Earth System Model & Human Collapse Assessment Tool

Calibration sources (2026): atmospheric CO₂ ≈ 428 ppm (Scripps/NOAA Mauna Loa — the Keeling Curve; Met Office forecasts 429.4 ppm as the 2026 annual mean); global temperature anomaly ≈ +1.5°C vs pre-industrial; ocean surface pH down to ≈ 8.05 from a pre-industrial ≈ 8.2; FAO 2025 finds 35.5% of marine fish stocks are overfished, with overfishing prevalence rising ~1 pt/yr; global vertebrate populations (Living Planet Index) have declined roughly 70–75% on average since 1970. Model logic: fossil + land-use emissions raise CO₂ → radiative forcing (log of CO₂ doubling) raises temperature with ocean-lag → warming lowers albedo (ice/snow loss), which feeds back into further warming → CO₂ uptake lowers ocean pH → warming + acidification bleach coral and shrink fish carrying capacity → deforestation and farming intensity erode topsoil → temperature (bell-curve response), CO₂ fertilization and topsoil jointly set crop yield → crop yield and fish stock set the food-limited carrying capacity for population → population and its energy demand set next year's emissions, closing the loop.

Permafrost feedback: a vulnerable carbon pool of 200 GtC (of the ~1,500 GtC total northern permafrost stock; IPCC AR6 cites 14–175 GtCO₂ released per 1°C of warming, other syntheses ~150–250 GtCO₂e by 2100) thaws each year at a rate set by a Stefan-equation-style relationship: an active-layer "melt depth" that grows with the square root of cumulative thawing-degree exposure, plus an extra term for the current rate of warming (faster warming outpaces the lagged thermal response and drives deeper, quicker thaw). Released carbon splits 97.7% CO₂ / 2.3% CH₄ by mass (Schuur et al. 2013). The CO₂ portion feeds directly into atmospheric CO₂ alongside fossil and deforestation emissions. The CH₄ portion accumulates in an atmospheric methane pool that decays at ~8.1%/yr (an 11.8-year lifetime) and adds its own warming via a simplified Myhre et al. (1998) radiative-forcing formula, converted to °C through the model's climate sensitivity — so a burst of methane warms fast and fades, while the CO₂ it's mixed with lingers. This closes a second loop: warming → deeper/faster permafrost thaw → CO₂ + CH₄ release → more warming.

RCP emissions pathways — an either-or alternative to the bottom-up fossil model: the radio buttons above let fossil emissions follow one of the four standard IPCC Representative Concentration Pathways instead of being derived from the Jevons/EROEI/reserve-depletion mechanics — selecting an RCP overrides Jevons rebound, renewable growth's effect on emissions, reserve depletion, and new discoveries entirely for the fossil-emissions number specifically (those controls visibly grey out when an RCP is selected). Each pathway's fossil-emissions trajectory is built from its well-documented peak-timing, peak-magnitude, and 2100-level features (RCP2.6: peaks ~2020-2050 then declines toward near-elimination; RCP4.5: peaks ~2040 then roughly halves by 2100; RCP6.0: peaks later, ~2080, before declining; RCP8.5: keeps rising through the century) — this is a piecewise-linear approximation informed by those documented features (van Vuuren et al. 2011 and companion RCP papers), not a reproduction of the full official Meinshausen et al. emissions database, which isn't loaded here.

Deforestation, permafrost, and every other feedback keep running independently — exactly as requested, only the fossil-fuel number is overridden. This has an important, honest consequence visible directly on the CO₂ chart's dashed reference line (the pathway's official published concentration, plotted for comparison only — it is not what the model's physics is forced to match): this model's own resulting CO₂ and temperature routinely run higher than the official RCP concentration curve, even when fossil emissions are following that RCP's trajectory exactly. Tested directly under RCP2.6 with the default deforestation rate: the tropical dieback tipping point still gets crossed by the 2050s-2070s from warming driven by permafrost, ice-albedo, and the model's other feedbacks — once tipped, self-reinforcing forest loss keeps adding CO₂ regardless of how clean the prescribed fossil trajectory is. This is not a bug; it's the direct, disclosed consequence of "all other feedbacks remain active": achieving RCP2.6-like outcomes in this model requires also controlling deforestation and other land-use drivers, not fossil emissions alone — a genuinely realistic lesson, since the original 2011-era RCP concentration curves were paired with simpler reference carbon-cycle assumptions than the permafrost, tipping-point, ice-albedo, cloud, and water-vapor physics this model has since accumulated.

Latent heat of fusion: why the Arctic Ocean stays cool until the ice is gone. Always active (unlike the experimental clathrate mechanism below), because this is well-established, textbook physics rather than a debated hypothesis. As long as sea ice is present, incoming heat that would otherwise raise ocean/atmosphere temperature instead goes into the phase change itself — melting ice — rather than showing up as a temperature rise, using the standard latent heat of fusion for water/ice (L_f = 334,000 J/kg). This is genuinely distinct from the ice-albedo feedback already modeled elsewhere: albedo governs how much solar energy is absorbed in the first place, while latent heat governs what happens to that energy once absorbed — whether it melts more ice or raises temperature. The buffering capacity is calculated directly from remaining September ice extent, representative ice thickness (~1.5m, consistent with PIOMAS/CryoSat volume estimates for recent Arctic ice), sea ice density (~900 kg/m³), and the Arctic Ocean's area (~14 million km²) — a genuine, physically-derived number in W/m², not an arbitrary one (currently ~4.4 W/m² of local Arctic buffering at today's ice extent). As ice declines, this buffering capacity shrinks in direct proportion (mass, and therefore the energy sink, scales with remaining ice extent), expressed as a smooth increase in the effective local Arctic amplification factor — referenced against today's already-observed 2.5x amplification (so there is no discontinuity: today's partial buffering is already baked into that empirically-cited figure), rising toward a higher ceiling (3.3x) once the buffer is fully exhausted. Tested directly: the transition is smooth and threshold-free, exactly tracking physical ice mass rather than jumping at any single "sq km" cutoff — checked across a full century-long run with no discontinuities or instability. This feeds directly into the methane clathrate mechanism below (the shallow Arctic trigger uses this latent-heat-adjusted amplification, not a fixed constant), and is deliberately left disconnected from the separate, longer-term boreal crop-zone amplification factor elsewhere, which represents an average climatological amplification rather than the more acute, ice-buffering-sensitive dynamic relevant here.

Jet stream weakening and AMOC: two different confidence levels, treated honestly as different. Losing the latent-heat buffer doesn't stop at the Arctic — it erodes the temperature gradient that drives the midlatitude jet stream, since the jet's strength is governed by the thermal-wind relationship (wind shear driven directly by horizontal temperature gradient, not by either region's absolute temperature). This model deliberately separates two claims with genuinely different scientific standing. The physical premise — latent heat keeps the Arctic relatively cool, maintaining the gradient; losing that buffer erodes it — is settled thermodynamics, calculated directly from the same physically-derived amplification factor above (referenced against today's gradient, so there's no discontinuity). The consequence — that a weaker gradient produces a slower, wavier jet stream favoring "blocking" patterns behind prolonged heatwaves, droughts, and cold-air outbreaks (Francis & Vavrus 2012; Coumou et al. 2018 on quasi-resonant amplification) — is real, published research, but remains an actively contested area, with some studies finding a more nuanced or weaker linkage than originally proposed. That's why its modeled consequence — an extreme-weather volatility penalty on crop yields, on top of the smooth mean-climate response already captured elsewhere — is kept deliberately modest (up to 15% for wheat, only 5% for rice, since wheat is temperate-zone-dominant and therefore far more exposed to midlatitude jet-stream dynamics than tropical-dominant rice) rather than treated with the same confidence as the underlying thermodynamics.

AMOC (the Atlantic Meridional Overturning Circulation) gets a smaller, more cautious treatment, matching the "possibly, but not as much" premise directly. AMOC is driven primarily by density differences in the North Atlantic — freshwater dilution weakens the sinking that drives it, and Greenland ice sheet melt (now tracked explicitly below, not merely implied) is the dominant real-world freshwater source for this system, modeled here as the primary driver. Arctic sea ice loss and general warming contribute smaller, secondary signals to the same system, deliberately scaled down relative to the jet-stream effect. Caesar et al. (2018) find observational evidence of roughly 15% AMOC weakening since the mid-20th century; IPCC AR6 assesses further weakening this century as very likely, with an abrupt collapse considered unlikely but a genuine, actively-debated tail risk — Ditlevsen & Ditlevsen's 2023 suggestion of possible mid-century collapse under high emissions is a contested claim, not the mainstream position, and this model does not adopt it as default behavior: AMOC weakens by up to 40 percentage points under sustained severe warming (tested directly to a floor around 61% even in an extreme 200-year scenario), never to full collapse. AMOC strength is tracked as a genuine state variable with a realistic multi-decadal lag (~50 years, reflecting that this is a large, slow ocean circulation system, not something that responds within a few years), and feeds back into the model concretely: weaker overturning means less surface water carrying dissolved CO2 is carried to ocean depth (the "solubility pump"), directly reducing the ocean carbon sink calculated elsewhere — independent of, and on top of, the acidification and warming effects already modeled there. This closes a genuine, multi-directional loop: ice loss weakens the latent-heat buffer, which weakens the jet stream (hurting crop yields) and modestly weakens AMOC (hurting ocean carbon uptake, adding CO2, adding warming, adding further ice loss) — tested directly to confirm all of this remains stable and bounded, not just superficially wired together, across both moderate and deliberately extreme multi-century scenarios.

Greenland ice sheet: its own mass balance, latent heat, freshwater runoff, and rising seas. A genuinely different ice population from Arctic sea ice — this is land ice, roughly 2.9 million km³ (~2,900,000 Gt), holding enough water to raise global sea level by ~7.4m if fully melted (IPCC AR5/AR6). Unlike floating sea ice, melting Greenland ice adds new water volume to the ocean rather than merely converting already-displacing ice to water, which is exactly why it (unlike Arctic sea ice) matters for sea level. Current mass loss is modeled at roughly 270 Gt/yr, matching IMBIE consortium satellite- gravimetry estimates, contributing today's real ~0.7-0.8mm/yr to global sea level. Melt rate accelerates with warming — observed melt roughly quintupled from ~50 Gt/yr in the 1990s to ~270 Gt/yr now, for well under 1°C of additional global warming over that period, a genuinely super-linear historical sensitivity — modeled here with a moderated fit to that acceleration (21st-century projections carry wide uncertainty in the literature, roughly a 6-fold range across IPCC AR6 scenarios and ice-sheet models, so this is an illustrative fit, not a reproduction of any single published projection). Tested directly to a deliberately extreme 200-year scenario: even reaching +25°C of warming, Greenland retains more than half its mass, correctly reflecting that this is a genuinely enormous ice reservoir that doesn't vanish on short timescales even under severe stress.

