Design Codes Under Nonstationarity

Frontier The prize is an error signal on a variable that has never had one. A drainage standard is a single number — a 25-year culvert, a 5-year storm drain, a 100-year bridge opening — chosen once and inherited by every asset built under it for the 50 to 100 years it lasts. Nothing measures whether the number was right. An engineer computes the design flow, sizes the barrel, and the file closes; the structure either survives its service life or it does not, and nobody aggregates the answer into a rate.

Established The one retrospective failure study this brief could open puts the observed rate roughly ninety times below the nominal design exceedance probability. Cook, Barr and Halling counted 92 publicly owned roadway bridge failures in New York State between 1987 and 2011 and derived an annual failure rate of 2.13 × 10⁻⁴, or one bridge in 4,700 per year, with a 95% interval from one in 6,900 to one in 2,700. Hydraulic causes account for 52.17% of those failures, which puts flood-driven bridge loss near 1.1 × 10⁻⁴ per year against a design exceedance probability of 1 in 100 per year for a 100-year opening. Frontier The two probabilities are not like for like — collapse is not exceedance, and 7 says what would make them comparable — but that gap is the margin the whole argument is fought over, and it is unmeasured.

Established The money at stake in getting the number wrong is already priced. Cook, McGinnis and Samaras costed the modelling choices behind a climate-corrected rainfall curve across 34 US cities: taking the design storm from the upper bound of the uncertainty distribution rather than its median moves a stormwater pipe by up to five standard size increments of about 75 mm each, which the paper says “would nearly double the unit cost of the pipe”. Between two defensible bias-correction techniques the median predicted change differs by more than 20 percentage points on average and by nearly 100 percentage points in places. That spread is the whole decision.

Speculative The gain accrues to whoever owns a large, well-documented asset fleet, and to almost nobody else. A state transportation department with tens of thousands of culverts of known age and known design basis can convert its maintenance records into a failure rate; a rural county with four hundred crossings and paper files cannot, and will inherit whatever number the standard-setters choose. The binding condition is stated first: none of this exists unless somebody assembles a register linking assets to their design return period and construction date, which 5 argues is the missing input rather than the missing statistics.

Established A design code converts a probability into a dimension. The chain is short and every link is a convention. A rainfall frequency product gives a depth for a duration at an annual exceedance probability; a design manual names which exceedance probability applies to which class of asset; a hydrologic method turns depth into peak flow; a hydraulic method turns flow into a barrel diameter or a bridge opening. Established The Florida Department of Transportation manual effective January 2025 is a clean specimen: 5-year for most storm drains, 10-year for Interstate and replacement work, 25-year for outfalls, 50-year for Interstate sag inlets and pump station outlets, 50-year for cross drains on facilities carrying more than 1,500 vehicles a day and 25-year below that threshold.

Established The exceedance probability is an estimate from a fitted distribution, not an observed frequency. A 1% annual exceedance depth at a given gauge is a quantile of a distribution — in North American practice a generalised extreme value, generalised normal, generalised Pareto, generalised logistic, Pearson type III, Kappa or Wakeby form, selected regionally by L-moment ratios — fitted to an annual maximum or partial duration series typically decades rather than centuries long. NOAA Atlas 14 volume 12, covering Idaho, Montana and Wyoming, used roughly 2,317 stations in the frequency analysis out of 15,328 collected, split as 1,521 daily, 566 hourly and 230 sub-hourly.

Established Stationarity is the assumption that the sample is exchangeable in time. Every volume of Atlas 14 states it in the same sentence: “The precipitation frequency analysis methods used in NOAA Atlas 14 are based on the assumption that the annual maximum series used in the analysis are stationary.” Frontier That sentence is what a nonstationary product proposes to delete. Under nonstationarity the distribution parameters become functions of a covariate — most commonly time itself, sometimes global mean temperature, sometimes a climate index — and the quantile the engineer wants becomes a quantile at a stated year.

Frontier Once the distribution moves, the return period stops being a single number. The reciprocal-of-probability definition assumes a fixed annual probability; with a drifting parameter, an asset has a different exceedance probability every year of its life, and the design question becomes a reliability question over a stated design life rather than a frequency question at a point. Several reformulations exist — expected waiting time, expected number of exceedances, and the “equivalent reliability” construction of Hu and colleagues, which fixes the reliability planned at commissioning and solves backwards for the design flood — and they do not agree with one another. 15 deals with the popular version of this confusion.

Established The record shows increases in extreme precipitation, and the operational US frequency product does not represent them. These are separate facts and the tension between them is the subject. NOAA Atlas 14 is the authoritative precipitation frequency standard for the United States, published in twelve volumes between 2004 and 2024, and every volume rests on the stationarity sentence quoted in 2. Established Its estimates are what state manuals cite by name.

Established Atlas 14 tested for trends and mostly did not find them at station level. Volume 12, version 2, published September 2024 for Idaho, Montana and Wyoming, applied “a parametric t-test and non-parametric Mann-Kendall test for trends in means and Levene’s test for trends in variance in the 1-day and 1-hour AMS data at the 5% significance level”. It reports that “neither test detected trends in approximately 90% of the stations” at both durations, that Levene’s test found no variance trend at more than 90% of stations, and that the regional null of no trend could not be rejected “for either region, except for the Northern Great Plains region at the 1-hour duration, where the relationship was statistically significant but very small”. Volume 11 for Texas, published 2018, reports the same pattern in nearly the same words.

