When Your Climate Model's Horizon Exceeds the Civilization That Built It

Imagine you're funding a climate model that aims to predict the year 2500. The project will cost millions, take a decade to build, and rely on institutions—universities, agencies, maybe a whole government—that might not exist in their current form by then. That's the central gamble of long-horizon predictive modeling: the horizon of your forecast can exceed the lifespan of the civilization that commissioned it. But people still have to choose. And they have to choose soon.

This article is for the people who sign the checks, write the proposals, and defend the budgets. Not a sales pitch for any one method—just a plain look at the choices, the trade-offs, and the risks when your model's horizon stretches past your own. We'll walk through eight sections: the decision frame, the option landscape, criteria, trade-offs, implementation, risks, FAQs, and a no-hype recommendation. Each section is short enough to read in one sitting, but dense enough to matter.

Who Must Choose the Horizon—and by When

The decision makers: funders, program managers, principal investigators

Three groups hold the knife. Funders—government agencies, foundations, corporate R&D arms—set the outermost constraint by deciding what counts as a deliverable. A three-year grant naturally nudges teams toward a fifty-year horizon; a decade-long program tolerates centuries. The program managers sit between, translating money into milestones. They pick the horizon when they write the solicitation, often without realizing it. Principal investigators then inherit that choice, sometimes inheriting a mismatch they can't fix. I have watched a brilliant team spend eighteen months building a 2100-capable model only to discover their funder expected operational forecasts for 2040. The seam blew out on delivery day.

Each group delegates upward or downward—and the handoff rarely includes a warning label.

The clock: funding cycles, political terms, and model development timelines

Funding cycles run three to five years. Political terms run four to eight. A decent global climate model takes two to four years to stabilize, validate, and begin producing useful output. That sounds fine until you line them up: by the time your first data set arrives, your funder may have pivoted priorities. The administration that greenlit a four-hundred-year projection might be gone. The odd part is—model development timelines are the slowest thing in the room, yet they're almost never the thing that adjusts. Teams rush science to fit a fiscal calendar. Wrong order. The horizon should drive the schedule, not the other way around.

Most teams skip this: map the decision deadlines that will outlive your first release. If your model horizon exceeds the tenure of the people paying for it, you need explicit sunset clauses or a transition plan. Not having one is a decision too.

Consequence of delay: locking into an inappropriate horizon

Delay doesn't preserve optionality. It fossilizes the default. If no one explicitly chooses a horizon by the end of the proposal-writing phase, the team defaults to whatever the previous model used—usually fifty or a hundred years. That default then resists change because changing the horizon mid-build means restarting the physics tuning, the boundary-condition selection, the verification suite. I have seen groups lock into a fifty-year horizon at month four and never revisit the question, even when their funder openly started asking for century-scale results two years later. The catch is: by then the model architecture can't stretch without breaking. Retrofitting a fifty-year emulator to handle four hundred years of aerosol feedback is not a feature update; it's a rebuild.

A short rhetorical question worth sitting with: if you don't pick the horizon before the coding starts, who picks it for you later?

'The most expensive horizon is the one nobody chose—because you pay for it twice: once in the build, once in the forced rewrite.'

— paraphrased from a program manager who inherited a rebudget

What usually breaks first is credibility. A model that can't see past the next funding cycle feeds skepticism, not science. Choose before the ink dries on the contract. Or accept that the clock chose for you—and it rarely picks the right number.

The Option Landscape: Three Paths to the Far Future

Ensemble perturbation methods: many runs, wide uncertainty

The brute-force approach. You take your climate model—the same one that simulates next week's weather reasonably well—and you run it hundreds of times, each with slightly nudged initial conditions, perturbed parameters, even alternate physics schemes. The idea is simple: if you can't know which future will unfold, map every plausible future. I have watched teams burn three months of compute budget doing exactly this, and the result is a fat probability cone that gets wider with every decade projected. That cone is honest. It tells you the system is chaotic, that small errors compound, and that by year 2100 your ensemble may span a 6°C range. The catch is cost. Each run demands hours on a supercomputer cluster. Scale that to a thousand members for a horizon beyond 2200, and you either own an exascale machine or you wait—years. The trade-off: maximum uncertainty quantification, minimum guesswork. But the seam blows out when your civilization's infrastructure can't sustain that compute over the decades needed. Most teams underestimate cooling and power costs. They plan for 200 ensemble members and get 40 before the grant cycles shift.

Not everyone needs that many runs. But everyone needs to know what they're buying.

