Oikoumene: Autonomous Agent Civilization Simulator

*A research project by *

GeoLambda GmbH

This simulation was developed primarily with Claude Code, Anthropic's agentic CLI, using both Claude Opus 4.6 and Opus 4.7. The collaboration served as a real-world stress test of the latest coding LLM through extensive prompt engineering.

A physics-based, AI-driven simulation of human civilization on Planet Earth — from the Out-of-Africa migration 70,000 years ago to climate futures beyond 2100.

Each autonomous agent uses a JEPA world model (LeCun 2022; Maes et al. 2026) to perceive its environment and plan goal-directed movement in latent space. Goal selection follows a Maslow-style needs hierarchy modulated by personality traits — a dual-process architecture (Kahneman 2011) combining symbolic utility-AI for what-to-do with neural latent planning for how-to-move. Settlements, nations, trade networks, conflicts, and trait evolution are fully emergent on our planet's surface with real climate data.

Status:v0.3.1 — asecurity & correctness patch: closed an LLM API-key exfiltration path and DOM XSS on the web surface, added SSRF/base-URL validation and input hardening, made the background tick loop reset-safe, and fixed a batch of simulation-core bugs (Present-Day nations no longer self-delete, seeded-RNG reproducibility, logger lifecycle, paleo→macro climate continuity, drought/regen composition). See the## [0.3.1]

section of[CHANGELOG.md]. v0.3.0 added the optionalPyTorch JEPA backend(paper-aligned toggles, NumPy/Torch parity tests). Substantialunreleased work toward v0.4remains in the[Unreleased]

section:

• the empirical-input download pipeline(CHELSA, SoilGrids, ETOPO, HYDE, UCDP-GED, HadCRUT5) andSobol sensitivity analysis— the downs and their tests are implemented, but the runtimeingestion is not yet wired(the simulation still runs on the syntheticdata/earth_*.npy

grids);anthropogenic CO₂ coupled to the civilization's industrialisation(with a continuous paleo→Industrial climate handoff),seeded-RNG reproducibility, engine hardening (agent/settlement pruning, conservative economy), and performance vectorisation that removed the periodic per-tick stalls.v0.2.x were calibration releases that fixed scientific-constant, coupling, and timing bugs. See

[CHANGELOG.md]for the full history. TheBuilt primarily with Claude Code

framing in the header applies to the v0.1.0 generation; subsequent calibration/feature passes were human-led reviews with LLM assistance.

Tested on:

Ubuntu 22.04 ARM, Python 3.11.

macOS: requires Python 3.11+ (e.g.

conda create -n oikoumene python=3.11

orpython3.11 -m venv .venv

); Python 3.9 from miniconda base will fail due to eventlet/kqueue incompatibility. Disable AirPlay Receiver or change port from 5000 to 5001 inapp.py

.Windows: untested, please file issues.

Oikoumene running the Present-Day scenario — autonomous agents on a real population-density-weighted distribution, with live macro state (CO₂, temperature, population), JEPA agent cognition, and emergent nations.

                          +-----------------------+
                          |   Leaflet.js Frontend  |
                          |  Satellite / OSM tiles |
                          +-----------+-----------+
                                      |  WebSocket (SocketIO)
                          +-----------v-----------+
                          |   Flask Server (app.py) |
                          +-----------+-----------+
                                      |
          +---------------------------v---------------------------+
          |                    World Engine (world.py)             |
          |  Tick loop: agents -> businesses -> settlements ->    |
          |  macro ODE -> geopolitics -> resources -> UI emit     |
          +---+----------+----------+----------+----------+------+
              |          |          |          |          |
    +---------v--+  +----v----+  +-v--------+ +v-------+ +v-----------+
    | Agents     |  | Macro   |  | Geo-     | | Bridge | | History    |
    | (agents.py)|  | (macro  |  | politics | | (bridge| | (history   |
    | JEPA world |  |  .py)   |  | (.py)    | |  .py)  | |  .py)      |
    | model,     |  | 14-state|  | Nations, | | Macro  | | Paleo-     |
    | traits,    |  | ODE:    |  | alliances| | <-> Agnt| | climate,   |
    | skills,    |  | CO2,    |  | trade,   | | <-> Geo | | migration, |
    | memory     |  | temp,   |  | conflict | |        | | Diamond,   |
    +---------+--+  | SLR,    |  | (IFs)    | +--------+ | Dawkins    |
              |     | tension |  +----------+             +------------+
    +---------v--+  +---------+
    | Shared     |
    | World Model|
    | (shared_   |
    |  world_    |
    |  model.py) |
    | Batch JEPA |
    +------------+

For step-by-step operational guidance— env vars, LLM setup, log analysis, troubleshooting — see[.]HowTo.md

• Python 3.11+
• ~17 MB disk for pre-computed Earth data

cd oikoumene-main
pip install numpy flask flask-socketio eventlet scipy networkx shapely requests
conda create -n oikoumene python=3.11 -y
conda activate oikoumene
pip install -r requirements.txt
python generate_landmask.py        # ~10s — rasterizes Natural Earth coastlines
python generate_earth_data.py      # ~1s  — climate zones, resources, biomes
python generate_present_day_data.py # ~15s — World Bank API, NOAA, NASA (needs internet)
python app.py

Agents begin as small bands in East Africa (~68,000 BCE). Over thousands of ticks they migrate through Arabia to Asia, Europe, Australia, and eventually the Americas via the Beringia land bridge. Agriculture emerges in the Fertile Crescent. Civilizations rise and fall. The Industrial Revolution triggers the macro ODE system (CO2, warming, resource depletion). The simulation continues into the future.