Frontier That result is routinely read as evidence of no change, and it is not. A station-level test at 5% on a 40- to 70-year record of annual maxima has low power against the size of trend the climate literature expects; failing to reject is not evidence for the null. The opposite reading is equally available and equally unearned: a finding that 10% of stations show significant trends at a 5% level is only twice what pure noise produces. The honest statement is that the product’s own diagnostics cannot settle the question at the resolution a design code needs.

Established NOAA Atlas 15 is the first national US precipitation frequency product built on explicitly nonstationary methods. The National Weather Service describes the shift in its own words as moving “from a stationary assumption (i.e., extreme precipitation events do not change significantly over time) to a nonstationary assumption (i.e., extreme precipitation events change over time)”. Frontier It is spatially continuous rather than station-interpolated, covers durations from 5 minutes to 60 days and exceedance probabilities from 50% to 0.1%, and ships in two volumes: volume 1 gives present-day estimates that account for historical temporal change and will supersede Atlas 14, and volume 2 gives model-based adjustment factors projecting to 2100. It carries 90% confidence intervals, quoted as the range between the 5% and 95% bounds. The work was paid for by one-time federal funding under the Bipartisan Infrastructure Law.

Established The schedule has slipped by about a year and the product is still pre-publication. A Montana-only pilot was released on 26 September 2024 covering durations of 1 hour to 10 days and exceedance probabilities of 50% to 1%, with feedback sought to 31 January 2025; the pilot explicitly “did not go through the peer review process”. The Association of State Floodplain Managers reported in 2025 that contracts were paused for roughly a month and that preliminary CONUS estimates were then expected in 2025 with publication in 2026. The NOAA information page now gives preliminary CONUS in September 2026, published CONUS in 2027, preliminary estimates for Alaska, Hawaii and the territories in 2027 and publication in 2028. Established As of this writing the operational standard for every design office in the United States is still Atlas 14.

Established The same agency published a stationary and a nonstationary frequency product for the same state in the same month. Atlas 14 volume 12, version 2, covering Montana under the stationarity assumption, is dated September 2024. The Atlas 15 pilot, covering Montana under a nonstationary assumption, was released on 26 September 2024. This is the cleanest available statement of where the field actually stands: not a transition completed, but two incompatible statistical premises live at the same address on the same day. Frontier It is also, as 7 argues, the closest thing to a controlled comparison anyone has.

Established The design manuals in force in September 2026 do not use nonstationary precipitation frequency, and most do not mention climate at all. This brief read six current design documents directly. The Maryland State Highway Administration manual, published September 2023 and revised March 2026, states that “MDOT SHA uses the National Oceanographic and Atmospheric Administration (NOAA) Atlas 14 Point Precipitation Frequency Estimates for development of project hydrology”, and carries a clause headed “Projecting Rainfall Data for Climate Change” with a table of “NOAA Atlas 14 Projected Rainfall Ratios/Change Factors” applied where a facility needs added resilience; the ratio values themselves could not be retrieved from the document and are not reported here. Established The Florida manual effective January 2025 specifies Atlas 14 partial duration rainfall. The Georgia Drainage Design Policy Manual, revision 1.1 dated 3 April 2025, surfaced no climate provision in the text retrieved. The Mile High Flood District Urban Storm Drainage Criteria Manual chapter on culverts and bridges, dated November 2025, states that a “100-year basis of design for conveyance (plus freeboard above the allowable headwater) is common” for arterial and critical crossings draining 130 acres or more, and contains no nonstationarity statement in the text retrieved. Established Minnesota’s transportation department has not adopted climate-adjusted design standards; it names Atlas 15 as the expected authoritative source and points designers at an external projected-precipitation tool it says is “not fully vetted” for water resources design.

Frontier The commission’s anchor needs one correction: adoption into state drainage manuals has not happened yet. What has happened is narrower and more interesting. Manuals cite the precipitation source by name rather than by edition and date, so when Atlas 15 volume 1 supersedes Atlas 14 the design values in every manual that cites “NOAA Atlas 14” will have to be republished, and every manual that cites the current NOAA estimates generically will change its answers without any state making a technical decision. Frontier The adoption mechanism is bibliographic, not deliberative. That is the finding, and it is not the one the commission expected.

Established Florida already mandates a nonstationary treatment of sea level while keeping a stationary treatment of rainfall. The same 2025 manual requires that coastal project design “must incorporate sea level rise analysis to assess the vulnerability of flooding over the design life”, using relative sea level trend data from the National Water Level Observation Network with “straight-line extrapolation based on the design service life of the project”. Established A linear trend extrapolated over the service life is a nonstationary design method, and it is in force. Whatever the obstacle to nonstationary rainfall design is, it is not that engineers refuse to extrapolate.

Established The federal design guidance for highway drainage treats climate as a risk framing rather than a method. The Federal Highway Administration’s urban drainage design manual, fourth edition, publication FHWA-HIF-24-006, dated February 2024, states in its introduction that “Resilient and reliable design of storm drainage systems for transportation facilities is also essential in addressing the significant and growing risk presented by climate change” and, in its risk management chapter, directs designers to identify risks from “current and future environmental conditions, such as extreme weather events, climate change, seismic activity”. Frontier No nonstationary estimation procedure and no named precipitation frequency source surfaced in the portions retrieved. The honest reading is that the sentence commits nobody to anything computable.

