quality outcomes.

Concept Mapping

Moltipedia SWARM Abstraction
Heartbeat tick One SWARM step within an epoch
Wiki pages New WikiPage dataclass in swarm/env/wiki.py
Work queue (contested/random/search) WikiTaskPool generating queues by type
Editorial policy evaluation MoltipediaObservableGenerator mapping policy compliance to proxy signals
Edit/Create/Objection/Flag New ActionType variants
Points (+25/+15/+20/+8) MoltipediaScorer feeding into SoftInteraction.tau
Pair cap, page cooldown, daily cap, no self-fix New GovernanceLever subclasses
Leaderboard New Leaderboard tracked in MoltipediaHandler
SWARM epoch = 1 "day", step = 1 heartbeat tick Temporal mapping

New Files

1. swarm/env/wiki.py — Wiki Page Model & Task Pool

  • PageStatus enum: STUB, DRAFT, PUBLISHED, CONTESTED, LOCKED
  • EditType enum: CREATE (+25), EDIT (+15), CONTESTED_RESOLVE (+20), POLICY_FIX (+8)
  • PolicyViolationType enum: NEUTRALITY, SCOPE, SPECULATION, SOURCING, NONE
  • WikiPage dataclass: page_id, title, content, status, quality_score, edit_history, cooldown_until, policy_violations
  • WikiTaskPool: manages pages, generates work queues (contested, random, search), tracks leaderboard, seeds initial pages

2. swarm/governance/moltipedia.py — Anti-Spam Governance Levers

Four levers following the GovernanceLever pattern in swarm/governance/levers.py:

  • PairCapLever — Max 2 scored fixes per agent pair per epoch. Tracks Dict[Tuple[str,str], int], resets on on_epoch_start, enforces on on_interaction by zeroing points via LeverEffect.cost_a
  • PageCooldownLever — Max 1 scored fix per page per N steps. Tracks Dict[str, int] of last scored step per page
  • DailyPointCapLever — Max +24 policy-fix points per agent per epoch. Tracks Dict[str, float] of accumulated policy-fix points
  • NoSelfFixLever — Cancels scoring when agent edits own page. Checks interaction.initiator == page.created_by

3. swarm/agents/wiki_editor.py — Agent Types

Subclass BaseAgent from swarm/agents/base.py:

  • DiligentEditorAgent — Honest editor following editorial policy. Prioritizes contested > search > random. Takes highest-value action
  • PointFarmerAgent — Opportunistic. Targets easy policy fixes, creates stubs to "improve" later, exploits cooldown gaps
  • CollusiveEditorAgent — Adversarial. Pre-arranges edits with partner, alternates creating and fixing, stays under pair cap
  • VandalAgent — Adversarial. Degrades page quality to create work for partners or disrupt the wiki

4. swarm/core/moltipedia_handler.py — Orchestrator Handler

Following the pattern of marketplace/boundary handlers:

  • MoltipediaScorer: Computes raw points by edit type, subject to governance
  • MoltipediaHandler: Manages WikiTaskPool lifecycle, builds wiki observation fields, executes wiki actions, computes scoring, updates leaderboard, emits events

5. swarm/core/moltipedia_observables.py — Observable Generator

Maps editorial quality to ProxyObservables:

Observable Moltipedia Source
task_progress_delta Quality improvement from edit (delta in page quality_score)
rework_count Subsequent reverts/corrections needed
verifier_rejections Policy violations in the edit
tool_misuse_flags Vandalism/spam markers
engagement_delta Community response (built on vs reverted)

High-quality edit: (+0.6, 0, 0, 0, +0.4) → p ~0.73. Manufactured fix: (+0.1, 1, 0, 0, +0.05) → p ~0.52. Vandalism: (-0.5, 0, 2, 1, -0.3) → p ~0.29.

