Historical Megaprojects

Everyone arguing about infrastructure eventually reaches for the same three projects. The Empire State Building went up in thirteen and a half months. Hoover Dam finished more than two years early at close to its bid price. The Panama Canal opened two years ahead of its target date. The argument they are used to make is that something has been lost — that megaprojects used to be delivered better than they are now.

That claim is unusually testable, and it has already been tested. This brief is about the test, about what the three canonical examples look like when the columns their advocates omit are put back, and about a deeper problem that most of the argument never reaches: whether the comparison between eras can be measured at all.

The modern outturn record belongs to Megaproject Governance — the iron law, the one-in-ten figures, reference-class forecasting, and the error-versus-incentive question. That brief says explicitly that it does not cover the history of individual megaprojects. This is that slot, and its two jobs are the historical comparison and the measurability of the historical comparison.

The result, stated once at the top because everything else is detail: the null hypothesis that year of decision has no effect on cost-estimate error cannot be rejected. That kills the nostalgia and the progress story symmetrically. The past was not better. It was the same — which makes the nostalgic framing an empirical error rather than a difference of values.

Established The direct test exists, it is in the field's foundational paper, and its sample begins in 1927. Flyvbjerg, Holm and Buhl, Underestimating Costs in Public Works Projects: Error or Lie?, Journal of the American Planning Association 68(3), Summer 2002, pp. 279–292. The sample is 258 transportation infrastructure projects worth US$90 billion, completed between 1927 and 1998 — a span that covers the entire golden age the nostalgic framing appeals to. Established Cost escalation by type: rail 44.7% (sd 38.4), fixed links — bridges and tunnels — 33.8% (sd 62.4), roads 20.4% (sd 29.9).

Established And then the sentence that settles the question this brief exists to answer. “The null hypothesis that year of decision has no effect on the difference between actual and estimated costs cannot be rejected (p=0.22, F-test).” “We therefore conclude that cost underestimation has not decreased over time.” Established Read the direction of the test carefully, because it is the whole finding. Flyvbjerg tested whether things had got better and found no trend. That result is symmetric. It refutes “they built better then” and it refutes “we have learned” with the same F-statistic. The 2002 paper does not say the past was worse. It says the past was the same, and that is what converts a nostalgic mood into an empirical error.

Established The paper's causal conclusion is harsher than the descriptive one, and it is the part that most often gets quoted without the first part. “Underestimation cannot be explained by error and is best explained by strategic misrepresentation, that is, lying.” Its policy line follows: readers “who value honest numbers should not trust cost estimates and cost-benefit analyses produced by project promoters and their analysts.” Frontier The error-versus-incentive question itself belongs to Megaproject Governance, which holds that the relative weight of the two mechanisms is unresolved. What is owned here is the historical reading: if strategic misrepresentation is the dominant mechanism and it has been time-invariant since 1927, then it is not a modern pathology of modern approval processes.

Established The single worst overrun in the canonical catalogue is a nineteenth-century project. The Suez Canal, opened 1869, came in at a 1,900% cost overrun. The Scottish Parliament Building, the standing modern byword for the genre, runs at 1,600%. Frontier That figure reaches this brief through the author's own summary of his database rather than through a primary accounting of the canal's finances, which is why it carries the flag it does — but the direction is not in dispute and no modern project in the catalogue exceeds it.

Established Now the three canonical examples, with the columns their advocates cite and the columns they do not. Start with the Empire State Building, because it is the field's favourite counterexample and it fails on a criterion nobody mentions. Construction started 17 March 1930; the building opened 1 May 1931 — about thirteen and a half months, within a budget of $50 million, on land bought for $16 million in 1929, at $7.93 per square foot across roughly 2.8 million gross square feet. That is real and it is impressive and this brief does not dispute any of it. Established The benefit side is the part that vanishes from the telling. The owners “did not make a profit until the early 1950s.” The upper half went unrented. It was nicknamed “The Empty State Building.” Barr and Ahlfeldt describe “the vacant offices and negligible cash flow.” It sold for $51 million in 1951, at which point the New York Times called it “one of the most profitable single building operations in the world.”

Frontier Score it on the three criteria the field actually uses and the picture changes. On budget. Dramatically early. And roughly twenty years late on benefits. Under the iron law's own definition — over budget, over time, under benefits — the Empire State Building is a partial success, and the fact that it is the standard counterexample to a literature whose central finding is benefit shortfall is itself informative. Frontier The strongest version of the other side deserves stating at full strength, and it is Jason Barr's: “to blame the developers for failing to see the arrival of the Great Depression is moving the goalpost,” and the building “has been consistently profitable for most of its 90-year history” at a long-run return of about 5.3% average annual growth. Frontier Barr is an academic urban economist whose broader project defends skyscraper economics — an interested party whose interest runs against the “Empty State Building” material he reports, which raises its weight.