Established The dam side of the problem is worse and much older. Spillway sizing in the United States rests on probable maximum precipitation, and the National Weather Service PMP document set runs from 1963 to 1999: hydrometeorological report 51 covering everything east of the 105th meridian is dated 1978, report 55A is 1988, report 57 for the Pacific Northwest is 1994, and reports 58 and 59 for California are 1998 and 1999. The agency’s design studies landing page states flatly that “Probable maximum precipitation updates are currently not being developed”, while the PMP page reports that “NOAA received the first-ever direct Federal funding to update PMP estimates for the entirety of the United States” with a National Academies study under way. Frontier Both statements were live on the same site in September 2026; this brief reports the contradiction rather than resolving it.

Established The statistical objection to nonstationary design is specific, published, and has not been answered. Serinaldi and Kilsby, in Advances in Water Resources volume 77, pages 17 to 36 (2015), examined nonstationary distribution functions and return periods across three case studies — Little Sugar Creek, the Red River of the North and Assunpink Creek — and concluded that where the model structure cannot be determined deductively, the uncertainty added by model complexity can negate the practical benefit, and that stationary models with an honest uncertainty assessment may be the more reliable option. Frontier The claim is not that the climate is unchanging. It is that a fitted trend in a short record is itself an estimate with a variance, and that the variance propagates into the quantile an engineer builds to.

Established The magnitude of that penalty has been measured in at least one careful sensitivity study. Xiong and colleagues, in Stochastic Environmental Research and Risk Assessment volume 38 (2024), report that nonstationary GEV models show heightened sensitivity to parameter perturbation relative to stationary ones: their entire-parameter sensitivity index under posterior uncertainty is 0.52 for the stationary fit against 1.41 and 1.30 for nonstationary fits driven by two emissions pathways — roughly two and a half times the sensitivity. Established Their 100-year design discharges over a 2000 to 2100 design life run from 8,661 m³/s under one emissions pathway to 8,945 m³/s under another, a spread of about 3% that sits inside the parameter uncertainty rather than outside it.

Established Trend detection itself is fragile in exactly the way that matters here. Bayazit’s review in Environmental Processes volume 2, pages 527 to 542 (2015), reports that under long-term persistence the Mann-Kendall test “is likely to report significant trends about half the time when there is no trend for H = 0.85–0.90”. The worked example is stark: northern hemisphere annual temperature from 1856 to 2004, a series with a real trend, returns a Mann-Kendall p-value of 2 × 10⁻²⁷ with no correction, 5 × 10⁻¹¹ under an AR(1) model, 5 × 10⁻³ under a long-term persistence model, and 0.07 under a combination of the two. Established The same data, four defensible null models, and significance survives or does not depending on which one the analyst picked.

Frontier The field’s own summary says nonstationary models win, and the criterion it wins on is the wrong one. A 2025 systematic review by Twinomuhangi and colleagues screened 1,329 articles, reviewed 181, and reports that “non-stationary models outperformed the stationary ones”. The review also reports the practice profile: 70.8% of studies used the Mann-Kendall test to detect nonstationarity, 41.4% fitted a GEV, 30.0% selected models by Akaike information criterion, and 50.5% used time itself as the covariate. Asked directly which criterion the outperformance claim rests on, the review does not consistently specify, and the dominant criteria in the corpus are in-sample information criteria. Frontier An information criterion rewards fit against the data used to fit; it is not an out-of-sample test of design value accuracy, and no part of this literature that the brief could read supplies one.

Established Milly and colleagues have answered the parameter-uncertainty objection, and the answer is a value judgement rather than a measurement. Their 2015 reply in Water Resources Research engages Koutsoyiannis and Montanari directly on long memory and on added parameters, and the operative sentence is a question: “Should we ignore the uncertainty introduced by climate change, just because it is difficult to quantify?” They argue the risk worth minimising is the type II error of missing a real change, and that thermodynamics implies the water cycle is changing whether or not a gauge record can prove it. Frontier That is a coherent position about which error to prefer. It is not a demonstration that a nonstationary fit produces a better design value, and the reply does not claim to be one.

Established Downscaling uncertainty is larger than the signal being downscaled. Cook, McGinnis and Samaras, in Climatic Change volume 159, pages 289 to 308 (2020), took a regional climate model ensemble across 34 US cities for 2020 to 2099 and found that the median predicted change in the design storm “between MOS techniques differs by more than 20 percentage points, and can differ by nearly 100 percentage points”. Translated into steel and concrete, taking the 90% confidence bound rather than the median “leads to an increase of one or two pipe sizes”, which “would nearly double the unit cost of the pipe”, and in some cases moves the design by five increments across a size ladder of 450, 525, 600, 675, 750, 900 and 1050 mm. Frontier This is volume 2 of Atlas 15 in miniature, and it is the part of Atlas 15 that design codes would have to decide whether to adopt.

Established The asset failure record that would discipline all of this exists in fragments. Cook, Barr and Halling assembled 1,745 bridge failures from 1920 to 2011, a restricted national set of 691 vehicular bridge failures from 1987 to 2011, and a New York State sample of 92 publicly owned roadway bridge failures over the same 25 years. From the New York sample they derive an annual failure rate of 2.13 × 10⁻⁴ with a 95% interval of 1.45 to 3.68 × 10⁻⁴, and project 87 to 222 US bridge collapses a year with an expected value of 128. Hydraulic causes are 52.17% of the New York failures; in the national set flood is 28.26% and scour 20.65%. Frontier Their conclusion is the uncomfortable one: “Enhanced bridge design regulations do not appear to significantly reduce bridge failures.”

Frontier Nobody has joined that record to the design basis of the assets in it. The failure rates are per bridge per year, not per bridge of known design return period per year, and they are not conditioned on construction date, catchment, or which edition of which manual sized the opening. Until that join is made, the observed rate and the design exceedance probability are two numbers about the same objects that cannot be compared. 7 treats making that join as the decisive test, and 5 explains why it has not been made.