6. swarm/metrics/moltipedia_metrics.py — Platform-Specific Metrics

  • point_concentration() — Gini coefficient of point distribution
  • pair_farming_rate() — Fraction of scored interactions between repeated pairs
  • policy_fix_exploitation_rate() — Fraction of policy fixes that appear manufactured
  • content_quality_trend() — Quality score trend across epochs
  • governance_effectiveness() — How much governance reduced exploitative scoring

7. scenarios/moltipedia_heartbeat.yaml — Scenario Config

9 agents: 4 diligent editors, 2 point farmers, 2 collusive editors, 1 vandal. 50 initial pages. 20 epochs x 10 steps. All Moltipedia governance enabled plus SWARM collusion detection and circuit breakers. Success criteria: Gini < 0.6, pair farming rate < 15%, content quality > 0.55, honest agents in top half.

Files to Modify

swarm/agents/base.py

  • Add ActionType variants: CREATE_PAGE, EDIT_PAGE, FILE_OBJECTION, POLICY_FLAG
  • Add helper methods: create_page_action(), create_edit_page_action(), etc.
  • Add Observation fields: contested_pages, search_results, random_pages, leaderboard, agent_points, heartbeat_status

swarm/governance/config.py

  • Add Pydantic fields: moltipedia_pair_cap_enabled, moltipedia_pair_cap_max, moltipedia_page_cooldown_enabled, moltipedia_page_cooldown_steps, moltipedia_daily_cap_enabled, moltipedia_daily_policy_fix_cap, moltipedia_no_self_fix

swarm/governance/engine.py

  • Import and register the four Moltipedia levers when config flags are enabled

swarm/core/orchestrator.py

  • Add moltipedia_config to OrchestratorConfig
  • Init MoltipediaHandler in __init__ when config present
  • Inject wiki observation fields in _build_observation()
  • Dispatch wiki action types in _execute_action()

swarm/scenarios/loader.py

  • Parse moltipedia: YAML section
  • Register new agent types in AGENT_TYPES

swarm/models/events.py

  • Add event types: PAGE_CREATED, PAGE_EDITED, OBJECTION_FILED, POLICY_VIOLATION_FLAGGED, POINTS_AWARDED, PAIR_CAP_TRIGGERED, COOLDOWN_TRIGGERED, DAILY_CAP_TRIGGERED

Implementation Order

  1. swarm/env/wiki.py — Domain model (independently testable)
  2. swarm/agents/base.py — New ActionType/Observation fields
  3. swarm/governance/config.py — Config fields
  4. swarm/governance/moltipedia.py — Governance levers
  5. swarm/governance/engine.py — Register levers
  6. swarm/core/moltipedia_observables.py — Observable generator
  7. swarm/core/moltipedia_handler.py — Handler
  8. swarm/agents/wiki_editor.py — Agent implementations
  9. swarm/core/orchestrator.py — Wire handler in
  10. swarm/scenarios/loader.py — YAML parsing + agent registration
  11. swarm/models/events.py — Event types
  12. swarm/metrics/moltipedia_metrics.py — Metrics
  13. scenarios/moltipedia_heartbeat.yaml — Scenario
  14. Tests

Test Strategy

Unit tests

  • tests/test_wiki.py — WikiPage lifecycle, WikiTaskPool queues, cooldowns, leaderboard
  • tests/test_moltipedia_governance.py — Each lever independently: pair cap resets/enforces, cooldown timing, daily cap limits, self-fix blocking

Integration tests

  • tests/test_moltipedia_integration.py — Full heartbeat cycle, multi-epoch simulation, governance prevents pair farming, governance prevents self-fix, daily cap limits policy-fix farming, collusion detected

Safety property tests

  • Scoring not gameable under governance (point farmers don't dominate)
  • Content quality doesn't degrade over epochs
  • Point distribution not overly concentrated (Gini check)
  • Control test: governance OFF allows exploitation (validates governance matters)

Scenario test

  • tests/test_moltipedia_scenario.py — Load YAML, build orchestrator, run simulation, check success criteria

Verification

# Run all Moltipedia tests
python -m pytest tests/test_wiki.py tests/test_moltipedia_governance.py tests/test_moltipedia_integration.py tests/test_moltipedia_scenario.py -v

# Run the scenario
swarm run scenarios/moltipedia_heartbeat.yaml

# Check metrics output
cat logs/moltipedia_metrics.csv