Inverse modeling: calibrate to past data, project forward

Here the philosophy flips: instead of exploring all futures, constrain the model with what already happened. You feed paleoclimate records—ice cores, sediment layers, tree rings—into a Bayesian framework that tunes parameters so the model reproduces the last 10,000 years. Then you project. The beauty: your uncertainty shrinks because the past acts as an anchor. The horror: the past may not repeat. A model calibrated to the Holocene's relative stability will miss abrupt shifts—ice-sheet collapse, methane burps, ocean current reorganization—that have no analog in the training data. I have seen inverse models produce beautiful, narrow cones that completely miss the 4°C jump the ensemble methods flagged. That hurts. The cost is lower than ensemble perturbation—you might need only 50–100 calibration runs—but the risk is structural: your assumptions about stationarity are wrong. The pitfall is subtle: inverse models look precise. They're not accurate when the system changes regime.

What usually breaks first is the assumption that past variability bounds future variability. Wrong order. The far future may invent surprises the past never saw.

Surrogate emulators: fast approximations of complex physics

You train a statistical model—a Gaussian process, a neural net, a random forest—on a sparse set of expensive full-model runs. Then the emulator mimics the climate model at a fraction of the cost. Run it a million times. Explore parameter space freely. Project to 2500 in an afternoon. The trick is that emulators interpolate well but extrapolate poorly. Push beyond the training range—say, to CO₂ levels twice what the full model ever saw—and the emulator drifts into fantasy. I have debugged emulators that confidently predicted ice-free Arctic winters by 2050 (true) and then an abrupt return to glacial conditions by 2150 (nonsense). The emulator was interpolating between training points that didn't exist. The fix: retrain on-the-fly with active learning, adding full-model runs only where the emulator shows high uncertainty. That works, but it requires real-time orchestration between emulator and parent model—a coupling most groups skip until they see a wild output. The cost is low initially, high later if you do it right. The trade-off: speed for vigilance. Emulators are the only path to truly long horizons—10,000-year projections, for instance—but they demand constant sanity checking against physical plausibility.

'An emulator that never disagrees with physics is an emulator that never learned anything new.'

— climate modeler, after watching his surrogate predict a frozen Sahara

Honestly — most data posts skip this.

Honestly — most data posts skip this.

One rhetorical question sits beneath all three paths: do you want a wide, honest error bar or a narrow, likely-wrong one? Ensemble methods give you the first. Inverse models tempt you with the second. Emulators let you pretend the problem is cheap—until it bites. The choice is not technical. It's a bet on how much surprise the future holds.

How to Compare: Criteria That Actually Matter

Computational cost vs. forecast depth

Accuracy tempts everyone. A 1,000-year model that runs on a supercomputer cluster might spit out beautiful decadal oscillations—but if the simulation takes nine months to finish, the policy window it was meant to inform has already closed. I have watched teams pour budget into horizon extension only to realize the output arrives too late for the IPCC cycle it was supposed to feed. The trade-off is brutal: you can push the model deeper, or you can keep it fast enough to iterate. Not both.

What usually breaks first is the calibration loop. A shallow horizon model can be tuned against historical data in hours; a deep one demands weeks of spin-up just to reach equilibrium. Most groups I've consulted with underestimate that spin-up cost by a factor of three. Worth noting: cloud credits burn faster than compute cores. The question isn't *can you simulate 500 years*—it's *can you simulate 500 years fifty times before the funding cycle ends?*

That sounds fine until you need to test sensitivity across five aerosol scenarios. Then the seam blows out.

Uncertainty quantification and interpretability

A black box that outputs a single number for 2150 is worse than no model at all. You need to know *how* the forecast spreads—the fat tails, the bifurcation points, the parameter where the system stops behaving. The catch is that most long-horizon models sacrifice uncertainty bands for runtime. They run one deterministic trajectory and call it a projection. Wrong order.

Interpretability matters differently at decade 90 versus decade 10. Early in the horizon, you can validate against paleoclimate proxies. Late in the horizon, you're extrapolating from parameterizations that were tuned for the 20th century. One concrete anecdote: a team I worked with found that their 300-year simulation produced a 2°C warming band in year 280—but when they inspected the code, the ocean mixing parameter had defaulted to a constant because the developer assumed no one would push that far. The forecast was mathematically correct and physically meaningless.

Policy relevance demands that a decision-maker can ask "what if we cut methane in 2035?" and get a distribution, not a line. Without that, the model is a toy.

