Oikoumene: Autonomous Agent Civilization Simulator

GeoLambda GmbH released Oikoumene v0.3.1, an autonomous agent civilization simulator that models human history from 70,000 years ago to 2100 using JEPA world models and a Maslow-style needs hierarchy. The simulation, built primarily with Anthropic's Claude Code, features emergent settlements, nations, trade networks, and conflicts on a real climate data-driven planet surface.

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 downloaders and their tests are implemented, but the runtimeingestion is not yet wired the simulation still runs on the synthetic data/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. The Built 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 or python3.11 -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 in app.py .Windows: untested, please file issues. /GeoLambdaAI/oikoumene/blob/main/static/oikoumene.jpg 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 Open http://localhost:5000 Select scenario → Click Start 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. Nature 429, 623-628. — CO2 record for 800 kyr. - Petit, J. R., et al. 1999 . Climate and atmospheric history from the Vostok ice core. Nature 399, 429-436. — Temperature record for 420 kyr. - Jouzel, J., et al. 2007 . Orbital and millennial Antarctic climate variability over the past 800,000 years. Science 317, 793-796. - Marcott, S. A., et al. 2013 . A reconstruction of regional and global temperature for the past 11,300 years. Science 339, 1198-1201. - Spratt, R. M., & Lisiecki, L. E. 2016 . A Late Pleistocene sea level stack. Climate of the Past 12, 1079-1092. - Clark, P. U., et al. 2009 . The Last Glacial Maximum. Science 325, 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 calibrates λ 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. Hardcoded age 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 /GeoLambdaAI/oikoumene/blob/main/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 and minerals 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-factor ResourceMap.initialize from terrain . Tech diffusion double-credit in — geopolitics.py diffuse technology iterated trade graph.edges on a DiGraph 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 recovery food 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 mismatch MacroModel dt years = 1/12 was instantiated for the per-tick calibration test, but world.step invokes macro every macro 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 setting dt years = macro update interval / 12 . The standalone test 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 Letters 25 14 , 2715-2718. - Held, I. M., et al. 2010 . Probing the fast and slow components of global warming. Journal of Climate 23 9 , 2418-2427. - Gregory, J. M. 2000 . Vertical heat transports in the ocean and their effect on time-dependent climate change. Climate Dynamics 16, 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. Nature 429, 623-628. - Petit, J. R., et al. 1999 . Climate and atmospheric history from the Vostok ice core. Nature 399, 429-436. - Lüthi, D., et al. 2008 . High-resolution carbon dioxide concentration record 650 000–800 000 years before present. Nature 453, 379-382. - Jouzel, J., et al. 2007 . Orbital and millennial Antarctic climate variability. Science 317, 793-796. - Marcott, S. A., et al. 2013 . A reconstruction of regional and global temperature for the past 11,300 years. Science 339, 1198-1201. - Spratt, R. M., & Lisiecki, L. E. 2016 . A Late Pleistocene sea level stack. Climate of the Past 12, 1079-1092. - Clark, P. U., et al. 2009 . The Last Glacial Maximum. Science 325, 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 Change 16-17, 3-24. - Crowley, T. J., & Baum, S. K. 1997 . Effect of vegetation on an ice-age climate model simulation. Journal of Geophysical Research 102, 16463-16480. - McEvedy, C., & Jones, R. 1978 . Atlas of World Population History . Penguin. - Biraben, J.-N. 2003 . An essay concerning mankind's evolution. Population & Societies 394, 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 Holocene 20, 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. PNAS 114 7 , 1518-1523. - Romer, P. M. 1990 . Endogenous technological change. Journal of Political Economy 98 5 , S71-S102. - Hubbert, M. K. 1956 . Nuclear energy and the fossil fuels. Shell Development Company Publication 95. - 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 Politics 52 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 Resolution 36 2 , 309–341. - Pettersson, T. 2024 . UCDP/PRIO Armed Conflict Dataset Codebook v24.1 . Uppsala Conflict Data Program. https://ucdp.uu.se/ 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 Organization 57 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 Biogeography 19 6 , 769-782. - Mueller, N. D., et al. 2012 . Closing yield gaps through nutrient and water management. Nature 490, 254-257. - Arndt, N. T., et al. 2017 . Future global mineral resources. Geochemical Perspectives 6 1 , 1-171. - Sillitoe, R. H. 2010 . Porphyry copper systems. Economic Geology 105 1 , 3-41. - Doll, P., et al. 2003 . A global hydrological model for deriving water availability indicators. Journal of Hydrology 270 1-2 , 105-134. - Vorosmarty, C. J., et al. 2010 . Global threats to human water security and river biodiversity. Nature 467, 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. Nature 513, 45-53. - Hoskins, B. J., & Valdes, P. J. 1990 . On the existence of storm-tracks. Journal of the Atmospheric Sciences 47 15 , 1854-1864. - Roe, G. H. 2005 . Orographic precipitation. Annual Review of Earth and Planetary Sciences 33, 645-671. - Schlenker, W., & Roberts, M. J. 2009 . Nonlinear temperature effects indicate severe damages to U.S. crop yields. PNAS 106 37 , 15594-15598. - Schewe, J., et al. 2014 . Multimodel assessment of water scarcity under climate change. PNAS 111 9 , 3245-3250. - NOAA Global Monitoring Laboratory — Mauna Loa CO2 record. https://gml.noaa.gov/ https://gml.noaa.gov/ - NASA GISS Surface Temperature Analysis GISTEMP v4 . https://data.giss.nasa.gov/gistemp/ https://data.giss.nasa.gov/gistemp/ - World Bank Open Data. https://data.worldbank.org/ https://data.worldbank.org/ - Natural Earth. https://www.naturalearthdata.com/ https://www.naturalearthdata.com/ public domain . - USGS Mineral Commodity Summaries. https://www.usgs.gov/ 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