Local latent heat uses the identical physics already established for Arctic sea ice (L_f = 334,000 J/kg), applied to Greenland's own actively-melting mass each quarter and expressed as an ice-sheet-area-averaged W/m² buffering rate — shown on its own chart as a real, physically-scaled quantity illustrating how much energy this melt is absorbing, consistent with the same principle that keeps the Arctic Ocean cool while sea ice persists.

Freshwater runoff feeding AMOC is the connection given the most weight in this update, since it's the best-established of the three requested links: Greenland's melt rate is normalized against the freshwater flux magnitude ("hosing") climate models typically use to explore substantial AMOC responses (~0.1 Sv ≈ 3,150 Gt/yr) — today's ~270 Gt/yr is a small fraction of that, correctly implying only a modest AMOC contribution at current rates, growing as melt accelerates toward and beyond that reference scale.

The jet stream link — "if any," addressed directly: this is real but considerably more speculative than the Arctic mechanism, and is treated with a correspondingly smaller, explicitly secondary weight (25% of the effect size of the Arctic term, at maximum AMOC weakening). A weakened AMOC is associated with a persistent cool "warming hole"/cold blob in the subpolar North Atlantic even amid overall global warming — a real, observed regional temperature anomaly — which some research links to further atmospheric circulation changes over the North Atlantic/European sector. This rests on an extra inferential step (AMOC weakening → regional cooling pattern → circulation response) beyond the more direct thermal-wind relationship already driving the primary Arctic-latent-heat jet-stream effect, which is precisely why it's included as a smaller, clearly-labeled secondary contribution rather than treated with equal confidence.

Rising seas and land use: sea level rise is tracked from two sources — Greenland's contribution (the standard 360 Gt ≈ 1mm glaciological conversion) plus a simple thermal-expansion term (ocean water expanding as it warms, calibrated so today's ~1.5°C anomaly gives ~1.3mm/yr, in line with commonly-cited estimates of thermal expansion's real-world share of total sea level rise). Antarctica and mountain glaciers are real additional contributors not separately modeled here, so this is a partial but substantial share of the true total, not the whole picture. Rising seas inundate low-lying coastal land — the UN-cited Low Elevation Coastal Zone (below 10m elevation) holds roughly 10% of global population and significant delta agriculture (Nile, Ganges-Brahmaputra, and Mekong deltas among the most exposed) — and this is modeled as a genuine reduction in available arable land, capped well below total loss, that feeds directly into the land-competition mechanism elsewhere: less available land intensifies the same human-vs-wild-habitat competition already modeled (more deforestation encroachment, more biodiversity fragmentation), rather than existing as an isolated, disconnected number. Tested directly to the extreme 200-year scenario above: sea level rise reaches ~4.5m and coastal land loss approaches but does not reach its 10% cap, remaining bounded and stable throughout.

Mountain glaciers: a third, genuinely different ice population, with real regional water-supply consequences. Global glacier ice outside the Greenland/Antarctic ice sheets totals roughly 158,000 km³ (~144,000 Gt, Farinotti et al. 2019's consolidated global inventory) — far smaller than Greenland, but critical for regional water security in ways the ice sheets aren't. Current global mass loss is modeled at ~280 Gt/yr (Hugonnet et al. 2021, Nature — satellite-derived, with clear acceleration through the 2015-2019 period). Four regions are tracked separately and explicitly, as requested: Hindu Kush Himalaya (~6,500 Gt — ICIMOD's Hindu Kush Himalaya Assessment, Wester et al. 2019), tropical Andes (~900 Gt), European Alps (~100 Gt, having already lost roughly 60% of their Little-Ice-Age volume), and US Rockies/Cascades (~30 Gt — genuinely small in absolute global terms, but locally significant). The remainder (~136,000 Gt) sits mostly in Alaska, Arctic Canada, and Patagonia — real ice, but outside the four ranges specifically requested here, a genuinely important point since the Himalaya, while critical for human water security, is not actually the largest ice reservoir by volume.

On "using the atmospheric column we modeled" for altitude — this was checked directly rather than assumed to work, and the honest result mattered more than forcing a connection. Querying the multi-layer grey-atmosphere model for how warming varies with layer height shows it implies less warming at typical glacier elevations (~5,500m) than at the surface — about 79% of the surface rate, the opposite of what's actually observed. That's because the layer model captures pure radiative column physics, not the surface-boundary-layer mechanisms that actually drive real "Elevation-Dependent Warming" (snow-albedo feedback at the retreating snowline, moisture/lapse-rate feedbacks, black carbon deposition — Pepin et al. 2015, Nature Climate Change, establishing that most high-mountain regions empirically warm 1.3-2x faster than nearby lowlands). Rather than force a connection that doesn't hold, the amplification used is grounded in that empirical literature; the atmospheric column model is used for what it's actually designed for instead — a genuine diagnostic of the realistic ambient temperature structure at altitude, visible on its own chart, not the source of the amplification factor itself.

Seasonal replenishment alternates net winter accumulation with net summer ablation around the underlying warming-driven trend, rather than a smooth monotonic decline — real glaciers gain mass seasonally even while losing it over the year as a whole. A genuine hemisphere split is applied: the Andes (Southern Hemisphere) get an opposite-phased cycle from the Northern Hemisphere ranges (Himalaya, Alps, Rockies) — verified directly by tracing quarterly mass changes: Himalaya shows its smallest loss in Q4 (Northern winter) and largest in Q2 (Northern summer), while the Andes show the exact opposite pattern, confirming the hemisphere phasing is genuinely different, not just labeled differently. Both cycles average to exactly 1.0 over a full year, so the underlying multi-year mass-loss trend is preserved exactly — only the intra-year distribution differs.

Albedo and atmospheric moisture: mountain glaciers cover a far smaller area than Arctic sea ice (~700,000 km² globally, Randolph Glacier Inventory, versus the Arctic's multi-million-km² seasonal range), so the albedo effect is real but genuinely small — calibrated to roughly 1/15th the Arctic ice-albedo term's magnitude at full loss, added as its own separately-attributable, independently-lagged contribution to the temperature decomposition above (now an eight-way split, verified to still sum exactly). Atmospheric moisture is treated even more conservatively: glacier retreat modestly reduces high-altitude sublimation/evaporation surface area, a real but small and more uncertain pathway, applied as a minor reduction to the land-side relative humidity factor already tracked elsewhere rather than a major independent forcing.

Himalaya, Pakistan/India/China, and the poor tier specifically. Glacier-fed rivers initially see increased flow as melt accelerates ("peak water") — this falls directly out of the model's own mass-balance mechanics (ablation is a fraction of remaining mass, so absolute melt flux rises before it eventually falls as the reservoir depletes), not a separately bolted-on assumption. But once a sufficient share of the ice buffer is gone, flow becomes less reliable, particularly damaging during dry seasons when glacial melt is normally the critical buffer. ICIMOD's assessment (Wester et al. 2019) cites roughly 1.9 billion people across South and East Asia depending on Hindu-Kush-Himalaya-fed river systems (Indus, Ganges, Brahmaputra, Yangtze, Yellow River) for water security. Modeled as a water-stress index that activates once the Himalayan pool falls below 60% of its baseline (a "past peak water" threshold) — verified directly: mortality stays at exactly zero for six full decades in the default scenario until that threshold is crossed around 2088, then rises smoothly to a ceiling of roughly 1.07%/yr excess mortality in the poor tier specifically (structured identically to the existing fish-dependent-famine mechanism, at comparable scale — roughly 1 billion of the poor tier's 2.8 billion baseline assumed meaningfully water-stressed).

Rockies, Andes, Alps: groundwater and aquifer depletion. As these smaller glacier-fed systems decline, agricultural and municipal users increasingly turn to groundwater to compensate for reduced and less reliable surface water — a well-documented adaptation response, though real aquifer depletion (e.g. northern India, the US High Plains) is genuinely multi-causal, driven by irrigation policy and pumping practices as much as by glacier loss itself; this model captures the glacier-loss contribution specifically, not the full multi-causal picture, and says so rather than overclaiming. Tracked as a combined aquifer-stress index once the temperate-region glacier pool (Andes+Alps+Rockies combined) falls below 65% remaining, feeding a modest yield penalty concentrated in wheat — temperate-zone dominant, matching where Rockies/Andes/Alps-fed irrigated agriculture actually sits — on top of, not instead of, the jet-stream volatility penalty already applied there.

An honest note on how this was built: the calculation logic above — regional pools, altitude diagnostics, seasonal hemisphere cycling, the Himalaya and aquifer feedbacks — was already fully written and, on inspection, genuinely well-reasoned. But checking it directly (rather than assuming a large, already-detailed block of code must be complete) found that the underlying glacier mass state variables had never actually been initialized or persisted: every reference to them was silently evaluating to undefined, which JavaScript propagates as NaN rather than throwing an error — meaning the entire model would have silently produced NaN for temperature from the very first step, with no visible error message. This was caught by actually running the model rather than trusting that a well-written calculation must also be correctly wired in, and fixed by adding the five missing state variables and confirming, by tracing actual output, that mass balance, seasonal cycling, and the water-stress mortality threshold all behave exactly as the (already-good) surrounding documentation described.

IPCC validation mode. With so many Earth-system feedbacks now layered into this model, it becomes genuinely hard to tell whether a large divergence from published IPCC projections comes from the core radiative physics or from one of the many additional mechanisms built on top of it. This mode isolates the former by neutralizing the latter: permafrost, deforestation, and soil-erosion CO2/CH4 all contribute zero; methane clathrate release is force-disabled regardless of its own checkbox; mountain glacier albedo contributes zero; AMOC-driven ocean-sink weakening is disabled (though AMOC strength itself still evolves and displays normally — only its effect on carbon uptake is severed); forest cover and soil carbon are held at their pristine 2026 baseline (checked directly and fixed once found: zeroing deforestation's *emissions* alone wasn't enough, since forest cover could still decline from temperature-driven fire and tipping-point mechanisms, independently weakening the land carbon sink even with its emissions already zeroed — the same issue existed for soil carbon via wind erosion); and population is frozen at its 2026 baseline in all three tiers, which also prevents bottom-up fossil demand and the methane baseline from drifting away from a clean comparison. Water vapor, cloud, and ice-albedo feedbacks are deliberately left active, since these are "fast feedbacks" already included within IPCC's own definition of climate sensitivity — turning them off would make this a worse comparison, not a better one. Verified directly: with the box unchecked, results match this model's ordinary behavior exactly (no regression at all); with it checked, the model produces a visibly closer, but not perfect, match to the dashed official-RCP reference line already on the CO2 chart.