Time scale: 200 years/tick (Paleolithic) → 1 month/tick (Modern)

Initializes from real-world data (World Bank API, NOAA, NASA GISS):

• 300 agents distributed proportional to real population density
• 140 nations from World Bank economic indicators
• 10 active conflicts with geolocation (Ukraine, Gaza, Sudan, Myanmar...)
• CO2 = 427 ppm, temperature = +1.19°C (actual 2025 values)
• Macro ODE active from tick 0

Time scale: 1 month/tick

The simulation integrates research from seven distinct scientific domains. Every equation in the codebase cites its source.

Each agent perceives the world through a Joint Embedding Predictive Architecture (JEPA), as proposed by Yann LeCun.

Core papers:

• LeCun, Y. (2022). A Path Towards Autonomous Machine Intelligence. Position paper, Meta AI. — Sections 3.1-3.3: cognitive architecture with world model, cost module, actor, and configurator. - Maes, L., Le Lidec, Q., Scieur, D., LeCun, Y., & Balestriero, R. (2026). LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels. arXiv:2603.19312. — AdaLN action conditioning, SIGReg regularization principle, temporal path straightness metric. (Our v0.2 SIGReg implementation is a moments-based variant — see Implementation Notes below.) - Qu, H., Morel, M., McCabe, M., Bietti, A., Lanusse, F., Ho, S., & LeCun, Y. (2026). Representation Learning for Spatiotemporal Physical Systems. arXiv:2603.13227. — Linear probing of latent embeddings to test if physical parameters are captured.

Implementation (world_model.py

, shared_world_model.py

):

Component	Architecture	Reference
Encoder	3-layer MLP (obs → hidden → hidden → latent), RMSNorm + GELU	LeCun 2022, Section 3.1
Predictor	MLP with Adaptive Layer Normalization (AdaLN) — action conditions each layer's scale and shift; zero-init scale/shift weights (DiT-style)
Maes et al. 2026, Section 3.2; Peebles & Xie 2022
SIGReg (v0.2)	Differentiable moments-matching variant: skewness² + kurtosis² + variance penalty along random unit-norm projections, in the spirit of Cramer-Wold gaussianity testing	Adapted from Maes et al. 2026, Section 4
CEM Planner	Cross-Entropy Method: sample action sequences, rollout in latent space, select elites, refine	LeCun 2022, Section 3.4
Training	L = L_pred + λ · SIGReg(Z), analytic backpropagation (hand-implemented in NumPy, gradient-checked against central finite differences to <1e-8 in test_world_model_gradcheck.py ), Adam optimizer with gradient clipping at 5.0. An optional PyTorch backend (world_model_torch.py ) uses autograd.
LeCun 2022

Loss function:

L = ||z_hat_{t+1} - z_{t+1}||^2 + lambda * SIGReg(Z)

where z_hat_{t+1} = Predictor(Encoder(x_t), a_t)

and z_{t+1} = Encoder(x_{t+1})

.

Backends (NumPy default, PyTorch optional). The reference implementation in world_model.py

is pure NumPy with hand-written, gradient-checked backprop. An opt-in PyTorch backend (world_model_torch.py

) implements the identical architecture with autograd and optional CUDA. At its default settings it reproduces the NumPy model — weights copied across backends match encode/predict outputs to < 1e-4 — so it is a true drop-in, not a different model. It also adds opt-in paper-aligned toggles: an Epps–Pulley characteristic-function SIGReg (Maes et al. 2026) with λ = 0.1, and predictor dropout. Install with pip install -e ".[torch]"

; select at runtime in code (SharedWorldModel(backend="torch")

) or from the dashboard's JEPA tab (NumPy / PyTorch × Repo-default / Paper × device). Switching preserves the shared experience buffer; if PyTorch is not installed the option degrades gracefully and the NumPy backend keeps running.

Agent decision loop (Kahneman's Dual Process Theory):

System 1, symbolic(every tick): Maslow-style needs hierarchy weights eleven goal candidates by trait-modulated priorities (eat, heal, work, trade, build_business, socialize, reproduce, explore, research, govern, migrate); argmax selects the active goal.System 1, neural(JEPA, every PLAN_INTERVAL=3 ticks): observe → encode current state → encode goal-target observation → CEM plan in latent space → extract movement bias and intensity; cached between re-plans.System 2(LLM, optional, social actions only): trade negotiation, governance speech, social dialogue.