Established The institutional fight in 2026 is about floors, not about distributions. The Federal Flood Risk Management Standard, adopted by the Department of Housing and Urban Development in 2024, requires new or rebuilt HUD-funded homes to be elevated two feet above the local flood level and widens the zones in which elevation applies. HUD has proposed rescinding it and reverting to National Flood Insurance Program minimums that the Association of State Floodplain Managers describes as “more than 50 years old”. In comments dated 8 September 2026 the association — (an interested party; it is the professional body for the officials who administer the standard) — states that “average annual flood losses have roughly doubled each decade since the 1990s, with annual losses averaging more than $40 billion so far this decade”. Its executive director argued a week later for adopting ASCE 24-24 in the 2027 International Building Code. Frontier None of this turns on whether a GEV location parameter has a time covariate.

Frontier Does a nonstationary fit produce a better design value out of sample? This is the question the whole dispute reduces to and the literature this brief could read does not answer it. The standard comparison is an information criterion computed on the same series used for fitting, which measures parsimony of description rather than predictive accuracy of a far-tail quantile. The test that would settle it is simple to state and 7 gives it; the obstacle is record length, since a 1% quantile validated on a 30-year holdout expects 0.3 exceedances.

Frontier Which covariate? Half the published nonstationary analyses use time itself, which is a proxy for an unnamed mechanism and extrapolates without constraint. Physically motivated covariates — global or regional mean temperature, dew point, a climate index — extrapolate under a scenario rather than under a ruler, but import the scenario’s uncertainty. Kim and colleagues, screening 59 GEV candidates over South Korea, found seasonal Atlantic and North Atlantic indices improved fit; whether an index fitted on one region’s record transfers to design practice elsewhere is untested.

Frontier Is the Clausius-Clapeyron expectation a prior strong enough to override the record? Milly and colleagues argue that thermodynamics tells us the water cycle is changing and that a gauge record’s silence is therefore a power problem rather than evidence. The counter-argument is not that thermodynamics is wrong but that the step from a global moisture-holding relation to the 1% annual exceedance 1-hour depth at a specific culvert is where the argument works by assertion. Handwave The version of that step which appears in policy documents — warmer air holds more water, therefore raise the design storm — skips convective organisation, storm duration, catchment response and the local moisture supply, every one of which can move the answer in either direction.

Frontier Is long-term persistence an alternative explanation or a nuisance parameter? Koutsoyiannis and Montanari’s position, which Milly and colleagues engage rather than dismiss, is that Hurst-Kolmogorov behaviour in hydroclimatic series produces apparent trends in finite samples from a stationary process. Bayazit’s figures show the effect is large enough to flip a significance verdict. The frontier question is whether long memory and forced trend are separable at all in records of the length available, and this brief found no analysis claiming they are.

Frontier What is the right reliability target for an asset whose hazard rate moves? Design life and return period are currently independent choices: a 50-year culvert is sized to a 25-year storm without anyone computing the cumulative probability of exceedance over the life. The equivalent-reliability construction fixes the planned reliability and solves for the design flood, which makes design life a first-class input and changes answers materially. Whether codes should specify a reliability rather than a return period is an open institutional question with a real cost attached, discussed further in 10.

Speculative Could the mixed-population problem be more important than the trend? Grego and Yates, in the Journal of Agricultural, Biological and Environmental Statistics, argue that annual maxima drawn from genuinely different generating processes — snowmelt against tropical storm against convective cell — break standard estimators of the 1% flood in ways that look like nonstationarity but are not, and develop robust local likelihood estimation against it using the Congaree, Illinois and Winooski rivers. If mixture rather than drift explains a meaningful share of apparent nonstationarity, then adding a time covariate is fitting the wrong model more flexibly.

Frontier Should volume 2 of Atlas 15 enter a code at all? Volume 1 is a re-analysis of the observed record; volume 2 is a projection built from downscaled climate model output. The measured spread between bias-correction methods in the published literature is of the same order as the projected change itself. A code that adopts volume 1 is adopting a contested statistical method on observed data; a code that adopts volume 2 is adopting a scenario. Those are different decisions and no manual this brief read distinguishes them.

Established Record length is the binding constraint and it cannot be bought. Atlas 14 volume 12 required a minimum of 70 years of data to test 1-day trends and had to drop the threshold to 40 years for 1-hour data to get a usable sample. A design code addresses the 1% and 0.1% annual exceedance quantiles; the record available to estimate them is typically 40 to 100 years long. Every statistical dispute in 3 is downstream of that ratio, and no method fixes it.

Established Sub-hourly gauge density is thin where culverts are dense. Volume 12 used 1,521 daily stations but only 230 sub-hourly ones across three large states. Small road crossings respond to 15-minute and 1-hour intensities, which is the duration band with the weakest observational base and the strongest claimed climate signal. Atlas 15 extends downward to 5-minute durations; the gauges underwriting those estimates are the same sparse set.

Frontier There is no register joining assets to their design basis. National inventories record what a structure is and what condition it is in, not which return period sized it or which manual edition was in force when it was designed. Without that field the retrospective test in 7 cannot be run at any scale, and the field is cheap to add going forward and expensive to reconstruct backwards. This is an archival bottleneck wearing a statistical costume.