Established Hoover Dam is the strongest case the nostalgic framing has, and it is also the clearest illustration of why the comparison fails. Six Companies, Inc. won with a bid of $48,890,955; construction began in early 1931; the dam was turned over to the federal government on 1 March 1936, more than two years ahead of schedule, with the first concrete pour on 6 June 1933 running eighteen months early. Early, at bid price, on a project of that scale — nothing in this brief takes that away. Established And then the accounting. 112 deaths were reported as associated with construction — 91 Six Companies employees, three Bureau of Reclamation employees, one visitor. A further “42 workers were recorded as having died from pneumonia and were not included in the above total,” with workers alleging at the time that these were misclassified carbon-monoxide deaths from the diversion tunnels. An independent modern figure puts 37 tunnel workers dead from inhaling carbon monoxide released by idling gasoline engines. Frontier The recorded death toll of the framing's best example depends on a classification decision that the workforce disputed at the time, made by the contractor being paid.

Established The labour record on the same project is the second omitted column. In August 1931 the company cut tunnel workers' wages; when workers presented demands, management laid off the entire workforce and gave them until 5 p.m. to vacate. The strike ended with a guarantee of no further wage reductions and none of the workers' demands met. Families initially housed themselves in camps such as “Ragtown.” Frontier That is part of the production function that delivered the schedule, and it is never in the column that gets compared to a 2026 project.

Established The Panama Canal is a survivorship-bias machine sitting on a single site, and it is the most useful of the three. The French effort ran 1 January 1881 to 15 May 1889, cost US$287 million, killed an estimated over 22,000 people — as many as 5,000 of them French citizens — and failed completely. The American effort ran 4 May 1904 to 15 August 1914, cost “almost $500 million,” killed approximately 5,600 workers, and completed two years ahead of the target date of 10 June 1916. Established Of the American-era dead, “the great majority were West Indian labourers, particularly those from Barbados,” and “the number of Americans who died was about 350.”

Frontier The same canal, at the same place, was attempted twice, and only one attempt is in the reference class. The attempt that killed twenty-two thousand people and went bankrupt is not what “Panama Canal” means in the sentence “they built better then.” The remembered project is the successful second attempt, and the reference class has been silently pruned to n=1. Frontier And the differential mortality tracking — 5,600 dead, of whom about 350 were Americans, recorded that way in the sources themselves — tells you what the historical record was set up to count.

Established Which brings the decisive methodological point, and it is the largest single change between the eras and the one that is never priced. US construction fatality rates: roughly 150–200 deaths per 100,000 workers in the 1930s and 1940s, against 13–15 per 100,000 in 2023. A decline of more than 90% since 1940, averaging 3.2% a year, against 2.6% in manufacturing. For scale: in 1912 ironworkers suffered “109 deaths due to accidents, out of a membership of 10,928” — about 1,000 per 100,000, against roughly 27 per 100,000 for ironworkers today. Overall US workplace fatalities in 1908 ran around 61 per 100,000. Established The mechanism is documented: forty-four states passed workers' compensation laws between 1911 and 1921, and “the sharp rise in accident costs that resulted from compensation laws and tighter employers' liability initiated the modern concern with work safety.” Steel-industry fatalities fell from 0.40 per million man-hours in 1910–13 to 0.13 in 1937–39. The Hawk's Nest Tunnel of 1936, where 476 of 3,000 tunnel workers died of silicosis, is the outer bound of what the old regime tolerated.

Frontier So comparing a 1931 project with a 2026 project is comparing two production functions that price human life more than an order of magnitude apart. A 1930s schedule was bought partly with a fatality rate that is now illegal. Any honest “they built faster then” has to carry the clause “at roughly ten to fifteen times the death rate per worker” — and once that clause is attached the comparison stops being about competence and becomes a comparison of what each era was willing to spend. Established What this does not license is the claim that safety explains historical speed. No source fetched for this brief decomposes historical schedule performance into safety, regulation and other factors. The mechanism is a strong hypothesis with a large measured input; the contribution is unmeasured.

Established The open questions here are mostly measurement questions, and the honest headline is that the comparison the framing demands is only weakly measurable in principle. Five problems, in rough order of severity, and each of them is a live research question rather than a caveat.