The global state evolves via a 14-variable ODE system inspired by the Club of Rome.

Core references:

• Meadows, D. H., Meadows, D. L., Randers, J., & Behrens, W. W. (1972). The Limits to Growth. Universe Books. — World3 model structure: population-resource-pollution feedback loops. - Meadows, D. H., Randers, J., & Meadows, D. L. (2004). Limits to Growth: The 30-Year Update. Chelsea Green. — Calibrated depletion rates. - Dixson-Decleve, S., Gaffney, O., Ghosh, J., Randers, J., Rockstrom, J., & Stoknes, P. E. (2022). Earth for All: A Survival Guide for Humanity. New Society Publishers. — Social tension model: f(inequality, food insecurity, environmental degradation). - Nordhaus, W. D. (2017). Revisiting the social cost of carbon. PNAS 114(7). — DICE model: GDP growth sector, climate damage function D = a*T^2.

Climate sub-model (two-layer energy balance):

dT/dt = (1/C) * [F(CO2) - lambda*T - gamma*(T - T_deep)]
F = 5.35 * ln(CO2/280)   [Myhre et al. 1998]

Parameter	Value	Source
Climate sensitivity (ECS)	3.0°C / 2xCO2	IPCC AR6 WG1, Table 7.SM.1
Ocean heat capacity C	7.0 W·yr/m²/°C	Held et al. (2010), lower end
Climate feedback λ	1.236 W/m²/°C	Calibrated so emergent ECS = F_2x / λ = 3.00°C exactly (v0.2 fix)
Deep ocean coupling γ	0.7 W/m²/°C	Gregory (2000)
CO2 forcing coefficient	5.35 W/m²	Myhre et al. (1998)
Natural CO2 absorption	50% of emissions (decadal mean)	Friedlingstein et al. (2024)
Base emission rate	42 GtCO2/yr	Friedlingstein et al. (2024)
ppm per GtCO2	0.128	IPCC AR6 WG1 Annex VII (= 1/2.13 GtC × 1/3.67)

Resource depletion follows Hubbert-style curves (Hubbert, 1956), not linear depletion. Technology provides a balancing loop (S-curve growth per Romer 1990).

Validated: BAU scenario 2025→2100 produces 679 ppm CO2, +2.74 °C, 0.61 m sea-level rise — sitting between IPCC AR6 SSP2-4.5 and SSP3-7.0 envelopes. Carbon-cycle calibration verified against the Mauna Loa observed growth rate (~2.5 ppm/yr at 2025 emissions). 9/9 IPCC validation checks plus 2 unit tests for ECS consistency and the carbon-cycle anchor.

In the historical scenario nations are not pre-defined — they emerge organically when agent settlements grow large enough and merge. The Present Day scenario instead seeds real-world nations (flagged seeded

) that persist as macro-actors and then evolve, absorbing nearby emergent settlements. Both share the same interstate dynamics, which follow established models.

Core references:

• Hughes, B. B. (2019). International Futures (IFs): Building and Using Global Models. Pardee Center, University of Denver. — Conflict probability model with calibrated logistic regression coefficients. - Liberal peace theory (Russett, 1993; Oneal & Russett, 1999): trade interdependence reduces interstate conflict probability.
• Bremer, S. A. (1992). Dangerous Dyads: Conditions Affecting the Likelihood of Interstate War, 1816–1965. Journal of Conflict Resolution 36(2), 309–341. — Power-parity effect. - Pettersson, T. (2024). UCDP/PRIO Armed Conflict Dataset Codebook v24.1. Uppsala Conflict Data Program. — Empirical conflict-duration anchor. - Tinbergen, J. (1962). Shaping the World Economy. — Gravity model of trade.

Conflict probability (per nation-dyad per macro tick):

P(conflict) = sigmoid(
    β₀                                  # base rate (v0.2: −7.5)
    + β₁ · resource_competition         # scarce resources → conflict
    + β₂ · power_parity                 # near-peer → more likely (Bremer)
    + β₃ · (1 − trade_interdependence)  # liberal peace
    + β₄ · social_tension               # Earth4All link (v0.2: 1.5)
    + β₅ · territorial_overlap          # border proximity
    − β₆ · shared_alliances             # mutual allies → peace
    − β₇ · diplomatic_history           # positive history
)

Active-conflict intensity decays at 0.80/tick (~2.6-year half-life at the default 10-month macro tick), consistent with the UCDP/PRIO median armed conflict duration of ~3 years (Pettersson 2024).