'A forecast you can't interrogate is a forecast you can't trust—and a forecast you can't trust won't be used.'

— climate modeler, after a failed stakeholder briefing

Policy relevance and institutional stability

The odd part is—horizon depth is often a political problem dressed as a technical one. A 500-year projection assumes the agency that commissioned it still exists in 500 years. That assumption holds poorly for small nations, coastal municipalities, or any institution with a five-year election cycle. Most teams skip this: they optimize for scientific rigor and forget that the model's half-life must outlast the funding body's attention span.

I have seen a beautiful 200-year ocean circulation model shelved because the ministry that funded it was restructured mid-project. The new director wanted decadal outputs, not millennial. The team had no fallback—their architecture could not produce actionable short-term projections without re-running the entire spin-up. That hurts.

The fix? Design the horizon as a stack: shallow layers that produce usable output in year one, deep layers that publish in year five. Make the model useful at every checkpoint. That way, if the institution wobbles—and it will—the work survives the transition. Returns spike when you build for institutional fragility, not scientific ambition alone.

Trade-offs at a Glance: A Structured Comparison

When Each Method Fails: Worst-Case Scenarios

Every approach has a breaking point—and the failure mode tells you more than the sales pitch. Neural network ensembles trained on paleoclimate proxies? They extrapolate beautifully until an aerosol regime shift that never appeared in the Holocene record. Then the loss spikes, the confidence intervals implode, and your stakeholders ask why nobody flagged the unknown-unknown. I have seen a team burn six months on a hybrid PDE-ML model that performed flawlessly on validation—and then a volcanic winter analogue dismantled the training distribution in week one of deployment. The catch: they had prioritized horizon over robustness, and the seam blew out at year 47.

The simpler statistical baselines fail differently. Linear inverse models are cheap, interpretable, and catastrophically wrong when tipping cascades activate. They miss the knee in the curve entirely. That hurts more than a high-variance prediction because the false precision lulls you into action—wrong action. One concrete anecdote: a water authority used a 200-year linear forecast to justify reservoir capacity. The real climate trajectory bent at year 30. They had no fallback trigger. The institutional cost wasn't the model; it was the decade of sunk infrastructure.

The best possible forecast is useless if the organization that commissioned it can't survive the first major discontinuity.

— observation from a model reviewer who watched two agencies fold

Table Comparing Methods on Cost, Uncertainty, Horizon, and Institutional Risk

Let's get concrete. Here is the trade-off matrix that matters—not abstract theory, but the four dimensions that break projects. Cost means total compute + personnel over three years. Uncertainty is the 90th-percentile prediction interval width at half the target horizon. Horizon is the plausible upper bound before skill decays below random. Institutional risk measures how badly a model rewrite would hurt if the approach fails mid-deployment.

• Deep learning (Transformer + Earth System Emulator): High cost ($500k+), moderate uncertainty, horizon 80–120 years, institutional risk severe — the pipeline is brittle and the talent pool shallow.
• Statistical downscaling + regime-switching model: Moderate cost ($80k–150k), high uncertainty, horizon 40–60 years, institutional risk moderate — easier to refit but harder to explain to regulators.
• Physics-based reduced-order model (ROM): Low cost ($20k–50k), low uncertainty, horizon 30–50 years, institutional risk low — breaks gracefully, but can't stretch beyond its parameterization.
• Hybrid (ROM + ML correction): Moderate–high cost ($100k–300k), moderate uncertainty, horizon 50–80 years, institutional risk moderate — the seam between physics and learned terms is where trust frays.

The matrix collapses when you force a method outside its natural domain. I have watched teams pick the deep-learning path because it promised the longest horizon, then discovered that their institutional memory couldn't sustain the data pipeline past year two. Wrong order. The horizon you can actually support is shorter than the horizon you can technically predict.

Not every data checklist earns its ink.

Not every data checklist earns its ink.

The Middle Ground: Hybrid Approaches That Buy Optionality

Most teams skip this: design for the transition, not the destination. A hybrid model — say, a coarse PDE core with a lightweight neural correction that retrains every 10 simulated years — doesn't win any single metric. But it survives. The PDE ensures the output stays physically plausible even when the ML layer hallucinates; the ML layer catches emergent patterns that the PDE cannot represent. The trade-off is operational complexity — you now maintain two codebases and a handshake protocol that can silently drift.