Frontier One: survivorship in what is remembered. The three canonical examples were selected because they succeeded. The French Panama attempt is the visible control at the same site and it is not counted. Speculative Nobody has assembled a reference class of historical megaprojects that includes the abandoned and the bankrupt, and until someone does, the nostalgic sample is a selection on the dependent variable.

Frontier Two: the baseline was never measured. Megaproject Governance records that outturn data collection in a schema comparable across projects and jurisdictions is mostly absent today. It was more absent in 1931. Flyvbjerg's sample reaches back to 1927 only for transport projects that had recorded estimates at all. Speculative The set of historical projects with a documented pre-commitment estimate is a biased subset of historical projects, and the direction of that bias is not known.

Frontier Three: the estimate being compared against is not the same object across eras. A 1931 contractor bid, a 1975 authorisation estimate and a 2020 business-case P50 are three different artefacts produced under three different accountability regimes for three different audiences. Speculative Calling all three “the estimate” and differencing them against outturn is the central methodological hazard of the whole comparison, and no paper fetched for this brief addresses it directly.

Speculative Four: the regulatory, environmental and consultative baseline changed, and nobody has quantified it. Scope that was simply absent in 1931 — environmental review, seismic standards, accessibility, archaeology, public inquiry — now sits inside the project and inside the cost. Handwave This is asserted everywhere and quantified nowhere in any source fetched for this brief. It is flagged at the weakest available strength for that reason, and the flag is the finding: one of the most confidently repeated claims in the “we have forgotten how to build” literature has no measured magnitude attached to it.

Frontier Five: benefits were rarely forecast, so benefit shortfall is largely unmeasurable historically. The Empire State Building is unusual precisely because occupancy is a public number, which is why its twenty-year benefit lag is visible at all. Speculative For most historical projects there is no forecast to compare against, which means the criterion on which the modern record looks worst is the criterion on which the historical record is emptiest. That asymmetry alone could generate a spurious impression of historical superiority with no change in underlying performance.

Speculative And one testable hypothesis worth putting in print because nobody has tested it. If historical projects were forecast less formally and approved by fewer people, the strategic-misrepresentation channel had fewer places to operate — which predicts smaller historical overruns. The 2002 F-test says that prediction fails. Speculative That is mild evidence that optimism bias is the more time-invariant of the two mechanisms, and it is a cleaner test of the error-versus-incentive split than anything currently in the literature. Offered as a research question, not a finding, and the design would need historical approval-process data that this brief has not established exists.

Established The binding constraint on this subject is data, and it is a constraint on the question rather than on any technology. There is no instrument to build and no discovery to wait for. There is a historical record that was never assembled for this purpose and largely cannot be retrofitted.

Established First bottleneck: comparable outturn data across eras does not exist in a common schema. Flyvbjerg's 258-project sample is the largest thing of its kind reaching back to 1927, and it is transport infrastructure only — rail, roads and fixed links. Established Buildings, dams and canals are outside it. That is a direct limit on this brief: the Empire State Building, Hoover Dam and the Panama Canal are the three examples the argument runs on, and none of them is in the sample that tests the argument.

Frontier Second: the historical fatality record is contested at the project level even where the aggregate rate is solid. The national series — 150–200 per 100,000 in the 1930s and 1940s against 13–15 today — is well sourced. The project-level counts are not. Hoover's 112 excludes 42 pneumonia deaths that workers disputed at the time. Frontier No defensible Empire State Building construction fatality figure was obtained for this brief. The commonly cited official number is five, with higher figures in circulation, and no reliable source was found. This brief therefore prints none.

Frontier Third: benefit forecasts barely exist before the late twentieth century. The criterion on which modern projects perform worst is the one for which there is almost no historical counterfactual, which biases any era comparison in a direction nobody has estimated.

Speculative Fourth, and least tractable: the counterfactual scope problem. To price what modern regulatory, environmental and consultative requirements add, someone would have to specify what a 1931 project would have cost under 2026 rules or vice versa. Handwave That is a counterfactual construction, not a measurement, and treating any number produced that way as an observation is the error the whole “scope creep” argument rests on.

Established Fifth, a bottleneck that is nobody's fault and worth naming anyway: the surviving record is the promotional one. Photographs, opening-day coverage, commemorative volumes and company histories survive; internal cost variances, contractor claims and injury registers mostly do not. Frontier The historical record of megaprojects was largely produced by the people who built them, which is precisely the sourcing condition the field's own foundational paper warns against.

Established This brief depends on Megaproject Governance for the modern half of the comparison, and the dependency is clean because that brief cedes project history explicitly. FR-VII-19 owns the iron law, the one-in-ten and one-in-a-thousand figures, the benefit-shortfall record, reference-class forecasting, the error-versus-incentive question and the accountability gap. Without those, “better than now” has no “now” to be measured against.