Calibration (v0.2): tuned against UCDP-style active-conflict prevalence in a 5-nation neighbour cluster: ~10–25% / 30–50% / 50–80% prevalence at low / mid / high social tension. v0.1 produced ~99% / 100% / 100% prevalence because the previous decay (0.95) gave an effective conflict lifetime of ~38 years; the v0.2 calibration is a coupled re-tuning of decay, lifetime cap, base rate, and tension coefficient. Distance calculations use haversine (degree-equivalents) to preserve threshold semantics while correcting the polar distortion of euclidean lat/lng.

The simulation runs on our planet's surface derived from multiple datasets.

Data sources:

• Natural Earth (naturalearthdata.com) — 110m land polygons, rivers, lakes. Rasterized to 0.25° land mask (720x1440) via Shapely point-in-polygon.
• Climate zones classified via Whittaker biome diagram(Whittaker 1975): temperature x precipitation → 12 biome types.

Climate model components (in generate_earth_data.py

):

Layer	Method	References
Temperature	Latitude + elevation lapse rate (-6.5°C/km) + continentality + ocean currents (Gulf Stream, Kuroshio, Humboldt, Benguela)	Hartmann (2016) Global Physical Climatology; Peixoto & Oort (1992)
Precipitation	ITCZ + Hadley cell subsidence + mid-latitude storm tracks + monsoon regions + orographic effects	Schneider et al. (2014); Hoskins & Valdes (1990); Roe (2005)
Soil fertility	FAO GAEZ methodology: biome + precipitation + temperature + known breadbaskets	Licker et al. (2010); Mueller et al. (2012); Schlenker & Roberts (2009)
Mineral deposits	Tectonic/orogenic belts + known provinces	USGS; Marshak (2019); Arndt et al. (2017); Sillitoe (2010)
Freshwater	Precipitation + Natural Earth rivers/lakes + known aquifer regions	Doll et al. (2003); Vorosmarty et al. (2010); Schewe et al. (2014)
Fossil fuels	Known sedimentary basin locations	USGS World Petroleum Assessment; BGR (2019)

The historical scenario models 70,000 years of climate oscillation.

Data sources:

• EPICA Community Members (2004). Eight glacial cycles from an Antarctic ice core. Nature429, 623-628. — CO2 record for 800 kyr. - Petit, J. R., et al. (1999). Climate and atmospheric history from the Vostok ice core. Nature399, 429-436. — Temperature record for 420 kyr. - Jouzel, J., et al. (2007). Orbital and millennial Antarctic climate variability over the past 800,000 years. Science317, 793-796. - Marcott, S. A., et al. (2013). A reconstruction of regional and global temperature for the past 11,300 years. Science339, 1198-1201. - Spratt, R. M., & Lisiecki, L. E. (2016). A Late Pleistocene sea level stack. Climate of the Past12, 1079-1092. - Clark, P. U., et al. (2009). The Last Glacial Maximum. Science325, 710-714. — LGM ice sheet reconstructions. - Stringer, C. (2012). The Origin of Our Species. Penguin. — Anatomically modern human dispersal timeline (Out of Africa, ~70 kya); informs Scenario A initial conditions.

Key events modeled:

Event	Year (BP)	CO2 (ppm)	Temp (°C)	Sea Level (m)
Out of Africa	70,000	200	-6.0	-80
Last Glacial Maximum	21,000	185	-8.0	-130
Younger Dryas	12,000	235	-5.0	-65
Holocene Optimum	6,000	270	+0.5	-5
Pre-industrial	200	280	-0.3	0

Agents experience different agricultural potential, disease resistance, and technology diffusion rates based on their geographic location.

Core reference:

• Diamond, J. (1997). Guns, Germs, and Steel: The Fates of Human Societies. W. W. Norton.

Implementation (history.py: GeographicAdvantage

):

Factor	Eurasia	Africa	Americas	Source
Continental axis multiplier	1.5x (E-W)	0.7x (N-S)	0.5-0.6x (N-S)	Diamond Ch. 10
Domesticable large mammals	4 (sheep, goat, cattle, pig)	0	1 (llama)	Diamond Ch. 9
Founder crops	8 (wheat, barley, lentils...)	3 (sorghum, millet, cowpea)	3 (maize, squash, beans)	Diamond Ch. 8
Disease resistance (from animal proximity)	High (zoonotic exposure)	Low	Very low	Diamond Ch. 11

Agricultural origins: Fertile Crescent (11,500 BP), Yellow River (10,000 BP), Yangtze (9,000 BP), Mesoamerica (9,000 BP), Andes (8,000 BP). Diffusion modeled at ~1 km/year along latitude, slower across climate barriers.

Over many generations, agent populations accumulate adaptations to local environments (cold tolerance, altitude adaptation, disease resistance).

Core reference:

• Dawkins, R. (2009). The Greatest Show on Earth: The Evidence for Evolution. Transworld Publishers.

Implementation: Trait inheritance via crossover + mutation (mutation rate 15%). Environmental selection pressure: agents better adapted to local temperature, altitude, and disease environment have higher survival and reproduction rates.