Frontier Failure is not defined consistently enough to count. A bridge collapse is unambiguous and rare. A culvert that overtops without damage, a storm drain that surcharges for twenty minutes, an embankment that scours and is repaired next season — these are the events that a design return period is actually about, and they are recorded, if at all, in maintenance systems that were built for work orders rather than for reliability statistics. Minnesota’s transportation department is building a flood event database precisely because this record does not exist.

Established Bias correction and downscaling dominate the projected signal. More than 20 percentage points of median difference between defensible methods, and up to nearly 100 in places, is larger than the change most adaptation plans assume. Until that spread narrows, a projected design value is a choice among methods dressed as a measurement.

Frontier Professional liability points the wrong way. An engineer who designs to the code is defensible; an engineer who exceeds it on the strength of a contested statistical argument is exposed both to a client asking why the pipe is bigger and, if something else fails, to the question of why a non-standard method was used. The moral hazard argument runs in the opposite direction at the developer end. Both point at the code as the only thing that moves behaviour, which is the premise of this brief.

Established The probable maximum precipitation basis for dams has not been maintained. A design basis whose newest national document is from 1999 and whose custodian says updates are not currently being developed is a bottleneck regardless of which side of the stationarity argument is right.

Frontier Extreme event attribution has to become quantitative at design durations. Attribution statements are usually made about events at synoptic scales and daily durations. A design code needs a statement about the 1-hour 1% quantile at a point, with an uncertainty, and the attribution literature does not currently deliver that resolution.

Frontier Storm-resolving climate simulation is the only path to a physically grounded covariate. A kilometre-scale model that resolves convection rather than parameterising it would let the covariate be a simulated process rather than a calendar year. This is a compute and model-development dependency, and the work sits alongside what Earth-System Digital Twins covers; this brief does not duplicate it.

Frontier Machine-learning weather models have not yet been shown to help with far-tail design quantiles. The forecasting gains discussed in AI Weather Prediction are about skill at lead times of days, not about the 0.1% annual exceedance depth. Whether a learned emulator can generate physically consistent synthetic extremes long enough to estimate design quantiles is an open research question and a genuine dependency.

Established Historical gauge digitisation is a prerequisite nobody funds. The single largest available gain in effective record length is rescuing pre-digital charts and paper records. It is slow, unglamorous and produces exactly the resource the statistical dispute is starved of.

Frontier Reliability theory for time-varying hazard needs a code-ready form. Structural codes already use load and resistance factor design with target reliability indices. Hydrologic design does not, and the translation from a drifting exceedance probability to a partial safety factor an engineer can apply is unfinished work.

Speculative Discounting decides more of this than the statistics do. Whether to spend on a larger culvert now against damages in 2075 is a discount rate question before it is a hydrology question; Long-Run Discount Rates owns that argument. A high discount rate makes the debate economically moot for short-lived assets; a low one makes it decisive for long-lived ones.

Frontier The decisive experiment is a retrospective survival analysis on assets whose design return period and construction date are known, comparing the observed failure hazard against the design exceedance probability. Nothing else this brief examined can adjudicate between the two statistical camps, because both camps are arguing about a quantity — the reliability actually delivered by a design standard — that neither has measured. Established The study is a join, not a new instrument: take an asset inventory, add a design-basis field, attach the maintenance and failure record, fit a hazard model with construction era and design return period as covariates, and read off whether assets built to a 25-year standard fail at a rate consistent with a 25-year standard, and whether that rate has drifted with construction decade.

Established The pieces exist separately and have never been joined. Cook, Barr and Halling show that the failure side is assemblable: 1,745 failures from 1920 to 2011, a rate of 2.13 × 10⁻⁴ per bridge per year in New York, and hydraulic causes at 52.17%. What their data lack is the design side. Adding design return period and design-manual edition to an asset register is an administrative act, not a research programme.

Frontier The expected result would discipline both camps. If assets built to a 25-year standard fail hydraulically at far below 4% a year — as the bridge figures suggest, where a hydraulic rate near 1.1 × 10⁻⁴ sits roughly ninety times under a 1-in-100 design exceedance — then the codes carry a large undocumented margin, the stationary estimate is not the binding constraint, and arguing about the third decimal place of a GEV location parameter is not where the safety is. If instead the hazard rate rises sharply with construction decade after controlling for design return period, that is direct evidence of nonstationarity in the quantity that matters, obtained without fitting a single covariate to a rainfall series.

Frontier The second experiment is a split-sample validation of nonstationary against stationary design values. Take the longest available gauge records, fit both model families to the first two thirds, and score predicted against observed exceedances in the withheld third by a proper scoring rule rather than by an information criterion. This is the test the systematic review’s “outperformed” claim does not supply. It is cheap, it needs no new data, and its weakness is stated in 4: at design exceedance probabilities the expected count in a holdout is under one, so the test has to pool across many stations and thereby imports the regionalisation assumptions it was meant to avoid.

Established The third is a synthetic-truth simulation, and part of it is already reported. Generate series from a known long-memory stationary process and from a known trending process, hand both to the standard nonstationary pipeline, and count how often it recovers the truth. Bayazit’s figure — significant trends about half the time under a stationary process with H between 0.85 and 0.90 — is one cell of that table. The full table, run over the estimator chain actually used in a national product, has not been published.

Established A natural experiment is already running in Montana. Two official frequency products now exist for the same state, built on opposite premises, dated the same month: Atlas 14 volume 12 version 2 and the Atlas 15 pilot. Comparing the two grids against the subsequent observed record, and against the design decisions each would have produced for the same set of crossings, is a study that requires no new measurement and no new funding. That it has not been published is a statement about who is incentivised to run it.