The trick I have seen work: budget 20% of your compute for a "watchdog" ensemble of simpler models that run alongside the hybrid. When the watchdog's error exceeds a threshold, you freeze the hybrid and retrain the correction layer. That introduces a four-month lag for retraining and validation. It also prevents the disaster of a silently degraded forecast running for three years while everyone assumes the numbers are correct. The middle ground is not a compromise between cost and horizon — it's a hedge against your own ignorance about when the world will change faster than your training data.

Choose the method whose failure you can survive. Then build the escape hatch before you need it.

Implementation Path After You Choose

Data curation: which records to trust for calibration

You have chosen your method—statistical downscaling, hybrid ML, or full-physics emulation. Now the real work begins, and it starts with a brutal question: which data do you actually trust? Not the data you wish existed. Not the reanalysis product that extends back to 1950 with convenient gaps filled by someone else's model. The raw instrumental record for most regions barely reaches 150 years—shorter than a single tree's lifespan in a Pacific Northwest forest. For calibration at century-plus horizons, that's a sliver of signal drowning in noise. I have watched teams load the full CMIP6 archive, run their code, and declare success. Then three months later they discover their model had essentially memorized the 1980s aerosol cooling spike. It nailed historic CO2 ramp-up but failed the first test against a medieval warm period reconstruction. Painful.

So you start with paleoclimate proxies. Not as decoration—as ground truth. Tree rings, ice cores, speleothem records: each comes with its own error bars, its own dating uncertainties, its own regional biases. The catch is that no single proxy covers enough variables. A Greenland ice core gives you temperature and atmospheric dust, but nothing about ocean heat uptake. A sediment core from the Santa Barbara Basin gives you marine productivity but zero information about atmospheric circulation shifts. You splice them. You cross-validate against the handful of long instrumental records—Central England Temperature, the Nile flood heights—and you accept that your calibration window is a patchwork, not a pristine dataset. The odd part is that this mess helps. Imperfect but diverse data forces your model to learn dynamics, not artifacts.

Wrong order here kills everything downstream. If you calibrate a hybrid ML model on the modern satellite era (1979–present) and then try to extrapolate to 2100, you're essentially asking a system that has never seen a volcanic winter to predict one. The trade-off is stark: cleaner short data gives you tighter validation metrics, but dirty long data gives you robustness. Most teams skip this.

Model calibration and validation against paleoclimate proxies

Now you hold the curated proxy stack. The next step is not to train end-to-end. It's to stress test your method against known climatic shifts—the 8.2 ka event, the Medieval Climate Anomaly, the Little Ice Age. Each is a stress case your model must handle without retuning. A typical statistical emulator from the literature, trained on 1850–2020, will overfit to the last century's warming trend. Feed it an 800-year cold period and watch it panic. I have seen this happen: the output suddenly oscillates, then flatlines, then returns garbage. The fix was not a better architecture. It was forcing the model to first reconstruct the LIA from boundary conditions alone, then freezing those weights before training on modern data. That's an iterative step, not a one-shot.

Validation here uses metrics that hurt to look at: spectral coherence over decadal bands, phase alignment of ENSO-like variability, the ratio of forced to internal variance. A model that nails mean temperature but wiggles at the wrong rhythm is broken. You will rebuild it. Three, four, five times. Each iteration tightens the paleo-constraints or reveals a structural flaw—a missing feedback, a misparameterized ocean mixing scheme, an overly smooth land-surface response. The process is humbling. That's the point.

“A model that passes 20th-century tests but fails a single paleoclimate check is not a long-horizon model. It's a curve fit.”

— comment overheard at a model intercomparison workshop, 2023

Iterative refinement: updating as new data arrives

You have a calibrated model. It validates against the past millennium. Good. Now what? The horizon doesn't stop moving. Every year the observational record extends, every major volcanic eruption or ENSO extreme adds a data point your initial calibration never saw. Iterative refinement means you re-run the validation loop every 12–18 months. Not retrain from scratch—that destroys the long-memory structure you tuned. But you update the boundary conditions, check for drift in the internal dynamics, and prune proxy records that newer, higher-resolution data supersede. One concrete practice: maintain a "holdout stack" of the three most recent decades of satellite data. Never touch it during calibration. Only use it to test whether your model's predictions for the current period still hold. The first time that holdout score degrades significantly, you know your model has started wandering outside its training manifold—a clear signal to pause and recalibrate.