Module	Lines	Purpose
agents.py
1,208	Autonomous agents: JEPA cognition, physics, traits, skills, memory, social actions
world.py
1,189	World engine: tick loop, resources, businesses, settlements, scenario dispatch, era-aware UI summaries, runtime JEPA backend swap
world_model.py
701	JEPA implementation (NumPy, hand-written backprop): encoder, predictor (AdaLN), SIGReg, CEM planner, deterministic batch sampling
world_model_torch.py
534	Optional PyTorch JEPA backend (autograd): same architecture, CUDA-ready, Epps–Pulley SIGReg toggle, NumPy weight bridge
shared_world_model.py
263	Single shared JEPA for all agents with batch encode/plan; selects NumPy or PyTorch backend
macro.py
512	14-state ODE: climate, resources, pollution, socioeconomics
geopolitics.py
705	Emergent nations, alliances, trade (gravity model), conflict (IFs)
bridge.py
456	Bidirectional coupling: agents <-> macro <-> geopolitics; per-cell regen baselines
history.py
849	70,000-year timeline: paleoclimate, migration, Diamond, Dawkins

Module	Lines	Purpose
llm_module.py
713	LLM social cognition (Ollama/OpenAI): trade negotiation, governance speech, social dialogue
god_mode.py
450	Interventional experiments: whisper, commandment, drought, plague, climate nudge
scenarios.py
334	Scenario A (historical) and B (present-day) configuration
earth.py
478	Real geography: Natural Earth land mask, Whittaker biomes, resource lookup

Script	Purpose
generate_landmask.py
Rasterize Natural Earth 110m polygons to 0.25° land mask
generate_earth_data.py
Compute 9 Earth system grids at 0.5° (temp, precip, biome, fertility, minerals, freshwater, fossil)
generate_present_day_data.py
Fetch World Bank API + NOAA + NASA data for Scenario B

Module	Purpose
agent_state.py
Structure-of-Arrays storage + cKDTree — benchmarked at 173 tps (2000 agents)

Test	Validates	Count
test_macro.py
BAU 2025–2100 vs. IPCC AR6 SSP2-4.5/SSP3-7.0 envelope; carbon-cycle vs. Mauna Loa decadal mean; ECS-consistency unit test	9 + 2
test_world_model.py
JEPA training: prediction-loss reduction, learned action-conditioning, anti-collapse, linear probe R², CEM planner output validity	5
test_world_model_gradcheck.py
Backward implementations (linear, GELU, RMSNorm, AdaLN, SIGReg) verified against central finite differences (measured relative error <1e-8)	5
test_shared_world_model.py
Single vs. batch equivalence (max diff 1e-15), per-agent vs. plan_batch identity, edge cases	6
test_world_model_torch.py
PyTorch backend (opt-in): single/batch parity, NumPy↔Torch weight cross-check (<1e-4), Epps–Pulley toggle, device handling (skips if torch absent)	9 (+1 CUDA-gated)
test_agents_lifecycle.py
Era-aware lifecycle thresholds across 4 eras, modern drift bounds, paleolithic 1-tick floor	7
test_geopolitics.py
Haversine correctness, conflict monotonicity, 5-nation BAU prevalence calibration, summit-cadence independence	5
test_world.py
Haversine threshold semantics, snapshot iteration safety	4
test_bridge.py
Behavioural identity of optimised lookups (410-agent run, 0 diffs), 6.2× hot-path speedup, edge cases	6
test_llm_module.py
Fallback mode, JSON parsing, rate limiting	9
test_agent_state.py
SoA operations, KDTree, batch metabolism + benchmarks	4
test_security.py
API-key/base_url binding (exfiltration), SSRF/host allowlist, input clamping, escapeHtml on every innerHTML sink
12
test_simcore_fixes.py
Tick-loop generation token, Present-Day nation persistence, seeded-RNG reproducibility, logger lifecycle, macro handoff continuity, drought/regen composition, bounded growth, agent fixes	24

File	Resolution	Source	Size
landmask.npy
0.25° (720x1440)	Natural Earth 110m + Shapely	1.0 MB
earth_terrain.npy
0.5° (360x720)	Whittaker biome diagram	253 KB
earth_temperature.npy
0.5°	Latitude + lapse rate + ocean currents	2.0 MB
earth_precipitation.npy
0.5°	ITCZ + Hadley + monsoon + orographic	2.0 MB
earth_biome.npy
0.5°	Whittaker: temp x precip -> 12 biomes	253 KB
earth_fertility.npy
0.5°	FAO GAEZ-inspired + breadbaskets	2.0 MB
earth_minerals.npy
0.5°	USGS provinces + tectonic belts	2.0 MB
earth_freshwater.npy
0.5°	Precipitation + rivers + aquifers	2.0 MB
earth_fossil_fuels.npy
0.5°	USGS petroleum basins	2.0 MB
ne_110m_land.geojson
110m	Natural Earth (public domain)	138 KB
ne_110m_rivers.geojson
110m	Natural Earth	38 KB
ne_110m_lakes.geojson
110m	Natural Earth	37 KB
present_day_*.json/npy
2° / country	World Bank API + NOAA + NASA	~5 MB