Established It depends on one paper for its central result. The 2002 F-test is the direct test of this brief's framing and there is no independent replication of it fetched here. Frontier That is a single-source dependency on the load-bearing claim, and it should be visible. The mitigation is that the result is a null — the paper failed to find a trend it was well powered to find — and null results of that shape are harder to manufacture than positive ones.

Established It depends on labour and safety history for the omitted column. The construction fatality series and the workers' compensation mechanism come from independent sources with no stake in the megaproject argument — a practising engineer writing quantitatively, and a peer-reviewed economic history encyclopaedia. Frontier Sources with no interest in the nostalgia question are the most valuable kind available here, because almost everything else in this literature is written by someone making the argument.

Frontier It depends on encyclopaedic sources for all three canonical project chronologies. Dates, bid prices, death tolls and completion figures for the Empire State Building, Hoover Dam and the Panama Canal are taken from general-reference entries rather than from project archives. Nothing in them is contested, and none of them is primary.

Established And it depends on being allowed to declare a tie. The finding is that the record shows no trend in either direction. Frontier That is a terminal position, not a failure to reach one, and it is only publishable because the same test that refutes the nostalgic claim also refutes the progress claim — which is a stronger and more useful result than either side wanted.

Established The experiments this subject needs are archival and statistical, and the most valuable one has a specification that could be written today.

Frontier One: extend the F-test outside transport. The 2002 sample is rail, roads and fixed links. The three examples that carry the popular argument are a building, a dam and a canal. A comparable outturn series for buildings, dams and canals spanning the same period would test the claim on the projects the claim is actually made about, and it does not exist. Speculative A predicted result is worth recording in advance so it can be wrong: no time trend, wider variance than transport, and a fatter tail on the nineteenth-century end.

Frontier Two: assemble a historical reference class that includes the failures. The French Panama attempt is the model: same site, same objective, twenty-two thousand dead, bankrupt. Speculative An honest historical base rate requires the abandoned projects, and constructing one is a documentary exercise rather than a statistical one — the hard part is finding the projects that left no monument.

Frontier Three: obtain a defensible Empire State Building construction fatality figure, or establish that none exists. This brief could not, and prints no number. Established Either outcome is publishable and the second is the more interesting: a project cited constantly as proof of historical delivery excellence, whose human cost cannot be established from the sources in general circulation.

Frontier Four: resolve the Hoover Dam count. 112 recorded, 42 pneumonia deaths excluded, workers alleging carbon-monoxide misclassification at the time, and an independent modern figure of 37 tunnel deaths from carbon monoxide. Bureau of Reclamation and contractor records are the route. Speculative The value is not the number. It is that a disputed classification made by the paying contractor is the kind of artefact that any era comparison has to handle, and handling it once properly would show how.

Speculative Five, the design nobody has attempted: test the strategic-misrepresentation hypothesis against approval-process complexity over time. Code historical projects by how many approving bodies saw the estimate, and regress overrun on that. Handwave Whether the historical approval-process data needed to do this survives is not established anywhere in this brief, which is why the design is offered as a question and not as a programme.

Established The engineering requirement this brief cares about is measurement infrastructure, not construction technology. The construction question — whether modern methods are better — is not seriously disputed by anyone; the argument is about delivery, and delivery is measured or it is not.

Established What a usable historical comparison would require is a schema. For each project: the date of decision, the estimate as it stood at that date, what kind of artefact that estimate was, the outturn cost in comparable terms, the schedule at decision and at completion, the forecast benefits if any, the realised benefits, and the fatality count with its classification rules. Frontier No such schema exists for the historical record, and Megaproject Governance records that it barely exists for the modern one.

Established The one column that can be reconstructed reliably across eras is the safety one, and it is the one the argument omits. National fatality rates per 100,000 workers exist as a continuous series from the early twentieth century, and the change is enormous: 150–200 to 13–15, more than 90% since 1940. Frontier That is the single most defensible cross-era quantity available to this argument, and it is the one that most damages the nostalgic reading.

Frontier The engineering achievements themselves are not in dispute and this brief does not diminish them. Thirteen and a half months for a 2.8 million square foot tower at $7.93 per square foot. A $48.9 million dam handed over more than two years early. A canal completed two years ahead of a 1916 target. The framing does not fail because its examples are false. It fails because of what has been left out of the comparison.