Speculative A fourth test is institutional rather than hydrological. Where a jurisdiction has changed its design standard on a known date, the change is a policy shock; comparing hazard rates for assets built either side of it, controlling for exposure, would measure what a code change actually buys. The Maryland projected rainfall ratios and the 2024 federal flood standard are both candidate shocks, and the proposed rescission of the latter would supply a second edge.

Established The engineering consequence of a changed design value is a size increment, and the increments are coarse. Pipe comes in a ladder — 450, 525, 600, 675, 750, 900, 1050 mm in the sizes Cook and colleagues cost — so a 12% change in design flow may buy nothing or may buy a whole size. This quantisation is why modest differences in method can produce large differences in capital cost and also why many arguments about the third significant figure of a design rainfall are operationally empty.

Frontier Retrofit is where the cost actually lands. New construction pays the difference between two adjacent sizes. An existing crossing upsized mid-life pays for traffic control, excavation, utility conflicts and stream restoration, and those costs are not proportional to the diameter change. A code change is therefore cheap for the pipeline of new work and expensive for the inherited stock, which is the asymmetry Infrastructure Resilience treats at system level and this brief treats at the level of the standard that sizes each element.

Established Check-storm practice is the existing engineering answer and it is weaker than it sounds. A facility is designed to its nominal storm and checked against a larger one for tolerable consequences. Minnesota encourages designers to use projected precipitation estimates in check storm analysis rather than in the primary design, which is a reasonable way to use a contested number — and it changes nothing about the dimension unless the check fails.

Speculative Staged or adaptive design is the response that fits the uncertainty. Build the headwall, wingwalls and roadway embankment for a barrel larger than the one installed, so a future upsize is a barrel swap rather than a reconstruction. The engineering is straightforward; the obstacles are budget rules that do not capitalise optionality and a design standard that has no way to specify it.

Frontier Scour, not capacity, is the dominant hydraulic failure mode for bridges. With scour at 20.65% and flood at 28.26% of the national failure set, a design rainfall change propagates to bridge safety mainly through bed shear and foundation depth, not through freeboard. A code that raises the design storm without revisiting scour depth has moved the number that is easiest to change rather than the one that kills bridges.

Frontier Insurance prices risk on a different clock than codes do. A reinsurer reprices annually and can act on a contested trend without anybody amending a manual; a design code changes on a decade cycle and binds for a century. Insurance as Regulator owns the mechanism by which the faster instrument disciplines the slower one, and it is the most likely route by which a nonstationary estimate reaches a built asset without passing through a code committee.

Frontier The standards machinery is itself a governance technology. Adoption by reference and the question of who is in the room when a number changes are treated in Standards as Governance. This brief adds the case where the referenced document is a statistical product whose method is disputed by the discipline that produced it.

Frontier Hazard co-occurrence breaks the single-variable design storm. A culvert sized on rainfall alone is not sized for rain on a saturated catchment behind a burned hillslope during a high tide; Compound Climate Hazards and Wildfire Systems carry that argument, and it is orthogonal to whether the marginal rainfall distribution is stationary.

Speculative Early-warning systems are the cheap substitute for capacity. Where a design storm cannot be settled, operational detection and closure can hold consequences down at a fraction of the capital cost; the pattern is the one Earthquake Early Warning sets out for a hazard nobody proposes to design out entirely.

Frontier Coastal design has already made the move inland design has not. Coastal Defense Systems covers structures whose governing standards routinely extrapolate a trend; the Florida manual quoted in 3 shows the same agency doing it for water level and declining to do it for rainfall.

Established Nobody who sets the design standard produces the statistics, and nobody who produces the statistics sets the standard. The National Weather Service estimates precipitation frequency. State transportation departments and flood districts choose which exceedance probability applies to which asset. Model code bodies write the floor. The consequence is that a methodological change inside a frequency product propagates into thousands of design decisions without any of the bodies involved holding a debate about whether the method is sound.

Established Adoption is by reference, which makes it automatic. Maryland names Atlas 14 explicitly; Florida names Atlas 14 partial duration data; Minnesota names Atlas 15 as the expected source it has not yet adopted. A manual that cites the source by name rather than by dated edition inherits the successor product on publication. This is why the interesting institutional moment is the day Atlas 15 volume 1 is published, not the day a state committee votes.

Frontier The peer review that would legitimate the method is running on the product, not the method. The reviewer questionnaire this brief read asks whether Atlas 15 communicates uncertainty more clearly than Atlas 14, whether it enables better informed decisions, and whether it improves guidance on necessary actions. Those are usability questions. They are not the question Serinaldi and Kilsby raised, and a favourable answer to them does not address it.

Established The funding is one-time and the product needs maintenance. Atlas 15 was paid for by a single infrastructure appropriation. A nonstationary product is, by construction, one that goes stale faster than a stationary one, because its whole claim is that the distribution is moving. A one-time appropriation for a product that must be revised on a shorter cycle than its predecessor is an institutional mismatch, and the frozen probable maximum precipitation set is the worked example of what happens next.

Frontier Liability, not science, decides adoption speed. The standard of care is defined by what a reasonable practitioner does, which is to follow the manual. Until a manual changes, using a nonstationary estimate is discretionary and exposed; after it changes, not using one is exposed. The switch is binary and sits with a small number of state hydraulics engineers, a point Megaproject Governance makes about approval gatekeepers more generally.

Speculative The reform with the best ratio of value to cost is a failure register, not a better estimator. A mandatory, public, structured record of hydraulic failures keyed to design basis and construction date would make the decisive experiment in 7 routine rather than heroic. It costs a data field and a reporting rule. No jurisdiction this brief examined maintains one.