Agents	ms/tick	tps	Scenario
25	28	35.7	Historical (Out of Africa)
37	39	25.7	Historical (after 200 ticks)
300	328	3.0	Present Day

Shared JEPA World Model— 1 model for N agents (not N copies)** Tick-skipping**— CEM plan every 3 ticks, cached behavior between** cKDTree**— O(log N) spatial queries (was O(N) hash grid)** Vectorised macro/ice coupling**—apply_macro_to_world

and_apply_ice_age_effects

rewritten from per-cell Python loops to NumPy (removed periodic multi-100 ms stalls; verified numerically identical)Bounded entities— dead agents and settlements are pruned every tick, and nation statistics are aggregated in O(agents + memberships) rather than O(settlements × agents)

Note: the benchmark table above predates optimizations 4–5; re-run on your hardware for current numbers.

Component	2000 agents	Source
Physics + spatial	5.8 ms	agent_state.py
JEPA encode + plan	294 ms	shared_world_model.py
Total (projected)
~300 ms (3.3 tps)

The v0.1.0 release was generated primarily with Claude Code in roughly two weeks. A subsequent domain-review pass identified six categories of bugs that affected scientific correctness without breaking the runtime:

JEPA training did not actually train. v0.1 estimated gradients with 3 random search directions per weight matrix. For an encoder layer with 16 384 parameters this gave an effective signal-to-noise ratio of ~2 × 10⁻⁴, so the prediction loss decreased only on the bias terms and the AdaLN action-conditioning weights were never updated at all. v0.2 replaces this with hand-written analytic backpropagation in pure NumPy, verified against central finite differences to <1e-8 relative error (test_world_model_gradcheck.py

). On a synthetic toy problem with hidden physical parameters, prediction loss now decreases 103× and a linear probe recovers the hidden physics with R² = 0.98.Carbon-cycle unit conversion was off by a factor of 3.67 because the v0.1 code applied a GtCO₂→GtC division and then multiplied by a ppm/GtCO₂ constant, double-converting. The model produced ~0.8 ppm/yr vs. the Mauna Loa observed 2.5 ppm/yr. v0.2 fixes the conversion and verifies against NOAA GML decadal mean.Climate sensitivity was inconsistent— the declared 3.0 °C ECS constant was unused by the ODE; emergent ECS was 3.37 °C. v0.2 calibratesλ

so the emergent value matches the declaration exactly.Conflict prevalence saturated at ~99% in a 5-nation BAU run because conflicts decayed too slowly (38-year effective lifetime vs. UCDP median ~3 years). v0.2 re-calibrates decay, lifetime cap, and logit coefficients against UCDP prevalence targets.Lifecycle thresholds were not era-scaled. Hardcodedage > 40 ticks

reproduction threshold meant agents reproduced at 3.3 years in Modern era and never reached reproductive age in Paleolithic era. v0.2 parameterises in real-world years with runtime conversion.The macro/agent coupling layer had quartisch lookups in its hot path, costing ~50 ms/tick at 300 agents. v0.2 reduces to linear complexity (~8 ms/tick).

We document this honestly because we think the conclusion is interesting: LLM-generated code can produce scientifically-flavoured architectures faster than humans can write them, but the physical and empirical calibration requires domain expertise that LLMs (at least currently) do not reliably substitute for. Every fix in this list required a domain-grounded judgment call that the original generation pass got wrong despite confident-sounding code comments. The full v0.2 calibration pass is documented in CHANGELOG.md.

A same-day follow-up review pass on v0.2.0 surfaced five additional bugs in the integration glue between the (now correctly calibrated) scientific modules and the simulation loop, plus several UI-payload issues. The themes are different from v0.2.0: where v0.2.0 was about scientific constants and equation correctness, v0.2.1 is about coupling, timing, and presentation correctness. The main entries:

Regen-array ratchets in—bridge.py

water_regen

andminerals_regen

were multiplied by macro factors each tick with no baseline reset, underflowing to zero independently of macro state.food_regen

had a related but distinct bug: a two-factor terrain approximation (plains vs. all-else) that silently inflated mountain, desert, and tundra regen by 5×, 10×, 3.3× relative to the five-factorResourceMap.initialize_from_terrain

.Tech diffusion double-credit in—geopolitics.py

_diffuse_technology

iteratedtrade_graph.edges()

on aDiGraph

carrying both directions of every dyad, so each tick credited the lower-tech nation twice.Stale phantom trade edges— when a dyad's volume fell below the retention threshold, the previous tick's edge persisted with its old weight, feeding phantom values into liberal-peace, alliance-affinity, and tech-diffusion calculations.Paleo—_apply_ice_age_effects

ratchet + ice-retreat recoveryfood_regen *= cold_factor

compounded across tens of thousands of paleo ticks, and cells once covered by ice never recovered productivity when the ice retreated, inconsistent with the post-LGM recolonisation record. Fixed via per-cell baselines and a_was_iced

transition flag.Macro—dt_years

10× rate mismatchMacroModel(dt_years = 1/12)

was instantiated for the per-tick calibration test, butworld.step

invokes macro everymacro_update_interval = 10

world ticks. The ODE therefore integrated only one month per ten sim months, running at one-tenth of the calibrated rate. Fixed by settingdt_years = macro_update_interval / 12

. The standalonetest_macro.py

path is unaffected.