An honest result from testing this, not a claim of validation success: even with every Earth-system extra neutralized, a residual gap remains between this model's CO2 trajectory and the official RCP curve — and testing multiple pathways found that gap isn't even consistently signed. Under RCP4.5, the model runs about 26% *above* the official 2126 concentration; under RCP8.5, it runs about 8% *below* it. That inconsistency is itself informative: it points to this model's simplified ocean-plus-land sink formulation having a different net-absorption calibration and different saturation behavior than the more sophisticated, multi-model-ensemble carbon-cycle assumptions underlying the official RCP curves, rather than to any single missing feedback. This is reported honestly rather than tuned away — the purpose of this mode is to let you see where the core physics actually stands relative to IPCC, not to manufacture agreement with it.

Methane clathrate release — optional, experimental, off by default. Methane clathrates (hydrates) are ice-like structures trapping methane, stable only under specific pressure/temperature conditions — the mechanism behind the "clathrate gun hypothesis" occasionally raised as a catastrophic risk. Current expert consensus (Ruppel & Kessler 2017; IPCC assessments) treats a large, abrupt, climate-significant release as unlikely this century — a real but low-probability tail risk, not a mainstream expectation — which is why this is opt-in and clearly labeled as high-uncertainty rather than presented as settled science. Two genuinely distinct populations are tracked separately, because their physical response timescales differ enormously — this is what "location" means here. Deep-ocean hydrates along continental margins hold a large total inventory (recent estimates ~1,500-1,800 GtC, substantially revised down from older ~10,000+ GtC figures — Archer et al. 2009's "slow tipping point" analysis) but sit under hundreds of meters of sediment, so warming has to diffuse downward to reach them: a centuries-scale process, modeled via an explicit, extremely slowly-lagged "deep ocean heat proxy" state (a ~300-year characteristic response time) distinct from surface temperature. Shallow Arctic subsea hydrates — particularly the East Siberian Arctic Shelf (ESAS), averaging only ~45-50m deep — sit in a thin, already-marginal stability zone, submerged since sea level rose after the last glaciation and now further stressed by Arctic-amplified warming; this is the one area of genuine, active scientific debate (Shakhova & Semiletov's observed methane seeps vs. more cautious hydrate-specialist assessments), and responds on a much faster decades-scale timescale, tracked using the same polar-amplification factor (2.5x global warming) already used for the boreal crop zone elsewhere in this model. Both populations are triggered only by the model's own warming crossing cited thresholds — never a slider — and once triggered, release only a small, temperature-scaled fraction of the remaining pool per year (a gradual, chronic increase, not an instantaneous "methane bomb" burst; the more alarmist ~50 Gt-over-a-decade scenario proposed by Whiteman, Hope & Wadhams (2013) was itself heavily contested by hydrate specialists and is not what this model reproduces by default). Released methane feeds directly into the same atmospheric CH₄ pool and Myhre forcing pathway already used for permafrost and baseline sources elsewhere — genuinely affecting global temperature, which affects ocean and Arctic warming, which affects further release — rather than existing as an isolated bolt-on. Tested directly: in a default, moderate scenario the shallow Arctic pool triggers around the 2050s once regional warming crosses its threshold, with the released methane measurably pushing the shared atmospheric CH₄ pool past its own already-tracked threshold events downstream, confirming genuine integration; the deep-ocean pool, tested even under an extreme, unmitigated 200-year scenario, still only triggers after roughly a century and barely depletes even once active — correctly representing that this population is effectively inert on any timescale shorter than centuries, exactly the mainstream scientific position this mechanism is built to reflect.

Fisheries collapse: more than a fishing-rate calculation. Fish stock previously responded to fishing effort, ocean acidification, and coral loss — a real but incomplete list. Three further well-documented stressors are now included. Ocean deoxygenation: dissolved oxygen solubility in seawater falls as water warms (the same basic physical chemistry already used for the ocean's CO2 solubility decline elsewhere in this model), and global ocean oxygen content has already measurably fallen — Schmidtko et al. (2017) find roughly 2% lost since 1960 — expanding low-oxygen "dead zones" and compressing habitable water column. Food-chain/primary-productivity collapse: warming increases ocean stratification, reducing the nutrient upwelling that feeds phytoplankton and zooplankton at the base of the marine food web — most visibly documented in Southern Ocean krill, whose sea-ice-linked larval habitat has shrunk with Antarctic warming, cascading up through fish, penguins and whales (used here as a general warming-linked productivity proxy rather than a separately-modeled Antarctic sea-ice system, which is beyond this model's scope). Both reduce carrying capacity (K) alongside the existing acidification and coral terms. Migration and international disputes: as stocks shift poleward and deeper to escape warming waters, they cross into other nations' waters or unregulated international zones — a real, documented pattern (the Northeast Atlantic mackerel dispute among the EU, Norway, Iceland and the Faroe Islands is the clearest case) that empirically tends to *increase* cumulative fishing pressure through uncoordinated, competitive harvesting rather than reduce it, since no existing quota framework covers the newly-shifted stock. Modeled as an extra harvest multiplier (up to +30%) that grows specifically as the stock becomes scarcer — scarcity intensifying competition for what's left, exactly the dynamic seen in real disputes. Migration itself isn't separately tracked as a spatial process (this model has no regional/ geographic resolution for fisheries), but its most consequential real-world effect — disputes eroding cooperative management once stocks relocate — is. Tested directly: even under a low fishing-effort scenario (well below sustainable MSY), fish stocks still decline substantially from climate stressors alone (temperature still rising to +3°C even under otherwise favorable settings), confirming these mechanisms operate independently of — not merely alongside — fishing pressure itself.

CO2 per annum and CO2e per annum, plotted together: two new lines make the model's total emission rate directly comparable to how global emissions are conventionally reported. CO2 per annum annualizes the model's quarterly total emissions (fossil + deforestation + permafrost + soil erosion) to GtCO2/yr; CO2e per annum additionally converts this quarter's actual methane mass emissions (from permafrost thaw and the agricultural/wetland baseline) into their CO2-equivalent warming contribution using methane's 100-year Global Warming Potential (GWP100 = 27 for non-fossil methane, IPCC AR6) — the standard metric for putting different greenhouse gases on one comparable scale. Plotted concurrently, the gap between the two lines is a direct visual measure of how much extra warming-equivalent methane adds on top of pure CO2 in any given year (today's baseline: ~50 GtCO2/yr vs. ~62 GtCO2e/yr).

Human-released vs. natural (warming-driven) CO2, on the same chart: fossil fuel combustion and deforestation are direct anthropogenic actions — categorized as human-released. Permafrost thaw and soil-carbon loss from wind erosion/desertification are Earth-system feedbacks that occur as a consequence of warming (both are modeled elsewhere in this simulation as responding to temperature and desertification, not to direct human combustion) — categorized as natural, due to warming. This split is not perfectly clean — some deforestation is itself climate-driven (heat/drought fire loss, tropical dieback once tipped) and some soil erosion is human-driven (agricultural intensification) — but it captures the dominant, most policy-relevant distinction: emissions humans directly control by burning less fuel and clearing less land, versus emissions the Earth system generates on its own once warming is underway, which humans can no longer simply switch off. The two lines are verified to always sum exactly to the total CO2-per-annum line above (checked directly across a full 100-year run) — a genuine decomposition, not an approximation. Watch how the balance shifts over a long run: today, human sources dominate by a wide margin (~85% of the total); in unmitigated scenarios, the natural/feedback share can still shrink in absolute terms even as the human share grows much faster (permafrost's *own* carbon pool is finite and depletes, while fossil-driven emissions can keep climbing) — a reminder that "natural" emissions being smaller doesn't mean they're harmless, only that they're not the largest lever humans still directly control.

Building this exposed and fixed a real, pre-existing initialization bug: the very first simulated step showed permafrost contributing an implausible ~36 GtCO2/yr — nearly as much as all fossil emissions combined — because the permafrost model's cumulative "thaw index" started at a literal zero, creating an artificial rate-of-change spike (going from "no accumulated thaw exposure at all" to some nonzero value in a single step, rather than the smooth, gradual buildup a real Stefan-equation-style accumulator should show). This had been present since the permafrost mechanism was first built, but only became obvious once an annualized emissions-rate chart existed to expose a single-step transient that longer-run trajectory charts had been smoothing over. Fixed by starting the thaw index at a value representing roughly two decades of already-accumulated exposure (permafrost hasn't been at "zero cumulative thaw" for all of history — it's already been warming for decades before this simulation begins in 2026), which brought the first-step permafrost contribution down to a defensible ~5.5 GtCO2/yr — still on the higher end of published current-day estimates (order of magnitude ~0.3-several GtCO2e/yr today, growing substantially by century's end in high-emissions scenarios), consistent with this model's general tendency to represent the more comprehensive end of feedback assessments, but no longer an outright initialization artifact.

Ocean pH lag and genuine two-way exchange — checked directly, found the ocean was recovering unrealistically fast and couldn't outgas at all. The previous pH formula computed pH directly from the same step's atmospheric CO2 with no memory whatsoever — if CO2 dropped for any reason (RCP2.6's late- century decline, or now direct air capture), pH would fully recover to match the new, lower level in a single simulated step. A study measuring exactly this (surface ocean pH response to atmospheric CO2 perturbation) found a single exponential relaxation time of 52 ± 10 years (36 ± 7 year half-life), explicitly confirming the process is genuinely reversible — real, but nowhere near instant. pH now lags toward its CO2-implied target at that measured rate. This also closed a second, related gap: the ocean sink previously only ever absorbed CO2, with no mechanism for releasing any back — physically wrong, since an ocean that has already absorbed a large pool of carbon genuinely resists a sudden atmospheric CO2 decline by outgassing some of it back, exactly what real marine carbon-dioxide-removal research describes ("the solubility pump... yields CO2 outgassing... in low or zero emission scenarios"). Modeled as a direct consequence of the same pH lag rather than a separately-invented mechanism: whenever the ocean is more acidified than current CO2 would be in equilibrium with, it releases CO2 proportional to that deficit, calibrated from the same pH-CO2 relationship and the same 52-year timescale, not a new assumption. Tested directly with a controlled scenario (CO2 suddenly dropped from 600 to 450 ppm): pH recovered by only 0.002 over 19 months rather than snapping instantly, and genuine outgassing of ~0.95 GtCO2/month was confirmed independent of any other emissions source in the test — the ocean measurably pushing back against the sudden decline, exactly as real ocean chemistry does. This directly matters for the direct- air-capture mechanism added earlier: removing atmospheric CO2 doesn't erase the ocean's already-absorbed carbon burden, which was previously invisible to this model and is now a real, quantified resistance.