Established Design standards already allocate safety by traffic volume. Florida sizes cross drains to a 50-year storm above 1,500 vehicles a day and a 25-year storm below it. That is a defensible economic rule and it is also a rule that gives rural crossings twice the nominal exceedance probability of urban ones. Any change to the underlying frequency estimate is applied on top of this existing allocation and inherits it.

Frontier The cost of raising a standard lands hardest where the capacity to absorb it is lowest. A large urban district amortises a code change across a capital programme. A small county upsizing culverts on a fixed levy defers other work to do it. Arguments that present a stricter standard as costless are not honest about this, and arguments that present it as unaffordable are usually silent about the losses in 3.

Frontier The moral hazard argument is real and cuts in a specific direction. A developer who builds to a minimum and sells leaves the residual risk with the purchaser and the public purse. The argument that minimum standards function as a subsidy from future owners to present builders is made explicitly by the floodplain managers’ association — (an interested party) — and it does not depend on any claim about nonstationarity being true.

Speculative Over-design has victims too. Oversized culverts alter stream morphology and sediment transport; oversized detention changes downstream timing; capital spent on hydraulic margin is capital not spent on maintenance backlogs that kill people more reliably than hundred-year storms do. The framing that treats a larger pipe as strictly safer is a category error.

Frontier Intergenerational allocation is decided by an unexamined convention. Choosing a 50-year design life and a 25-year return period without computing the cumulative exceedance probability over that life is a decision about how much risk to leave to the people who will use the asset in 2070, made implicitly by two independently chosen numbers.

Established The design code is the largest single control variable in the built environment, and it is set by a small number of people who are not watched. Every culvert, storm drain, detention basin, levee and spillway in a jurisdiction traces its dimensions to a table in a manual. No other lever touches as much physical capital per page of text. The corollary is that an error in the table is replicated across the entire stock and is discovered, if at all, on a timescale longer than the careers of everyone involved.

Frontier A civilization that cannot measure its own design margins is running an uncontrolled experiment either way. The stationarity camp and the nonstationarity camp both propose to set a number that governs a century of construction, and neither can state the reliability its number delivers. This is the same failure that Infrastructure Resilience identifies for restoration time: an elaborate optimisation literature sitting on top of a quantity nobody publishes.

Frontier Lock-in is the real asymmetry. An under-sized culvert built in 2026 is in the ground in 2086. A code that is too strict wastes capital at a rate that is visible and correctable on the next cycle; a code that is too loose accumulates a liability that becomes visible only through failures and cannot be corrected without touching every asset. That asymmetry, not a claim about the climate, is the strongest argument for erring high, and it is an argument about irreversibility rather than about statistics.

Speculative The institutional memory problem outlasts the technical one. The probable maximum precipitation documents governing US dam spillways were written between 1963 and 1999 by an office that now says it is not developing updates. Whatever is decided about nonstationary rainfall in 2027 will be maintained, or not, by an organisation with a different budget and a different mandate in 2057. Designing the maintenance obligation is as important as designing the method, and it gets a fraction of the attention.

Handwave The claim that better standards will produce a resilient built environment is the step that works by assertion. Cook, Barr and Halling found that bridge failure rates in their data were roughly constant across decades of strengthening design regulation and concluded that enhanced regulations “do not appear to significantly reduce bridge failures”. That is one dataset with known limitations, and it is the only direct evidence this brief found on the question of whether changing the code changes the outcome.

These horizons track the design standard as the control variable: when the numbers in the manuals change, when anyone can say whether the change helped, and what remains unsettled.

• 10 yr: Established Atlas 15 volume 1 publishes and supersedes Atlas 14, on the current schedule in 2027 for the contiguous states and 2028 for the rest. Frontier State manuals that cite the source generically switch without a vote; those naming Atlas 14 explicitly, as Maryland does, require an amendment cycle, and the spread between the fastest and slowest jurisdictions is the thing to watch. Frontier Whether volume 2 projections enter any binding standard, as opposed to check-storm guidance of the kind Minnesota uses, is genuinely open.
• 25 yr: Frontier Long enough for the first cohort of assets designed under a nonstationary standard to accumulate exposure, and therefore the earliest date at which the survival analysis in 7 could compare cohorts rather than eras — but only if the design-basis field is added to asset registers now. Speculative Reliability-based hydrologic design, specifying a target reliability over a stated design life instead of a return period, reaches at least one national manual.
• 50 yr: Speculative The stock built under Atlas 14 stationarity reaches the end of its design life and is replaced under whatever succeeded it, making the comparison of failure rates between the two design regimes possible for the first time on a large sample. Speculative Probable maximum precipitation is either re-established on a modern basis or quietly abandoned in favour of a very low exceedance probability; the current documents will be between 75 and 110 years old.
• 100 / 250+ yr: Handwave Any claim about design practice at this horizon is a claim about institutions rather than hydrology. The defensible statement is narrow: assets being sized today under a 100-year standard are the ones whose performance will constitute the evidence, and whether anyone will have retained the records to read it is the open question, which is why Infrastructure Resilience and this brief both end at the same place — the missing measurement, not the missing model.