Plus several frontend issues: top-header / sidebar climate sources diverged in modern era; the right-sidebar Macro panel showed frozen 2025 values across the entire 70 000-yr history view; temperature was rendered as +${value}

(yielding +-5.13 °C

in paleo); sea-level always in cm (yielding -13 000 cm

for LGM); chart lines crossed through their own Max:

labels at peak values. v0.2.1 introduces an era-aware payload helper, paleodemographic population from McEvedy & Jones (1978) / Biraben (2003) / HYDE 3.1 (Klein Goldewijk et al. 2010) for the paleo panel, and small format helpers (fmtSigned

, fmtSeaLevel

).

The pattern echoes v0.2.0: even after a calibration pass that fixed the "science layer", a second pass at the coupling and presentation layers still found real bugs that affect what a reader of the simulation output would see. We document this not to claim every bug has now been found — it almost certainly hasn't — but to be honest about the cost of auditing LLM-generated code.

Where multiple plausible parameter sets exist, we anchor against observation rather than to round-number defaults:

• Carbon cycle to Mauna Loa decadal mean (NOAA GML 2014–2024).
• Climate physics to IPCC AR6 SSP-envelope projections and Held et al. two-layer energy balance.
• Conflict to UCDP/PRIO prevalence in regional clusters (the equivalent of "all 5 nations are neighbours on one continent").
• Agent lifecycle to anthropological / demographic ranges (15-year reproduction, 80-year lifespan, 60-year senescence onset).

Every calibration choice is verified by a test that would catch a regression from a future refactor.

• LeCun, Y. (2022). A Path Towards Autonomous Machine Intelligence. Meta AI Position Paper. - Maes, L., Le Lidec, Q., Scieur, D., LeCun, Y., & Balestriero, R. (2026). LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels. arXiv:2603.19312. - Qu, H., Morel, M., McCabe, M., Bietti, A., Lanusse, F., Ho, S., & LeCun, Y. (2026). Representation Learning for Spatiotemporal Physical Systems. arXiv:2603.13227. - Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.

• IPCC (2021). Climate Change 2021: The Physical Science Basis(AR6 WG1). Cambridge University Press. - Myhre, G., et al. (1998). New estimates of radiative forcing due to well mixed greenhouse gases. Geophysical Research Letters25(14), 2715-2718. - Held, I. M., et al. (2010). Probing the fast and slow components of global warming. Journal of Climate23(9), 2418-2427. - Gregory, J. M. (2000). Vertical heat transports in the ocean and their effect on time-dependent climate change. Climate Dynamics16, 501-515. - Friedlingstein, P., et al. (2024). Global Carbon Budget 2024. Earth System Science Data.

• EPICA Community Members (2004). Eight glacial cycles from an Antarctic ice core. Nature429, 623-628. - Petit, J. R., et al. (1999). Climate and atmospheric history from the Vostok ice core. Nature399, 429-436. - Lüthi, D., et al. (2008). High-resolution carbon dioxide concentration record 650 000–800 000 years before present. Nature453, 379-382. - Jouzel, J., et al. (2007). Orbital and millennial Antarctic climate variability. Science317, 793-796. - Marcott, S. A., et al. (2013). A reconstruction of regional and global temperature for the past 11,300 years. Science339, 1198-1201. - Spratt, R. M., & Lisiecki, L. E. (2016). A Late Pleistocene sea level stack. Climate of the Past12, 1079-1092. - Clark, P. U., et al. (2009). The Last Glacial Maximum. Science325, 710-714. - Adams, J. M., & Faure, H. (1998). A new estimate of changing carbon storage on land since the last glacial maximum. Global and Planetary Change16-17, 3-24. - Crowley, T. J., & Baum, S. K. (1997). Effect of vegetation on an ice-age climate model simulation. Journal of Geophysical Research102, 16463-16480.

• McEvedy, C., & Jones, R. (1978). Atlas of World Population History. Penguin. - Biraben, J.-N. (2003). An essay concerning mankind's evolution. Population & Societies394, 1-4. - Klein Goldewijk, K., Beusen, A., van Drecht, G., & de Vos, M. (2010). HYDE 3.1: Long-term dynamic modelling of global population and built-up area. The Holocene20, 565-573.