Ocean CO2 absorption, modeled as genuinely independent of forest cover: the ocean and land carbon sinks were previously combined into a single formula gated entirely by forest cover — meaning if forests collapsed toward zero (which happens in several tested scenarios once the tropical dieback tipping point is crossed), the model incorrectly implied ocean absorption collapsed too, down to an arbitrary 6% floor. That's a real physical error: the ocean absorbs CO2 through its own chemistry and physics — gas dissolving into seawater, mixing, the biological pump — none of which depends on how much rainforest exists on land. Ocean and land sinks are now two independent terms that sum to the total, split from the Global Carbon Project's long-term partition of emissions (~45% remaining airborne, ~25% absorbed by the ocean, ~30% by land) so today's baseline is unchanged. The ocean term weakens with acidification (the same acidStress used for coral/fish elsewhere — as the ocean absorbs more CO2 its buffering capacity for absorbing further CO2 genuinely declines, the Revelle-factor effect) and independently with ocean warming, since CO2 solubility in seawater falls as temperature rises (basic physical chemistry — the reason polar oceans are stronger CO2 sinks than tropical ones) — but never with forest cover. Tested directly across a full 100-year run in which forest collapses from 99.9% to 5%: the ocean sink's own contribution stays meaningfully positive throughout (weakening from ~12% to ~4% as the ocean itself acidifies and warms, not because trees disappeared), on top of whatever the separately-computed land sink still contributes. This is not merely a cosmetic fix — it measurably reduces resulting CO2 and warming in already-tested scenarios compared to the earlier, buggy version, since the ocean's real, ongoing absorption is now properly credited rather than incorrectly zeroed out alongside deforestation.

Regenerative land management — a deliberate positive, added under Option 3. Three real, citable practices: recovered organic waste (human and livestock) applied as fertilizer/soil amendment instead of lost to lagoons or landfill; no-till farming (planting into previous rootstock rather than tilling after every harvest); and farmland deliberately given over to rewilding and hedgerows. Calibrated conservatively against real research rather than the most generous available figures. No-till: 0.12-0.47 Mg C/ha/yr (Carbon Balance and Management, 2020), scaled to ~0.3 Gt C/yr global technical potential at full adoption — deliberately on the conservative end, since a 2024 World Resources Institute review found some widely- cited older no-till carbon estimates "are no longer justified" by current evidence, and a global meta- analysis found no-till can increase some other emissions (CH₄, N₂O) even while building surface carbon; this model uses the newer, more conservative per-hectare figures rather than the larger, contested older ones, and also reduces wind erosion directly (feeding the same erosion mechanism already modeled elsewhere). Organic waste recovery: given a comparably conservative share of Lal (2004)'s combined 0.4-1.2 Gt C/yr multi-practice estimate for improved cropland management overall, plus a modest methane reduction (anaerobic lagoon storage — the alternative to recovery — is a real, documented CH4 source; kept small since livestock enteric fermentation, not manure storage, is the dominant agricultural methane source and isn't affected by this slider). Rewilding: directly boosts forest cover and biodiversity (hedgerow biodiversity gains of ~32% are documented; Strassburg et al. 2020, Nature, found restoring priority ecosystems on 15% of current farmland could spare 60% of expected species extinctions while sequestering a cumulative 299 Gt CO2, with diminishing but still substantial returns to 30% coverage) — but genuinely trades off against arable land, reducing food-producing area exactly as much as the percentage chosen, the same land-competition pressure already modeled elsewhere on this page. Defaults (15% manure recovery, 15% no-till, 0% rewilding) reflect realistic current global adoption, not an idealized starting point.

Checked directly, after being asked, whether these sliders' soil-health benefits actually reached crop yield or only affected the carbon-cycle sink: they only reached the sink. soilCarbonGtC and topsoil were both being computed and correctly feeding the land carbon sink's strength, but the crop yield formula referenced only local temperature and a fixed insolation constant — soil health had no effect on yield at all, despite real research being explicit that they're linked (Lal 2004, Science: "an increase of 1 ton of soil carbon pool of degraded cropland soils may increase crop yield by 20-40 kg/ha for wheat, 10-20 kg/ha for maize"). Fixed by blending organic soil carbon content and physical topsoil depth into a yield multiplier, referenced against today so it's exactly 1.0 at 2026 baseline (verified directly: 0.9996 after one month of ordinary baseline erosion, not a discontinuity). Verified the fix actually changes yield, not just theoretically: with identical 2%/yr deforestation pressure, rice yield reaches 25.2 by 2076 with these three sliders at zero adoption versus 48.2 at full adoption — soil health degradation or recovery now measurably drives food production, not only the carbon budget.

Worth being direct about the scale of effect when these are pushed toward their maximums, tested directly rather than assumed: a bottom-up scenario at default settings reaches ~6°C by 2126, while an otherwise identical run with all three sliders maximized (100% recovery, 100% no-till, 30% rewilding) stays close to 1.5°C the entire century. That gap is much larger than the direct carbon-sequestration numbers above would suggest on their own, and the reason is specific and checked directly rather than assumed: with forest cover protected from ever net-declining, this model's existing dieback/tipping-point cascade (forest loss weakening the land carbon sink, which raises CO2, which drives more heat-and-drought-driven dieback) simply never gets triggered in the first place. This isn't these three sliders being overpowered on their own terms — it's a genuine emergent consequence of preventing a threshold from being crossed in a model that, like the real climate system it's approximating, treats several of its feedbacks as compounding rather than independent. The scale of the difference is a real property of tipping-point dynamics, not a calibration error, but it's worth understanding why the number is so large rather than taking it at face value.

Six further additions under Option 3 — direct air capture, solar/wind side-effects, AI/data centers, and two "is this even modeled" checks. Direct air capture is modeled with its genuine energy cost, not as free removal — published figures put real DAC technology at 5-15 GJ/tonne CO2 (Belfer Center/Harvard, 2023), and the share of that energy not sourced from DAC's own dedicated low-carbon supply draws on the same grid mix as everything else. Tested directly: at a low renewables share and 0% dedicated clean energy, roughly three-quarters of captured CO2 is re-emitted generating the power to capture it, leaving only about a quarter as genuine net removal — closely matching the literature's own warning that DAC on a fossil grid "may even result in net emissions." One real bug was caught and fixed while verifying this: an early version of the energy-to-emissions conversion used the wrong kg-to-gigatonne factor (1 Gt = 10^12 kg, not 10^9 — the earlier code divided by 10^9), which made DAC's own emissions come out roughly 1000x too large before being caught by checking the actual numbers rather than trusting the formula looked right.

Solar and wind side-effects are deliberately linked to the existing renewable growth slider rather than separately adjustable, since they're consequences of that same deployment, not an independent choice. Solar: real, satellite-measured local albedo reduction (Yuan et al. 2024, Communications Earth & Environment) applied over an illustrative, conservative estimate of global land coverage, with agrivoltaics reducing but not eliminating the effect. Wind: genuine kinetic energy extraction from the atmosphere, checked directly against the literature and modeled at the small magnitude that research concludes is appropriate at global-average scale, even though real local/regional effects are documented. AI/data center growth is linked to the same demand multiplier Jevons rebound already drives, exactly as requested, plus genuine (if bounded) water and land-use pressure feeding the same aquifer-depletion and arable-land mechanisms already in this model — not a disconnected new resource.

The two checkboxes answer a direct question: is fossil combustion's literal waste heat, and human metabolic heat, already represented in this model's forcing? Checked directly: neither was — like essentially every mainstream climate model, this one represented fossil fuels' effect entirely through greenhouse forcing, never the literal thermodynamic heat combustion releases (true even though an internal combustion engine really does waste roughly 70-80% of fuel energy as heat). Calculated and added at their real, cited magnitudes when checked (Flanner 2009: global anthropogenic heat flux ~0.03 W/m², about 1% of greenhouse forcing; human metabolic output ~100W/person, ~0.85 TW globally at 8.5 billion people) rather than left as an assumption in either direction — both real, both genuinely small next to the ~460 TW already being trapped by the enhanced greenhouse effect, and both off by default for that reason, not because the physics isn't real.

Aerosol unmasking — the model's initial warming rate was checked directly against observations and found far too slow, even after the albedo fix above. Fixing albedo's initialization correctly introduced a genuine committed-warming gap, but checking the resulting t=0 rate of warming directly against real-world data found it came out to only ~0.08°C/decade — nowhere near the observed rate. Multiple 2026 sources converge on why: real-world warming has accelerated to roughly 0.3-0.4°C/decade, up from ~0.18°C/decade in 1970-2010 (Hansen, "2026 On Track for Warmest Year"; Foster & Rahmstorf, Geophysical Research Letters 2026, confirming ~0.35°C/decade since 2015 across five independent temperature datasets). The mechanism: industrial aerosol pollution has historically reflected sunlight, partially masking greenhouse warming that was already "committed" by accumulated CO2 — and aerosol emissions have fallen sharply since roughly 2020 (shipping-fuel sulfur regulations, China's air-quality improvements), unmasking warming that was always implied but suppressed from being observed. This model's CO2/CH4/water-vapor/cloud/albedo physics has no aerosol term of any kind, so it couldn't represent this. Modeled as a temporary, decaying addition to the equilibrium target — largest today (calibrated so the model's actual t=0 decadal rate lands inside the cited 0.3-0.4 range, verified directly rather than asserted) and fading over a multi-decade half-life as the global aerosol transition completes, since most of this reflects a one-time historical shift rather than an ongoing physical process. Verified the resulting trajectory rises smoothly with no artificial dip as the term fades (the lagged component actually continues rising for several decades even as its target has already begun declining, since it's still catching up to the initially-high target — ordinary lag behavior, not a bug).

