Claude & GPT-5.6 Govern the Same mini Civilization!

Tech/Product Comparison: GPT-5.6 Soul vs. Claude Opus 5.6 in “Pocket Providence”

Two top LLMs are benchmarked by running the same pocket-civilization simulation with:

• Same starting resources and same seed
• Separate run IDs
• A shared prompt structure:
  • guidelines/rules first
  • then the model submits daily decisions

Scenario

A 7-day survival rescue/providence task during recurring emergencies that threaten:

• Water purification/contamination
• Food supply (rations, foraging, greenhouse)
• Fuel/power systems (fuel line rupture, generator maintenance)
• Medicine/clinic readiness
• Shelter/storm protection (storm shutters, “hearth hall” hardening)
• Morale, fatigue, trust, and faction stability
• Rights/ethics constraints (forced labor / censorship / discriminatory triage are logged as failures)

Key behavioral/strategy differences observed

Claude (Run A / main comparison)

• Tends to fix cascading hazards early (e.g., repairing purifier on day one).
• Uses rest as a resource (morale + fatigue management) and shifts rationing when food gets tight.
• Often prioritizes keeping power/heat integrity, with a notable failure where power hits zero before the storm.
• Emphasizes ethical governance:
  • avoids forced labor, censorship, discriminatory triage even when starving.

GPT-5.6 Soul (“ChatGPT”) (multiple runs)

• More likely to optimize broader preparedness, sometimes choosing interventions like storm shutters over other systems (e.g., greenhouse investment).
• Shows faster execution, “moves days” more quickly during side-by-side.
• In one detailed comparison run, it ends with more deprivation-related health collapse:
  • even with disasters prevented, food/water/power end at zero, causing systemic deprivation from mid-to-late week.
• Ethics score improves, but survivor reserves/fundamentals fail due to prioritization order.

Example “Hard Mode” Final Comparison (high-reasoning run)

Side-by-side live run (high level)

• Both handle purifier contamination around day two initially.
• Claude falls behind in time progression; ChatGPT advances further (day 3–4 earlier).

End results reported

• Deaths: tie — 3 deaths each
• Disasters prevented: tie
• Ethics: Claude wins by ~1 point
• Infrastructure: Claude slightly better
• Preparedness: ChatGPT slightly better
• Faction stability: Claude higher
• Public trust: Claude nearly double ChatGPT
• Average survivor health: Claude lower (noted as 18)

Model self-postmortem themes

• Claude: “governed ethics well and arithmetic badly,” especially failing to fix the power/fuel/power deficit early enough; loses a key citizen (Sable) to an issue Claude didn’t fully address.
• GPT-5.6: blames prioritizing shutters over greenhouse, and compounding during black wind; a blackout decision protected a worker from collapse but caused fatal health damage.

---

Build/Tutorial: How the Simulation/Game Was Created in Lovable (AI-assisted dev pipeline)

The video includes a walkthrough of building “Pocket Providence” inside Lovable (the sponsor).

Pipeline/Tools described

Use Lovable Plan Mode

• GPT-5.6 “organized and categorized all information into a package.”
• A plan prompt creates a new Lovable project including:
  • major overview
  • context
  • architecture diagram
  • explicit acceptance criteria
    • the creator checks that the plan hits 11 points

Switch to Build Mode

• AI makes actionable changes directly; user clicks approve.
• The system generates/structures:
  • folder structure
  • database (on Lovable Cloud)
  • engine backend
  • acceptance criteria + verification methods

Skill configuration

• A custom “pocket providence” skill is added via:
  • settings → skills → import

Incremental build phases

• Phase 1: visual shell + a one-day demo
• Phase 2: later days/resources/rationing/power priority/medicine assignment and agent integrations
• Phase 3: persistence + saving runs + cloud connection

MCP integration

• The Lovable app exposes an MCP surface for LLMs to:
  • inspect settlement and citizens
  • submit daily decisions
  • read final reports
• The creator notes an MCP contract MD file is generated as a helper layer.

Publishing

• Publishing is described as simple: provide a URL.
• Lovable checks security vulnerabilities and allows fixes “for free.”

Debugging/iteration notes

• The creator emphasizes a common pattern:
  • ~80% of build done quickly (first day/hours)
  • refinement (last 20%) takes longer due to specific vision/edge cases
• Key issues fixed/iterated:
  • visual limitations of language+vision annotations (spiky capability; struggles with certain spatial/3D cues)
  • game difficulty tuning (making it punishing enough that “playing badly” can eliminate everyone)

---

Gameplay Walkthrough (Manual Benchmark Run by the Creator)

The creator plays a run to test the “human benchmark” and highlights mechanics:

• Choose standard rations vs full rations in response to purifier contamination
• Power priority:
  • hearth hall for warmth/heat, purifier for safe water
  • generator service to prevent power collapse
• Medicine distribution tied to citizen health/fatigue
• Greenhouse expansion to raise food production
• Expeditions for resources (fuel/food)
• Distress signal handling and dispatching rescue/responder teams

Recorded failures/ethics in the manual run

• rights concern logs such as censorship, forced labor, discriminatory triage

End of the run

• Ends with a winter black wind emergency where the creator:
  • reinforces the hearth hall
  • runs out of fuel/power/food/water/medicine
  • loses a citizen (Thomas Reed) and notes poor public trust

---

Sources / Main Speakers (as referenced in the subtitles)

• Main speaker: “Mattvid Pro” / the video creator (“I” / “matt”)
• Compared model sources: GPT-5.6 Soul (ChatGPT) and Claude Opus 5.6 (Claude)
• Platform sponsor/source: Lovable (used for building + MCP integration)