How do you know when to stop iterating? You don't, entirely. But a practical stopping rule: when three consecutive holdout windows each show error within your paleo-validation band, and when the model's decadal predictions stabilize rather than shifting with each new data ingestion, you can freeze a version for production. Tag it, document every proxy inclusion decision, and move to the risk management phase described next in this article. The final step is pragmatic—schedule a review cycle for 18 months out. Because the climate doesn't follow your timeline. It will break your model eventually. The question is whether you catch it first.

Risks of Choosing Wrong or Skipping Steps

Overconfident projections that mislead policy

A horizon set too long—say a 300-year model tasked with guiding next decade's coastal zoning—produces numbers that feel authoritative but aren't. Policymakers grab those smooth curves, slap them into environmental impact assessments, and suddenly a city's flood defense budget is anchored to something that cannot resolve the next El Niño. I have watched planning committees cite century-scale precipitation averages as if they were weather forecasts. The damage is twofold: bad policy gets written, and the modeling community absorbs the blame when reality diverges. That hurts. Institutional trust erodes faster than any single model run can repair.

The odd part is—the model itself may be technically flawless. It simply answered the wrong question. A perfectly tuned long-horizon diffusion model cannot tell you whether to reinforce a seawall next spring. Yet decision-makers, hungry for certainty, treat the output as universally applicable. The result? A credibility gap that takes years to close.

Wasted compute resources on inappropriate horizons

Most teams skip this: they provision GPU clusters based on the horizon they aspire to, not the horizon that fits their use case. I have seen a startup burn through six months of cloud budget training a 200-year land-use model when their actual client needed decade-scale crop rotation forecasts. The compute was not wasted in the sense of errors—it was wasted in the sense of irrelevance.

What usually breaks first is the cost-to-accuracy curve. Every extra decade of horizon demands exponentially more training tokens, finer resolution, and longer rollouts. That money could have bought ensemble runs at a shorter horizon—runs that actually help farmers pick planting dates. Instead, the team got a beautiful, useless artifact. Not yet. Not until the next funding round forced a painful retrain.

• Compute budget consumed by horizon length that exceeds any stakeholder's planning window
• No budget left for validation runs at decision-relevant timescales
• Model architecture locked into long rollouts, making mid-horizon variants expensive

Irrelevance to near-term decisions and loss of credibility

A climate model that cannot answer "Should I plant soybeans next month?" is a climate model that gets ignored. The catch is that long-horizon projects often produce their first meaningful output after the urgent questions have already been answered by simpler statistical tools. I have watched a research group present elegant 100-year projections to a water utility board that needed a five-year drought risk assessment by Friday. The board nodded politely, then went back to the spreadsheet that had never failed them.

Not every data checklist earns its ink.

Not every data checklist earns its ink.

The reputation damage is quiet but lethal. Once a modeling team earns the label "too far out," every subsequent output faces skepticism—even the short-horizon forecasts. Wrong order. You cannot build near-term credibility by leading with century-scale curves. A single line from a frustrated regulator stays with me: "These people predict the end of the century but can't tell me if my reservoir will dry up next summer."

'A model trusted for next month but ignored for next century is fixable. The reverse is not.'

— project manager, after a failed grant renewal

So the real risk is not technical failure. It's strategic misalignment that leaves the model precise, accurate, and useless—all at once. Fixing that requires you to walk the horizon choice backward from the decision timeline, not forward from the model's theoretical capability. Do that, or accept that your beautiful long-horizon work will decorate a slide deck while someone else's mediocre short-term model shapes the policy that matters.

Mini-FAQ: Common Doubts About Long Horizons

Can we validate a 500-year forecast?

No. Full stop. You cannot validate a forecast whose verifying data won't exist for half a millennium. The question itself betrays a misunderstanding of what long-horizon models actually do. They don't predict—they bound. They say: under these assumptions, the system cannot produce outcome X before year 480. Validation shifts from hit-rate to structural soundness. Does the physics hold? Are the feedback loops consistent with paleoclimate proxies? I have watched teams waste eighteen months trying to build a validation pipeline for year-2800 sea level. They ended up with a dashboard that compared their output against itself. That hurts.

What you can do: holdout tests on 50-year slices from reanalysis data, plus adversarial stress tests where you intentionally break boundary conditions. The model should fail visibly when you feed it absurd carbon paths. If it doesn't, your confidence is fake. The catch is that funders hate this answer. They want a number, a confidence interval, a pretty ROC curve. You have to tell them, flatly, that validation at century scale is about falsification, not confirmation. That conversation is often the one that kills the project.

What if the model outlasts its funder?