This has a real, substantial consequence worth stating plainly: faster initial warming pulls forward every temperature-threshold-dependent feedback in this model. Tested directly with a controlled comparison (identical scenario, aerosol term on vs. off): the Arctic Blue Ocean Event threshold is reached 27 years earlier with this term included (2065 vs. 2092) — which is exactly the point of adding it. A model whose starting conditions don't match the observed present will systematically understate how soon downstream feedbacks engage, regardless of how well-calibrated those feedbacks are individually.

Lapse-rate feedback and water vapor shortwave absorption — checked directly, found genuinely missing. Investigating the AR5 divergence (see the RCP2.6/4.5/6.0 comparison this was built to explain) led first to a hypothesis that turned out wrong: that water vapor and clouds were being lagged twice — once implicitly (their targets necessarily reference last quarter's temperature, since they feed into this same step's equilibrium calculation directly, so using the not-yet-computed current value would be circular) and again explicitly (their own decomposition-lag). Giving them a faster, decoupled lag was tested directly and made the divergence *worse*, not better — proof that the issue was never about timing, only about the equilibrium magnitude these feedbacks settle toward. That pointed to what was actually missing: real climate science essentially never reports water-vapor feedback in isolation — it's always assessed jointly with lapse-rate feedback (Soden & Held 2006; Vial et al. 2013; Sherwood et al. 2020's modern assessment uses "combined water vapor and lapse-rate feedback" as one of just six feedback categories) precisely because treating water vapor alone systematically overstates the net effect — a real, well-documented negative feedback this model was missing entirely. The mechanism: the upper troposphere (where outgoing longwave radiation actually leaves for space) warms more than the surface under moist-adiabatic dynamics, and a relatively warmer upper troposphere radiates more efficiently to space, substantially offsetting water vapor's positive effect. Standard multi-model-mean magnitudes — water vapor alone ~+1.7 W/m²/K, lapse rate alone ~-0.5 W/m²/K — mean lapse rate cancels roughly 25-30% of water vapor's raw effect; this model's own ratio holds at ~26% through 2100, matching that citation, before drifting up at extreme, later-century warming due to the two terms' different functional forms (log-based vs. linear), not a flaw in the underlying mechanism. This also stands in for upper-tropospheric humidity's outsized influence on outgoing longwave radiation (small absolute humidity changes high in a cold, optically-thin atmosphere have disproportionate effects — Held & Soden 2000) rather than building a fully separate vertically-resolved humidity tracker, since UTH and lapse-rate are driven by the same upper-tropospheric process and are conventionally assessed together for exactly that reason. A smaller water-vapor shortwave-absorption term is added alongside it — water vapor doesn't only trap outgoing longwave radiation, it also directly absorbs a modest, growing amount of incoming solar radiation, roughly an order of magnitude smaller than the longwave effect.

Genuinely state-dependent, as requested — sensitivity changes with temperature without touching the climate-sensitivity slider. Research on state-dependent climate sensitivity (Bloch-Johnson et al. 2021 and related work) finds the lapse-rate feedback's offsetting strength isn't constant — it weakens at higher warming as tropical moist-adiabatic dynamics change character, meaning progressively less of water vapor's raw warming gets cancelled as temperature rises. Modeled directly: the offset coefficient itself shrinks with warming (floored at 30% of its baseline -0.5 W/m²/K strength, never eliminated entirely) — verified directly at +6°C the coefficient has weakened to ~70% of its 2026 value, exactly as designed. This is the actual mechanism by which this model's effective sensitivity now increases with warming — a real physical process, not a slider being nudged. Tested against AR5 directly: this closed a genuine, if partial, share of the divergence, improving the 2081-2100 gap by 0.16-0.23°C across RCP2.6/4.5/6.0 (largest at RCP6.0, growing over time as the state-dependence compounds) — not a full resolution, but a real, physically-grounded step, verified with the full ten-way temperature decomposition still summing exactly and no regressions across every RCP mode, validation-mode, and clathrate-setting combination.

Non-linear carbon sink saturation, with population as the requested feedback. The acidification and warming effects above respond to *conditions*; three further terms make both sinks respond directly and non-linearly to *how much CO2 there already is* and to population, compounding with those condition-based effects rather than duplicating them. Ocean: the Revelle (buffer) factor — the share of absorbed CO2 that shows up as increased seawater pCO2 rather than staying safely as bicarbonate — genuinely rises as the ocean absorbs more carbon (Revelle & Suess 1957; Sabine et al. 2004), from roughly 9 pre-industrial to roughly 10.5 today, projected toward 13-15+ under sustained high CO2. A higher Revelle factor means progressively less of each additional unit of CO2 gets buffered — modeled with a genuine logarithmic dependence on CO2 concentration (smoothly accelerating decline, not a straight line), verified directly: ocean uptake efficiency falls to ~77% of today's value by the time CO2 reaches ~1,050ppm in a 100-year unmitigated test. Land: CO2-fertilization of photosynthesis is real but saturating — each additional ppm boosts plant growth less than the last, since nitrogen and phosphorus availability become the binding constraint as CO2 itself becomes less limiting (Fernández-Martínez et al. 2017; Wieder et al. 2015 on nutrient limitation of the fertilization effect). Modeled with a standard Michaelis-Menten saturating curve, capped at 1.8x today's land-sink strength — a moderately generous reading of a literature range that varies considerably by nutrient-limitation assumptions, disclosed as such rather than presented as precise. Population, the requested feedback: land-use pressure structurally reduces the land sink's *capacity*, not merely its current condition — every hectare converted to farmland or settlement to feed and house a larger population is a hectare that can no longer function as a carbon sink at all, independent of how healthy the remaining forest is. More people beyond today's 8.2B baseline erodes that capacity; fewer people (via the land-abandonment/rewilding dynamic already modeled elsewhere) allows some of it to be reclaimed, capped at a 15% recovery ceiling well short of full reversal, since land conversion has real permanence (soil structure, lost old-growth complexity, built infrastructure). All three terms are referenced against today's exact conditions (verified directly: each equals precisely 1.0 at 2026 baseline CO2 and population, so no discontinuity is introduced) and are correctly left active in IPCC validation mode, since the Revelle factor and CO2-fertilization saturation are standard carbon-cycle physics genuine IPCC/CMIP carbon-cycle models also include — only the population-capacity term is neutralized there, consistent with population itself being frozen in that mode. One implementation bug was caught and fixed while building this: an early version of the new terms referenced this quarter's not-yet-computed CO2 value to determine the very sink fraction that CO2 value itself depends on — a circular reference that would have thrown a runtime error — found by actually running the code rather than trusting the formula looked reasonable, and fixed by referencing last quarter's CO2 instead, consistent with the same circularity-avoidance pattern used throughout this model.

Fossil fuel as a finite, depleting resource (optional): toggled by the checkbox in the control panel; unchecked, the model reverts exactly to the original business-as-usual assumption of unlimited fossil supply (population and demand alone set emissions). Checked, fossil fuel can only be burned as fast as it can actually be extracted. Today's known, technically/economically recoverable reserves (oil+gas+coal combined) are set at ~900 GtC, consistent with published estimates that remaining proved+probable reserves would emit roughly 2,900–3,300 GtCO₂ if fully burned (McGlade & Ekins 2015; Carbon Tracker) — several times the remaining 1.5°C carbon budget. That 900 GtC is only a fraction of a much larger ~3,100 GtC geological resource base still undiscovered or currently uneconomic; the new fuel source discoveries slider controls how much of that larger pool is converted into known reserves each year through ongoing exploration, at a 22.5% recovery factor (most fossil carbon in the ground is never actually extractable even in principle, a standard feature of reserve accounting).

EROEI and reduced draw-down: energy-returned-on-energy-invested is tracked explicitly and plotted directly against reserves remaining in its own chart. Historical conventional oil/gas EROEI exceeded 30:1; today's blended global fossil EROEI sits in the widely-cited 15–30:1 range and is falling as easy reserves are exhausted, with unconventional sources (tar sands, tight oil) often below 5:1 and occasionally below the break-even point of 1:1 — a net energy loss not worth extracting regardless of price (Hall et al.'s "net energy cliff"). The model curves EROEI from an initial 20:1 down toward a floor of 2:1 as the reserve depletes, and — critically — uses that same curve to throttle the physical draw-down rate itself: harder, lower-grade deposits genuinely cannot be extracted as fast as easy ones, regardless of demand, so extraction slows on its own as EROEI falls, on top of the usual reserve-proportional decline-curve behavior (calibrated to today's real-world ~2%/yr reserve-to-production ratios for oil and gas).

Feeding back into emissions — and wealth-insensitive demand: actual fossil emissions equal actual extraction, not raw demand — if it can't be pulled out of the ground, it can't be burned. But demand isn't uniformly price-sensitive: roughly 36% of fossil consumption is modeled as effectively price-insensitive — concentrated among the wealthiest consumers and highest-value uses (private aviation, shipping, heavy industry, strategic/military use) that keep consuming regardless of rising extraction cost, reflecting the well-documented concentration of emissions among high-income consumers (Oxfam and related carbon-inequality research consistently find a small wealthy share of the population responsible for a disproportionate share of emissions). The remaining ~64% is price-sensitive and shrinks as EROEI falls (rising extraction cost/effort destroys demand through higher prices). Note that the model does not resolve exactly where the resulting energy shortfall is absorbed — foregone consumption, faster renewable buildout, or economic contraction are all plausible real-world responses it doesn't explicitly separate; it only guarantees that the emissions consequence of constrained supply is captured correctly.

Jevons paradox / energy rebound: renewable growth is not assumed to cleanly substitute for fossil fuels by default. As William Stanley Jevons first observed in 1865 (more efficient steam engines increased, rather than decreased, Britain's total coal consumption), and as formalized in the Khazzoom-Brookes postulate, cheaper and more abundant energy tends to stimulate extra consumption — new uses, more industrial activity, cheaper transport and computing — which can partially or, in some documented cases, more than fully ("backfire") offset the emissions reduction that fuel-switching alone would produce. The Jevons rebound slider controls this directly: growth in renewable share above the 2026 baseline (15%) inflates total energy demand by rebound × that growth, on top of ordinary population- driven demand. At 0, renewables substitute cleanly with no extra demand (the model's original assumption); at 1.0, each extra percentage point of renewable share adds a full extra percentage point of total energy demand; published estimates of the direct rebound effect for most efficiency/energy-source improvements cluster around 10-30%, while economy-wide "backfire" exceeding 100% is more contested but has historical precedent (Jevons' own coal example, and some readings of 20th-century electrification). Watch the Renewables share climb in the vitals strip alongside CO₂: with a high rebound setting, renewable adoption can look like real progress while emissions barely bend.