• Meadows, D. H., et al. (1972). The Limits to Growth. Universe Books. - Meadows, D. H., Randers, J., & Meadows, D. L. (2004). Limits to Growth: The 30-Year Update. Chelsea Green. - Dixson-Decleve, S., Gaffney, O., Ghosh, J., Randers, J., Rockstrom, J., & Stoknes, P. E. (2022). Earth for All: A Survival Guide for Humanity. New Society Publishers. - Piketty, T. (2014). Capital in the Twenty-First Century. Harvard University Press. - Nordhaus, W. D. (2017). Revisiting the social cost of carbon. PNAS114(7), 1518-1523. - Romer, P. M. (1990). Endogenous technological change. Journal of Political Economy98(5), S71-S102. - Hubbert, M. K. (1956). Nuclear energy and the fossil fuels. Shell Development Company Publication95.

• Diamond, J. (1997). Guns, Germs, and Steel: The Fates of Human Societies. W. W. Norton. - Dawkins, R. (2009). The Greatest Show on Earth: The Evidence for Evolution. Transworld Publishers. - Stringer, C. (2012). The Origin of Our Species. Penguin. - Marshak, S. (2019). Earth: Portrait of a Planet(6th ed.). W. W. Norton. ISBN 978-0393640137. - Whittaker, R. H. (1975). Communities and Ecosystems(2nd ed.). Macmillan.

• Hughes, B. B. (2019). International Futures (IFs): Building and Using Global Models. Elsevier Academic Press. - Russett, B. (1993). Grasping the Democratic Peace. Princeton University Press. - Oneal, J. R., & Russett, B. (1999). The Kantian peace: The pacific benefits of democracy, interdependence, and international organizations, 1885–1992. World Politics52(1), 1–37. - Tinbergen, J. (1962). Shaping the World Economy. Twentieth Century Fund. - Bremer, S. A. (1992). Dangerous dyads: Conditions affecting the likelihood of interstate war, 1816–1965. Journal of Conflict Resolution36(2), 309–341. - Pettersson, T. (2024). UCDP/PRIO Armed Conflict Dataset Codebook v24.1. Uppsala Conflict Data Program.https://ucdp.uu.se/ - Homer-Dixon, T. F. (1999). Environment, Scarcity, and Violence. Princeton University Press. - Leeds, B. A. (2003). Alliance reliability in times of war: Explaining state decisions to violate treaties. International Organization57(4), 801–827.

• Licker, R., et al. (2010). Mind the gap: how do climate and agricultural management explain the 'yield gap' of croplands? Global Ecology and Biogeography19(6), 769-782. - Mueller, N. D., et al. (2012). Closing yield gaps through nutrient and water management. Nature490, 254-257. - Arndt, N. T., et al. (2017). Future global mineral resources. Geochemical Perspectives6(1), 1-171. - Sillitoe, R. H. (2010). Porphyry copper systems. Economic Geology105(1), 3-41. - Doll, P., et al. (2003). A global hydrological model for deriving water availability indicators. Journal of Hydrology270(1-2), 105-134. - Vorosmarty, C. J., et al. (2010). Global threats to human water security and river biodiversity. Nature467, 555-561. - BGR (2019). Energy Study 2019: Data and Developments Concerning German and Global Energy Supplies. Bundesanstalt fur Geowissenschaften und Rohstoffe.

• Hartmann, D. L. (2016). Global Physical Climatology(2nd ed.). Elsevier. - Schneider, T., et al. (2014). Migrations and dynamics of the ITCZ. Nature513, 45-53. - Hoskins, B. J., & Valdes, P. J. (1990). On the existence of storm-tracks. Journal of the Atmospheric Sciences47(15), 1854-1864. - Roe, G. H. (2005). Orographic precipitation. Annual Review of Earth and Planetary Sciences33, 645-671. - Schlenker, W., & Roberts, M. J. (2009). Nonlinear temperature effects indicate severe damages to U.S. crop yields. PNAS106(37), 15594-15598. - Schewe, J., et al. (2014). Multimodel assessment of water scarcity under climate change. PNAS111(9), 3245-3250.

• NOAA Global Monitoring Laboratory — Mauna Loa CO2 record. https://gml.noaa.gov/ - NASA GISS Surface Temperature Analysis (GISTEMP v4). https://data.giss.nasa.gov/gistemp/ - World Bank Open Data. https://data.worldbank.org/ - Natural Earth. https://www.naturalearthdata.com/(public domain). - USGS Mineral Commodity Summaries. https://www.usgs.gov/ - Peixoto, J. P., & Oort, A. H. (1992). Physics of Climate. AIP Press.

• Pineau, J., et al. (2019). Improving Reproducibility in Machine Learning Research (a Report from the NeurIPS 2019 Reproducibility Program). arXiv:1906.06337.

AGPL-3.0-or-later

Dr. Gerrit Tombrink, GeoLambda GmbH

source & further reading

github.com — original article