It will. Assume that. The average grant cycle runs three to five years. The average government climate program gets restructured every election cycle. Your horizon—if you chose one beyond 2100—will almost certainly survive the institution that paid for it. The real risk is not abandonment but orphaned code: a model no one alive understands, running on infrastructure nobody maintains.

We fixed this once by embedding a "successor capsule"—a stripped-down version of the model that a single postdoc could rebuild from scratch in six months. The full model could die; the logic couldn't. That meant documenting every parameter choice as a plain-English rationale, not a Jupyter notebook comment. It meant writing the coupling scheme in a way that didn't require the original author's brain. Most teams skip this because it feels like overhead when the funding is secure. The odd part is—the successor capsule is often what gets cited, not the flagship run. Funders change. Papers stay.

Isn't it better to just focus on 2050?

Better for what? If your goal is actionable policy for the next budget cycle, yes. 2050 models have validation data, political relevance, and clear feedback loops. They also produce exactly the kind of short-term optimization that locks in long-term failure. I have seen infrastructure decisions made on 2050 projections that commit a city to a 2070 collapse because the model horizon stopped at the mayor's term limit.

The trade-off is brutal: a 2050 focus gives you precision you don't need for decisions you'll regret. A 500-year horizon gives you signal that's real but blurry. The mistake is framing it as either/or. What actually works is running both in parallel—one high-resolution near-term model for procurement decisions, one coarse long-horizon model to check whether the near-term choices create a dead end. The long-horizon model doesn't tell you what to build. It tells you what not to build. That alone justifies its existence.

'The long model is not a prophecy. It's a guardrail. You drive toward 2050, but the guardrail keeps you from driving off the cliff on the way.'

— principal investigator, decadal prediction program, speaking at a review panel I attended

Your next action: pull your current model's horizon. If it's shorter than the lifespan of the infrastructure your stakeholders are planning, flag it. That single mismatch is more dangerous than any parameter uncertainty you're currently fretting about.

Recommendation Recap Without the Hype

Hybrid approach: short-term ensembles + long-term reduced-complexity models

If I had to bet one architecture against the next century, it would not be a single model. It would be a split: high-resolution ensembles for the next three decades, paired with stripped-down reduced-complexity models that reach toward 2100 and beyond. The high-resolution side catches feedback loops you cannot parameterize away—cloud physics, ocean eddies, the kind of chaotic noise that kills a forecast at year forty. The reduced-complexity side trades granularity for survival. It runs cheaply, it runs fast, and it can be re-run by a small team with a laptop if the data center gets flooded. That's the trade-off nobody wants to admit: fidelity now versus resilience later.

The catch is integration. Most teams bolt a fast model onto a slow one and call it a day. The seam blows out. I have watched groups spend eighteen months coupling two codes only to discover that the boundary conditions drift apart by year fifteen. You need a shared, versioned input pipeline—and someone whose job description includes "will maintain this for a decade." Otherwise the hybrid is just two unrelated projects with matching logos.

When to lean one way or the other

Short-term ensembles dominate when you answer operational questions: should this coastal city evacuate next hurricane season? What reservoir release schedule minimizes flood risk through 2035? Here, reduced-complexity is a liability. You lose the spatial detail needed to defend a decision against a city council. But when the question is "what atmospheric carbon load triggers ice-sheet collapse in the Amundsen Sea?"—that's a long-horizon problem. Resolution becomes noise. The reduced-complexity model, tuned to the right slow variables, gives you a probability surface. The high-resolution model gives you a pretty movie of things already breaking.

The wrong tool for the horizon is worse than no tool at all—it produces confident falsehoods dressed as science.

— modeler who rebuilt his team's pipeline twice before getting the horizon split right

Final advice: invest in institutional continuity

Models outlive people. That sounds obvious. What usually breaks first is not the code—it's the person who knew why a particular parameter was clamped at 0.74 instead of 0.75. They leave. The model runs but nobody trusts it. Then it gets forked, re-tuned, broken. I have seen three separate climate groups lose seven years of horizon work because the original architect retired and the documentation assumed the reader already knew the unwritten rules. Fix this by forcing rotation: every module must be explainable to a competent graduate student in two afternoons. Not catchable. Explainable. And the budget line for "model maintenance" should be line-item veto-proof, not the first thing cut when funding tightens.

Wrong order? Hire the archivist before the modeler. Not yet. But if you start now, the model that runs in 2040 will still have someone alive who remembers why it was built that way. That matters more than the next parameterization scheme.