Fish-dependent famine mortality (distinct from the general food/population link): the blended food index below (70% crop yield, 30% fish stock) already fed into population via a standard carrying-capacity mechanism, but testing exposed a real gap: because crops and fish were blended into a single global average, a total fish-stock collapse sustained for over a decade produced almost no population response at all, so long as crop yields stayed healthy. That's not realistic — FAO estimates around 1 billion people worldwide rely on fish as their primary source of animal protein, and this group is disproportionately poor and coastal, with little practical ability to substitute in crop calories even when global crop production is, in aggregate, a surplus — that surplus doesn't reach them, whether from cost, distribution, or geography. A second, independent penalty now applies specifically to that ~1 billion (modeled as 12.2% of today's population), tied directly to fish stock with essentially no crop-surplus cushion, since the entire point is that the surplus doesn't reach them. It contributes nothing at today's fish level and scales up to ~4%/yr excess mortality within that sub-group at total, sustained collapse — a severe but not apocalyptic famine rate, comparable to historically documented severe famine episodes. Tested directly: the same fish-collapse scenario that previously left population essentially untouched for over a decade now shows real, immediate decline (population down to 7.88B by the same point that previously showed 8.20B) — a working feedback rather than a diluted-away one.

Cloud cover: low clouds cool, high clouds warm — modeled as genuinely distinct physical effects, not one number: satellite observations (CERES-type measurements) consistently show clouds pull in two directions. Low, warm clouds (marine stratocumulus, cumulus — tracked here starting at ~35% global coverage) are optically thick and reflect a lot of sunlight, but sit close to surface temperature and so trap little outgoing longwave radiation: net effect, cooling (today's observed low-cloud effect is roughly -35 W/m² of shortwave cooling globally). High, cold clouds (cirrus, convective anvils — tracked starting at ~22% coverage) reflect comparatively little sunlight, but being so cold they strongly suppress outgoing longwave radiation — the same physical mechanism as an added greenhouse-gas layer, which is why high clouds are folded directly into the multi-layer Stefan-Boltzmann calculation built earlier for CO2 and methane, rather than bolted on as a separate approximation: a change in high-cloud fraction adds to the effective layer count N exactly the way CO2 or methane does, calibrated to today's observed ~+10-15 W/m² net high-cloud warming effect. Low clouds are documented to thin with warming (Klein et al. 2017's review of marine boundary-layer cloud responses) — a real, moderate- confidence positive feedback, since less low cloud means less reflected sunlight. High clouds show a small, more uncertain increase, loosely consistent with the "Fixed Anvil Temperature" hypothesis (Hartmann & Larson 2002). IPCC AR6 explicitly flags net cloud feedback as the single largest source of uncertainty in climate sensitivity estimates; this implementation lands on the net-positive (mildly amplifying) side of that assessed range, and is clearly labeled as such rather than presented as settled. Both cloud fractions are hard-bounded (low cloud cannot fall below 15%, high cloud cannot exceed 32%) — tested directly to 400 simulated years under an extreme, unmitigated scenario: both fractions reach their bounds and hold there, with warming continuing (from CO2 and other still-active drivers) but at a visibly decelerating rate once the cloud contribution itself saturates, confirming genuine convergence rather than runaway.

Water vapor to cloud formation, and cloud "whiteness" — checked directly, and genuinely missing. Cloud fraction above previously responded only to temperature, with no explicit dependence on humidity at all — a real gap. Standard GCM cloud parameterization (Xu-Randall and Sundqvist diagnostic schemes, and the broader review in Zhang et al. 2023) computes cloud fraction directly from relative humidity, typically ramping up once RH crosses a critical threshold (~0.7-0.9) toward saturation: cloud fraction is fundamentally a function of moisture availability, not temperature alone. Added as a genuine second term layered on top of the existing temperature-driven mechanism (not replacing it): as this model's own blended land/ocean relative humidity — already tracked elsewhere, including the desertification-driven land-drying effect — moves above or below today's baseline, cloud formation potential moves with it, referenced against today's exact value so there's no discontinuity at 2026 baseline (verified directly: both the humidity boost and optical-thickness factor below compute to exactly zero/1.0 at t=0). This closes a real, previously-missing connection: worsening desertification now measurably suppresses cloud formation as well as land-surface humidity, rather than the two being tracked in isolation from each other. Low cloud (tied to boundary-layer moisture) gets a stronger humidity coupling than high cloud (convectively formed, less directly tied to bulk RH), reflecting that real distinction.

A cloud's reflectivity isn't fixed in reality either — cloud albedo depends on liquid water path and optical depth, which scale with how much moisture was available to condense into it. A cloud forming in a humid atmosphere is optically thicker and genuinely whiter (more reflective) than one forming in a drier atmosphere at the same areal fraction — the "whiter clouds" question directly. Modeled by scaling the shortwave reflectivity coefficient itself (not just cloud fraction) with the same blended humidity, again referenced against today so it's exactly 1.0 at baseline. Tested directly under a severe warming/desertification scenario: as blended humidity declined modestly over a century (0.750→0.735), both the humidity-driven cloud formation boost and the optical-thickness factor responded in the correct direction and stayed appropriately small and bounded — a real, physically-grounded refinement rather than a dramatic rewrite of the existing cloud mechanism.

Cloud shortwave absorption — distinct from reflection: the low-cloud term above is calibrated against satellite-measured net cloud radiative effect, which is a top-of-atmosphere quantity — it correctly captures energy reflected back to space (true albedo, a genuine loss from the whole Earth-atmosphere system) but does not, on its own, distinguish that from energy absorbed within the cloud itself. Clouds do both: some incoming sunlight bounces off cloud tops back to space, and some is absorbed by cloud droplets and ice crystals before ever reaching the ground. The absorbed portion has the opposite implication for the planetary energy budget — it stays in the system rather than escaping — but it still doesn't reach the surface, so it doesn't become surface-absorbed sunlight or surface-emitted infrared the way unobstructed sunlight would. This is now modeled as its own separate term (~22 W/m² of global cloud shortwave absorption at today's baseline coverage, smaller than the ~35-50 W/m² reflected — published estimates put absorption well below reflection for shortwave cloud effects), responding to total cloud cover rather than low cloud alone. The result is a genuinely nuanced, non-obvious behavior: tested directly, as warming proceeds and low-cloud cover thins faster than high-cloud cover grows, total cloud cover falls — meaning *less* shortwave absorption occurs, which this term correctly represents as a small additional cooling nudge (reaching about -0.3°C by 2126 in the default scenario) that partially offsets, without cancelling, the dominant reflection-driven warming from low-cloud thinning above.

Water vapor over land vs. ocean — only a finite amount of water can be drawn from the ground: the atmosphere's relative humidity was previously tracked as one global-average number, implicitly assuming unlimited moisture availability everywhere. In reality, oceans (71% of Earth's surface) are an effectively unlimited moisture reservoir, so ocean-air humidity can track the Clausius-Clapeyron relationship fairly freely — but land (29%) is not: evapotranspiration draws on finite soil moisture, and as that runs short — directly accelerated by the desertification already tracked in the soil/wind-erosion section above — land-air humidity falls short of what atmospheric demand alone would produce, a well-documented land-atmosphere feedback behind many severe heatwaves (Seneviratne et al.'s work connecting soil-moisture deficits to the 2003 European heatwave, among others). Land humidity is modeled as the ocean-equivalent value scaled down by a moisture-availability factor that shrinks with both existing desertification and a further, modest reduction as warming increases atmospheric moisture demand beyond what limited land supply can meet (floored so arid regions retain some minimum humidity from transported ocean air masses rather than reading zero). The two are area-weighted (71%/29%) into the blended humidity that actually drives the water-vapor greenhouse feedback — meaning the more the world desertifies, the weaker the overall water-vapor feedback becomes relative to a naive ocean-only calculation, a real dampening effect this model now captures rather than ignores.

Temperature anomaly, decomposed by cause: the chart no longer shows a single blended line — it shows the baseline warming from CO2 alone, then methane, water vapor, ice-albedo, low-cloud reflection, high-cloud, and cloud absorption feedback each as their own separate line, plus a total line that is the sum of all seven. This isn't just a display trick: because CO2, methane, and high clouds combine non-linearly through the multi-layer Stefan-Boltzmann calculation above (T⁴- based, not additive), methane's and high-cloud's lines specifically show their marginal contribution at each stage — what temperature would be with only the prior terms' effect, subtracted from the result with this term added — so that baseline + methane + high-cloud exactly reconstructs the combined multi-layer warming, term for term. Water vapor, low-cloud, cloud absorption, and albedo were already separate, linearly additive forcing terms, so no such correction was needed for them. Each of the seven components is then run through the model's ocean-heat-uptake lag independently, rather than lagging one blended total — and because that lag response is linear, the sum of the seven independently-lagged components is mathematically guaranteed to exactly equal the model's actual total temperature at every single step (verified directly: summed to within floating-point rounding error, ~1e-15°C, across a full 100-year test run) — a genuine, checkable decomposition rather than an approximation that might drift from the real total over time.

Genuinely separate sub-populations, by wealth tier: population was previously a single aggregate number. It's now three distinct tracked stocks — wealthy (~1.3B, ~16% of today's population, roughly matching a World Bank-style high-income split), middle (~4.1B), and poor (~2.8B) — each with its own baseline demographic rate (wealthy nations: near-zero/slightly negative natural growth; middle: moderate; poor/developing: highest, still in demographic transition, a real and well-documented pattern) and, critically, its own exposure to famine. The shared food-limited carrying-capacity mechanism still applies to total population as a broad demographic ceiling; what's new is that acute, price-driven famine no longer applies uniformly — it hits the tiers very differently, as it does in the real world.

Rice vs. wheat: distinct staple crops, not one blended index: reuses the same underlying zone temperature and arable-land data as before (land viability is a physical/climate property, not crop-specific) but applies each crop's own growing-region weighting and temperature-response curve. Rice is weighted toward warm tropical/subtropical zones with higher heat sensitivity (Peng et al. 2004: rice yields decline several percent per °C of night-time growing-season warming — a specific, well-documented vulnerability distinct from wheat's); wheat is weighted toward cooler temperate zones, more heat-tolerant but still vulnerable during grain-filling (Asseng et al. 2015). The two are tracked and charted separately; a simple average is kept as "crop yield" for backward compatibility with the arable-land plumbing, since rice and wheat provide roughly comparable shares of global cereal calories.