• Depends on the system-level failure accounting in Infrastructure Resilience, which establishes that the restoration and failure distributions this brief needs are not published anywhere; on the hazard co-occurrence treated in Compound Climate Hazards; and on the simulation capability discussed in Earth-System Digital Twins. Nothing on the map blocks the decisive experiment, which is the point: it is blocked by an absent data field, not by absent science.
• Requires (not on this map) a public register recording, for each hydraulic asset, the design return period and the manual edition that sized it, alongside its failure and repair history; a validation protocol that scores nonstationary against stationary design quantiles on withheld data rather than on in-sample information criteria; a sub-hourly gauge network dense enough outside cities to underwrite 5-minute to 1-hour design estimates, which currently rest on a few hundred stations per multi-state region; code language that references a dated edition of a frequency product so that a methodological change is a decision rather than a bibliographic accident; and a professional liability position under which designing above the code is not, by itself, evidence of departure from the standard of care.
• Enables a reliability-based hydrologic design code with a stated target over a stated design life; staged culvert and crossing designs whose headworks anticipate an upsize; actuarially grounded pricing of public infrastructure risk; and, most directly, the ability to answer whether a code change bought anything.
• Adjacent to Standards as Governance on how a referenced document becomes binding law, Insurance as Regulator on the faster-moving instrument that prices the same risk, Climate Adaptation Finance on who pays for the increment, and Coastal Defense Systems on the one design domain that already extrapolates a trend.

Handwave “Stationarity is dead, therefore stationary statistics are invalid.” This is the most widely repeated claim in the area and it is a misreading of the paper it comes from. Milly and colleagues defined stationarity as the property that “the joint probability distribution function (pdf) of [a stochastic process’s] state is invariant over time” and, in their 2015 reply, clarified that they meant the stationarity of bottom-line variables such as annual streamflow, not the analytical concept. The slogan travelled; the qualification did not.

Frontier “A nonstationary model is strictly better because the climate is changing.” The premise does not deliver the conclusion. A nonstationary GEV adds parameters estimated from the same short record, and Serinaldi and Kilsby’s result is that when the model structure cannot be derived deductively the added uncertainty can exceed the benefit. Xiong and colleagues put a number on the sensitivity penalty: an entire-parameter sensitivity index of 1.41 and 1.30 for nonstationary fits against 0.52 for the stationary one. The correct statement is that a nonstationary model is better when the trend is real, large relative to sampling noise, and correctly specified — and that whether those conditions hold at a given gauge is the contested part.

Established “A 100-year flood happens once a century.” It is an annual exceedance probability of 1%, so the probability of at least one exceedance in 100 years is about 63%, and two in successive years is entirely ordinary. Frontier This is the misconception that generates the newspaper arithmetic about three hundred-year storms in a decade, and correcting it removes a large part of the popular evidence for nonstationarity without touching the scientific question.

Frontier “Repeated 500-year events prove the standard is broken.” Mostly it proves that a 0.2% annual exceedance event somewhere in a large country is common, because the country contains thousands of quasi-independent locations. A 0.2% annual event at one gauge becomes a near-certainty across a continent in any given year. The residual after that correction is real and is what the literature argues about, but it is much smaller than the headline count suggests.

Established “The design return period is the failure rate.” It is not, and the gap is large. Exceeding a design flow means the assumed hydraulic condition is exceeded, which for most assets means ponding, overtopping or surcharge rather than loss of the structure. The New York bridge data put all-cause failure at 2.13 × 10⁻⁴ per bridge-year with hydraulic causes at 52.17% of that, which is roughly ninety times below a 1% design exceedance probability. Anyone arguing from design return periods to expected failure counts is off by about two orders of magnitude in the direction of alarm.

Frontier “Atlas 14 found no trends, so there are none.” Volume 12’s tests failed to reject the no-trend null at about 90% of stations. With records of 40 to 70 years and a 5% threshold, that is a statement about power as much as about the world. The symmetric error is also common: the 10% of stations that did reject are close to what a 5% test produces by chance, so they are not evidence of widespread trend either.

Handwave “Warmer air holds more moisture, so raise the design storm by the Clausius-Clapeyron rate.” The thermodynamic relation is not in dispute. The step from it to a point design quantile at a specific duration is, and it is usually made without stating the assumptions it rests on: that moisture supply is not limiting, that storm dynamics and duration are unchanged, and that sub-daily convective extremes scale like column moisture. Each of those is an active research question.

Established “Climate models can produce the design curve directly.” Between two defensible bias-correction techniques applied to the same regional model ensemble, the median predicted change in the design storm across 34 US cities differs by more than 20 percentage points on average and by nearly 100 percentage points in places. A number whose sign is stable but whose magnitude varies by that much between reasonable methods is not yet a design input.

Frontier “Atlas 15 tells engineers what to build.” It supplies precipitation depths with confidence intervals. The exceedance probability that applies to a given asset is chosen by a design manual, and the manuals reviewed in 3 set those probabilities by asset class and traffic volume, not by climate reasoning. A more accurate rainfall estimate leaves the allocation of safety entirely untouched.

Established “Codes update when the science updates.” They update when a reference is republished or a committee amends a table. Maryland’s manual was revised in March 2026 and still names Atlas 14. The Mile High Flood District republished its culverts and bridges chapter in November 2025 with no nonstationarity provision in the text retrieved. Georgia revised its drainage policy in April 2025 with no climate provision surfaced. The lag between a contested scientific claim and a changed dimension is measured in editions, not in papers.

Speculative “Building bigger is the safe default.” It is the defensible default and it is not free. Capital spent on hydraulic margin competes with maintenance, and the failure record attributes a majority of hydraulic bridge losses to scour rather than to insufficient opening, so a larger barrel may not address the dominant mode. The honest position is that erring high is justified by irreversibility, as 12 argues, rather than by a demonstrated frequency change.