The food supply-chain efficiency buffer, checked and corrected. A direct check of whether rice/wheat yield genuinely feeds into population — and how sharply — surfaced two real problems, both now fixed. Global food production exceeds baseline need, but a large share never reaches anyone: FAO estimates roughly 20-30% of food produced is lost to waste, spoilage and inefficient distribution, meaning today's ACTUAL delivered food is only about 80% of gross yield. The buffer is now modeled explicitly on that basis — assuming a best-case 100% delivery efficiency is achieved (all waste eliminated), gross yield can fall by up to 20% before any shortage registers at all, but once gross yield (times available land) drops below 80% of baseline, even perfect efficiency can no longer sustain today's consumption: a genuine, absolute shortage exists at that point, not before. The previous version used an abstract "20% overproduction" multiplier that, while similar in headline size, wasn't precisely anchored to this 80%-efficiency threshold and — the more significant problem — responded too gradually once triggered: tracing it step-by-step showed mortality still below 1%/yr even after yield had already collapsed to 57% of baseline, far too slow an onset for what a real, sudden price shock looks like (the 2007-08 and 2010-11 food crises were characterized by sharp, sudden spikes, not gradual creep). The response is now front-loaded — close to linear for the poor tier from the moment the buffer is exhausted — roughly quadrupling the mortality rate at the same shortage depth compared to the previous version, while keeping the same severe-but-not-apocalyptic ~8%/yr ceiling at total collapse. The middle tier is more resilient but still responds meaningfully; the wealthy tier remains largely insulated except at extreme, near-total shortage. The fish-dependent famine mechanism (FAO: ~1 billion people reliant on fish as primary protein) remains correctly assigned to the poor tier specifically, layered on top of this general mechanism rather than replacing it.

Population feedback into energy use, fishing and land clearing: fossil energy demand and the agricultural methane baseline already scaled directly with population (a population that has collapsed consumes proportionally less energy and supports less livestock). Deforestation and fishing effort did not — they were pure policy sliders, independent of how many people actually existed to drive that demand, so a population that had crashed by 90% would still clear forest and fish at unchanged intensity. Both sliders now set a per-capita intensity; actual land-clearing and fishing pressure scale with total population relative to the 2026 baseline (8.2B) — half the population, half the clearing/fishing pressure; double the population, double the pressure. Agricultural intensification (the wind-erosion driver) is now symmetric too: it no longer floors at zero once population falls below baseline, but goes negative — less land under cultivation, less tillage, some genuine relief for soil already under stress — rather than refusing to ever improve. Tested directly: a population collapse now measurably eases fishing pressure enough for stocks to recover from near-zero once effort has fallen far enough (confirmed in a 150-year stress test where fish stock rebuilt from 0.00 to double digits as population fell below ~1B) — a real, working feedback rather than a one-way relationship. Note that the self-reinforcing tropical-dieback and heat/drought fire-loss mechanisms are deliberately not scaled by population, since those represent ecological/climate processes that continue (and can still drive full forest loss) regardless of human population once triggered — population pressure affects only the human-driven clearing component, correctly leaving the irreversible ecological tipping dynamics untouched.

Deforestation is driven by the poor tier specifically, not aggregate population: subsistence forest clearing — fuelwood collection, smallholder farming, artisanal logging — is well documented as overwhelmingly a rural/poor-developing-world activity, not something wealthy or middle-income populations drive at anywhere near the same intensity. Using aggregate population for this pressure was too blunt: a population collapse concentrated among the wealthy would barely change real-world clearing, while one concentrated among the poor would reduce it sharply. Deforestation and the land-competition dynamic below are now both scaled specifically by the poor tier's population relative to its own 2026 baseline (~2.8B), not the whole population — closing a real gap where a decline in wealthy/middle populations alone would have misleadingly eased deforestation pressure that, in reality, has little to do with those tiers.

Fishing pressure: the "chicken and egg" question, checked and closed in both directions. Fish stock collapse driving famine mortality specifically in the poor tier was already modeled (they're the ~1 billion people most dependent on fish as primary protein). The reverse question — does *population decline* reduce fishing effort in the first place, and is it the poor tier that matters here too — was checked directly and found to be a real, if different, gap: the harvest calculation was still using *aggregate* population, not scaled to reflect which tier actually drives fishing pressure. Unlike deforestation, fishing genuinely is not primarily a poor/subsistence-driven activity: small-scale artisanal fisheries — largely poor and coastal — account for only about 25-40% of global catch (Science Advances 2018; a 2025 Nature Food archetypes study), while the majority (60-75%) is industrial fishing, empirically dominated by wealthy nations' distant-water fleets serving global export markets rather than local subsistence need (Science Advances 2018 is titled, tellingly, "Wealthy countries dominate industrial fishing") — and industrial fishing drives most large-scale depletion. So simply copying the deforestation fix (tying everything to the poor tier) would have been scientifically wrong here. Fishing pressure is now blended roughly 65% wealthy+middle tier / 35% poor tier, matching that literature split: a poor-tier population decline now eases fishing pressure by a real but modest amount (the subsistence share), while a wealthy/middle-tier decline eases the dominant majority of it (the industrial share) — verified directly: halving the wealthy+middle tiers cuts fishing pressure by ~33%, while halving the poor tier alone cuts it by only ~18%, with both directions now closing the loop against the already-modeled reverse relationship (fish collapse → poor-tier famine) rather than leaving it one-way.

Land competition: humans and wild habitat compete for a shrinking pool of usable land: as climate change, desertification, and heat stress shrink the land that's actually climatically usable (tracked via the zone-based arable-land index elsewhere in this model), poor/rural demand for farmland doesn't automatically shrink to match — population sets that demand independently. A land-pressure ratio (poor population relative to its baseline, divided by climate-available land) now drives two effects: when demand outstrips shrinking supply, the gap is filled by clearing more forest — an encroachment pressure on top of the population-scaled clearing rate already in place — and when the poor population falls faster than available land shrinks, land is freed for abandonment, enabling genuine reforestation-via-abandonment (a real, documented pattern — cropland abandonment and forest regrowth following demographic decline in parts of Eastern Europe post-1990, and in depopulating rural areas more generally). This abandonment term is real and directly offsets the deforestation rate in the same equation, not merely a documented intention — verified directly: at very low poor-tier population it reaches magnitudes comparable to the tipping-point dieback rate itself.

Can tipped tropical forest ever rebound, or is it permanent? Previously, once the dieback tipping point triggered, it was mathematically permanent — the self-reinforcing loss continued forever regardless of population, reforestation policy, or anything else, with no path back. That's now been softened to match the real ecological science: reversal is possible, but deliberately stringent, requiring forest to recover well past the original tipping point (85%, not merely back to 75% — the hysteresis gap real forest-savanna systems show, since restoring rainfall-recycling needs substantially more canopy than was lost to break it) and sustained cooling back below +1.0°C — reduced human pressure alone (via reforestation-via-abandonment above) is necessary but not sufficient; the climate driver of dieback has to genuinely improve too. Tested directly, and the honest answer to "will the equatorial regions rebound or still tip into savanna": in every realistic scenario tried — including an extreme, deliberately favorable one (RCP2.6 fossil emissions, aggressive reforestation, a rapidly shrinking poor population, sustained for 400 years) — global temperature kept climbing rather than falling back below the +1.0°C threshold, because permafrost, ice-albedo, and this model's other already-active feedbacks keep adding warming regardless of human mitigation effort. Forest still tipped into savanna in that test despite every favorable condition. Rebound is real in the code and not a hard mathematical impossibility, but for all practical purposes across the scenarios this model can produce, tropical forest tipping should be treated as effectively permanent once triggered — which is exactly what the peer-reviewed literature on this tipping point (Lovejoy & Nobre and others) concludes.

Biodiversity: direct human footprint, and a real (if slow) recovery mechanism: biodiversity previously declined only as a downstream consequence of forest, fish, and coral loss plus warming — it had no way to reflect human population or land occupation directly, and critically, no recovery mechanism at all: once damaged, it could only ever decay further, never improve, even if every pressure vanished completely. Two direct human-footprint terms are added — habitat fragmentation tied to the same land-pressure ratio above (encroachment is exactly when fragmentation is most acute), and a separate disturbance term tied to population density itself (hunting, infrastructure, pollution, scaling with population regardless of whether forest has literally been cleared yet). A genuine rewilding term is also added, active specifically when land pressure eases (population falling relative to available land) — calibrated deliberately slow relative to the decay rate, consistent with the real-world asymmetry that biodiversity loss is fast and recovery is slow (extinction is irreversible even where habitat itself recovers). Tested directly: in a sustained population-decline scenario, biodiversity — previously permanently stuck at its floor once collapsed — climbed measurably (1.0 → 1.9 over 75 years of relief, continuing to rise over longer horizons), confirming genuine, working recovery rather than a one-way ratchet. That recovery is real but modest by design: other still-active pressures (particularly tipping-point-driven forest loss, which persists regardless of population once triggered) continue to compete against it, an honest reflection of how deeply interconnected these feedbacks are rather than an overstated bounce-back.

Multi-layer grey atmosphere (replacing a single-slab treatment): CO2 and methane's warming effect is no longer computed as a fixed empirical sensitivity number; it's solved through the standard N-layer generalization of the grey-atmosphere model: T_surface⁴ = (N+1)·T_e⁴, where T_e is Earth's effective radiating temperature, set by genuine Stefan-Boltzmann physics from the solar energy balance — T_e = [(S₀/4)(1−albedo)/σ]^¼ ≈ 254.6 K today, and σ·254.6⁴ ≈ 238 W/m², matching the real observed ~240 W/m² outgoing longwave flux almost exactly. N is the atmosphere's effective number of infrared-opaque grey layers; adding CO2 or methane doesn't change the solar balance (T_e stays put) — it increases N, which is what raises the surface temperature above T_e. A one-time calibration snapshot (N_ref ≈ 0.65 "layers" today) reproduces the real ~33.5 K greenhouse effect between the actual global mean surface temperature (~288.65 K) and T_e, and CO2/methane changes beyond today are converted into additional layers calibrated so their marginal sensitivity at today's baseline exactly matches the model's existing climate-sensitivity slider and the Myhre methane formula — preserving continuity with previously-tested behavior while now running the actual response through genuine T⁴ physics. The practical difference this makes: successive CO2 doublings now show authentic diminishing marginal warming (tested at +2.95°C, then +2.87°C, +2.79°C, +2.71°C for four successive doublings from today's baseline) — a real, physically-motivated compounding of CO2's own already-logarithmic forcing with the multi-layer atmosphere's own saturating response, rather than a single fixed number extrapolated indefinitely. The "Multi-layer atmosphere" chart shows a live snapshot of the resulting vertical temperature structure (surface through six illustrative layers to the top of atmosphere), genuinely colder with height, as a real atmosphere is. Water vapor and ice-albedo feedback are deliberately kept as separate additive terms rather than folded into this same T_e/layer mechanism: piping the model's albedo index directly through the global solar-balance formula would overstate the real ice-albedo effect by roughly an order of magnitude (since that formula implicitly assumes the entire planet's reflectivity changed, not just the Arctic's small share of it), so both keep their own previously-validated, real-world-calibrated forcing pathways.

Water vapor feedback, with an explicit relative-humidity ceiling: the atmosphere's own largest amplifying feedback is modeled explicitly, not left implicit, and it is now genuinely bounded rather than growing without limit. Relative humidity is tracked as its own state, relaxing toward a ~75% global-mean baseline (nudged slightly by the rate of warming, since rapid warming can transiently outpace precipitation's ability to remove the extra evaporated moisture) but hard-clamped so it can never exceed 100% — air simply cannot hold more water than saturation; the excess condenses out immediately as precipitation, a real physical limit rather than a modeling convenience. Saturation vapor pressure itself follows the actual Clausius-Clapeyron relation, e_s(T) = e_s(T₀)·exp[(L/R_v)(1/T₀ − 1/T)] (L/R_v ≈ 5423 K), rather than the ~7%/°C linear shortcut used in an earlier draft. The resulting water-vapor index feeds a logarithmic forcing term calibrated so its slope at today's baseline matches Soden & Held's (2006) empirical estimate of ~+1.8 W/m² of extra forcing per °C of warming — historically about as large as the direct forcing from doubling CO2 itself. As before, only the warming beyond today's already-realized ~1.5°C drives this term, since that ~1.5°C already has its own water-vapor response baked into the "climate sensitivity" slider (standard practice — reported equilibrium climate sensitivity already includes water vapor feedback), so counting it again from zero would double-count it.

Stefan-Boltzmann / Planck feedback — heat genuinely lost to space: converting that water-vapor forcing into temperature now uses real physics rather than a fixed constant. Outgoing longwave radiation follows the actual Stefan-Boltzmann law, OLR = σT_e⁴ (σ = 5.67×10⁻⁸ W/m²/K⁴); at Earth's effective radiating temperature of ~255 K this gives σ·255⁴ ≈ 239.7 W/m², matching the real observed global outgoing longwave flux of ~240 W/m² almost exactly. The Planck feedback parameter — how much extra heat escapes to space per extra degree of warming, the fundamental blackbody brake that keeps climate stable — is its derivative, dOLR/dT_e = 4σT_e³, evaluated dynamically each step at the current temperature rather than frozen at a fixed value. That matters physically: the brake genuinely strengthens as the planet warms further (3.76 W/m²/K at today's baseline, rising to 4.72 W/m²/K twenty degrees warmer), which is what keeps the water-vapor loop's feedback gain safely below the runaway threshold of 1 (≈0.48 at baseline, falling further to ≈0.38 at extreme warming, i.e. becoming more stable, not less) at every temperature the model can reach, not just at today's. An earlier draft of this water-vapor mechanism used the CO2 conversion ratio instead of a genuine Planck parameter, which pushed the gain above 1 and produced runaway, unphysical warming (44°C by 2126) in testing — identified and corrected before release; the current version has been stress-tested across the full climate-sensitivity range with no instability.

Monthly timestep, and a direct test of whether granularity matters: the model steps every month (DT = 1/12 years) rather than the quarterly (DT = 0.25) resolution used through most of this build, so the fast-responding water-vapor/Planck loop can approach its quasi-equilibrium within each displayed year via twelve sub-steps rather than four, instead of being forced into coarser jumps alongside much slower processes (forests, ice sheets, permafrost) that genuinely do take years to respond. Every annual rate constant in the model — deforestation, permafrost thaw, wind erosion, fish growth and harvest, coral bleaching, biodiversity decay, population growth, renewable adoption, and more — was written from early in this model's development as (annual rate) × DT specifically so it could be re-run at a different step size without recalibration; this change exercised that design directly rather than assuming it held. One genuine bug was caught doing so: two glacier melt-rate safety bounds were hardcoded as "2% of remaining mass per step" rather than scaled by DT, meaning they'd silently permitted three times more annual change at monthly resolution than at quarterly (2%×12 months vs. 2%×4 quarters) — fixed to scale with DT like every other rate-of-change term. The actual answer to whether granularity matters: barely. Running the identical 100-year scenario at monthly versus quarterly resolution changed the outcome by about 0.004°C and 0.6ppm CO2 — consistently under 0.02°C and 2ppm across every RCP mode and validation- mode setting tested. That's a reassuring result for a model built entirely on simple forward-stepping (rate × DT, added to the previous state) rather than a higher-order numerical integrator: it suggests the quarterly resolution used throughout the rest of this build was already fine enough that discretization error was never a meaningful contributor to any of the findings in this documentation, distinct from the genuine calibration and structural issues that were found and fixed along the way. The "Step +1 month" button and year/step display (e.g. "2026 M7") reflect the current step size directly; "Run simulation" still asks for a span in years and converts it to the right number of steps internally, whatever DT currently is.

Blue Ocean Event / ice-albedo feedback: Arctic September sea-ice extent starts at 4.3 million km² (2025's minimum was 4.6M km², NSIDC) and declines toward a target set by the observed Notz & Stroeve (2016) relationship of roughly −3 million km² per °C of global warming, referenced against a ~1979 baseline of 7.2 million km². A "Blue Ocean Event" (BOE) — conventionally defined as September extent falling below 1 million km², since thick multi-year ice around the Canadian Arctic Archipelago resists full melt-out — is flagged as a threshold event once crossed. The fraction of 1979-era ice already lost subtracts directly from the model's planetary albedo (open ocean reflects only ~7% of sunlight vs ~50–70% for sea ice/snow), with a small permanent extra penalty once a BOE occurs (representing the documented lengthening of the open-water season and delayed refreeze). That albedo reduction is converted into an explicit radiative forcing (W/m²) the same way methane is, then run through the model's climate sensitivity and ocean-heat-uptake lag — so shrinking Arctic ice directly, and measurably, amplifies global warming, which then melts more ice, a self-reinforcing loop calibrated in the same broad range as published estimates (Pistone et al., ~0.3–0.7 W/m² of extra global forcing at full Arctic summer ice loss).

Tropical forest dieback tipping point: per Lovejoy & Nobre (2018), the Amazon (and tropical forest biomes more broadly) can cross a tipping point at roughly 20–25% cumulative deforestation combined with warming — beyond which reduced transpiration breaks the forest's own rainfall-recycling cycle, rainfall drops, the dry season lengthens, and the forest diebacks and burns even if human clearing stops. Here, once forest cover falls below 75% of its 2026 baseline (a proxy for that ~20–25% cumulative-loss threshold), a self-reinforcing dieback loss (~1.5%/yr of remaining forest) switches on and persists regardless of the deforestation slider — representing that irreversibility. Critically, the carbon-sink strength of whatever forest survives is also halved once tipped, since degraded, savanna-transitioning canopy sequesters far less carbon per unit area than intact rainforest. Combined with continuing area loss, this is what lets the biome flip from a net carbon sink to a net carbon source — the event log flags the year it happens.

Soil carbon / microbiome loss via wind erosion: global soil holds roughly 2,500 GtC (~1,550 GtC organic + ~950 GtC inorganic), and cultivated soils have already lost an estimated 60–75% of their original organic carbon to land conversion (Lal 2004; Sanderman et al. 2017 estimate ~78 PgC cumulative loss from land-use change). The model tracks the remaining ~350 GtC of still-vulnerable agricultural soil carbon — the living microbiome and organic matter that would otherwise keep locking atmospheric CO₂ into the ground. Clearing land for large-scale agriculture removes the hedgerows, forest margins and windbreaks that hold soil in place, leaving broad open, bare-fallow fields exposed to wind — the Dust-Bowl mechanism — with warming/drought further drying and loosening soil and worsening erosion. An estimated ~25% of the resulting eroded/disturbed carbon is oxidized directly to CO₂ (published estimates for the oxidized-vs-buried split vary widely by system and timescale); the rest is carried into rivers and the sea, lost either way from the local, microbially active pool. As that pool depletes, the land's ongoing ability to re-lock atmospheric carbon weakens too — closing the loop: more land cleared for farming → more wind exposure → more soil carbon lost as CO₂ and washed away → a weaker land carbon sink → more atmospheric CO₂ → more warming and drought → still more erosion. This also closes the loop back from the crop-yield/arable-land section below: land that heat and drought have pushed out of cultivation doesn't simply sit idle — without crop cover or irrigation holding it together it desertifies (bare, wind-scoured ground with a dying microbiome, the pattern seen from the Dust Bowl to the Sahel), so last year's arable-land loss is fed back in here as an extra driver of this year's wind erosion and soil carbon loss, on top of active-farmland clearance.

Zone-resolved crop yield & shrinking arable land: the model splits the world's ~1.5 billion ha of arable land (FAO) into three latitude zones — tropical (35% of cropland), temperate (55%), boreal (10%) — because the planet isn't a uniform box: the sun strikes different latitudes at different angles, so annual solar energy per hectare (and therefore the productivity ceiling of that land) falls off with latitude, and climate change doesn't warm every zone equally either (boreal/Arctic amplification means high latitudes warm roughly 2-3x faster than the global mean, per IPCC AR6, while the tropics warm more slowly). Each zone has its own crop-temperature optimum and heat-stress ceiling (Schlenker & Roberts-style nonlinear crop heat damage); once a zone's local growing-season temperature pushes past that ceiling, land is progressively lost from cultivation to heat and drought — the arable land series tracks this shrinkage directly and separately from per-hectare yield. Boreal land can nominally gain a little area as it thaws and the season lengthens, but the gain is small and capped, since newly-thawed soils are young, thin and carbon-poor (see the permafrost and soil-carbon sections above) and the low solar angle caps how productive that land can ever be even at its own local optimum. Total food production (used for the population's carrying capacity) is yield-per-remaining-hectare multiplied by how much land is still viable at all — so losing arable land hits total food output even where the surviving land still yields reasonably well per hectare.

All figures are deliberately simplified, single-parameter approximations of far more complex real systems, intended for exploring feedback dynamics rather than forecasting.