Speculative And one engineering-adjacent observation worth flagging weakly. The three canonical projects were all delivered by a small number of contractually dominant firms under conditions of extreme labour supply — the Depression for two of them — and with scope fixed early by an owner who could impose it. Handwave Whether that combination is a reproducible delivery model or a description of a labour market that no longer exists is exactly the question the record cannot answer, and asserting either is asserting a counterfactual.

Established The seam with Megaproject Governance is the defining boundary of this brief and it is drawn by era, not by strength of claim. FR-VII-19 owns the modern outturn record and the machinery built on it: the iron law, the one-in-ten and one-in-a-thousand figures, the rail demand-shortfall and ICT fat-tail results, reference-class forecasting, the error-versus-incentive question, and the accountability gap. That brief states in its own text that it does not cover the history of individual megaprojects. This brief owns the historical comparison and, more importantly, whether it can be measured at all.

Established There is exactly one legitimate overlap and it should be used rather than avoided. FR-VII-19 notes that the overrun rate “has stayed high and roughly constant across seventy years of comparable data.” That sentence is the 2002 F-test. This brief owns the historical reading of it — that it refutes nostalgia and progress symmetrically — and does not restate the governance argument built on the same result. Frontier The cross-link is worth making in both directions, because a reader who arrives at the governance brief wanting to know whether things used to be better is being sent here.

Established The second seam is with the historical space-settlement material in this category, and it is a shared instance rather than a shared subject. The National Research Council's 1981 finding that solar-power-satellite cost estimates were understated “by a factor of about 2 to 8” is a documented case of sponsor-forecast optimism, produced before commitment, on a project never built. Frontier Pre-commitment cost reviews are the rarest category of evidence in this literature, because the projects that fail them do not get built and therefore never enter an outturn database. That is a selection problem the megaproject record has and does not discuss.

Established Third, the delivery-technology seam. Automated Construction Systems owns the question of whether construction productivity is a machinery problem, and its own finding — that the case turns on organisational structure rather than equipment — is the technological counterpart of this brief's historical one. Frontier Both arrive at the same place from opposite directions: the binding variable is not the tooling.

Speculative And a seam this brief stays out of deliberately. The political theory of infrastructure — who decides, on whose authority, with what legitimacy — belongs to the governance category. Handwave It is worth naming the exclusion because the “we have forgotten how to build” argument is usually a political argument wearing a delivery-performance costume, and separating the two is most of what this brief is for.

Established The institutional finding is that the historical record was produced by institutions with no obligation to be comparable, and the modern record is barely better. A 1931 contractor bid existed to win a contract. A 2020 business case exists to obtain approval. Neither was produced to be differenced against outturn twenty years later, and treating them as the same instrument is a category error the whole comparison rests on.

Established The single largest institutional change between the eras is the one this argument omits, and it is legal. Forty-four states passed workers' compensation laws between 1911 and 1921, and the resulting rise in accident costs “initiated the modern concern with work safety.” Federal steps followed: the 1893 Safety Appliance Act, the Bureau of Mines in 1910, OSHA in 1970. Established The fatality rate fell by more than 90% from 1940, at about 3.2% a year, and it fell because liability moved. Frontier An institution that internalised the cost of killing workers changed the production function of construction, and that change is inside every schedule comparison anyone makes between 1931 and now.

Established Second, the counting institution and the paying institution were the same one, on the framing's best example. Hoover Dam's 112 recorded deaths exclude 42 pneumonia cases that the workforce alleged were carbon-monoxide poisoning from the diversion tunnels, and the classification was made by the contractor. Frontier Any historical performance figure produced by the party being measured inherits that problem, and the megaproject literature's own foundational advice — do not trust numbers produced by project promoters — applies with more force to 1936 than to 2026.

Frontier Third, the labour institutions of the period are part of the delivery record and are never counted as such. A workforce laid off in its entirety and given until 5 p.m. to vacate, in a Depression labour market, with none of its demands met, is a component of how a dam came in more than two years early. Speculative Pricing that component is not possible from any source fetched here, and its absence from the comparison is a choice rather than a limitation.

Established Fourth, the institution that would settle this does not exist. No body collects megaproject outturn data in a schema comparable across projects, jurisdictions and eras. Frontier That is recorded as an absence by Megaproject Governance for the present day, and this brief simply notes that the absence extends backwards — which means the historical half of every era comparison is assembled by individual researchers from whatever survived.

Speculative And fifth, the institutional reason the nostalgic claim persists despite being tested. The test is in a 2002 planning journal; the claim is in newspapers, books and political speeches. Handwave A null result in a specialist journal does not propagate, and there is no institution whose job is to make it propagate — an observation about how evidence moves, not a measurement of why this particular claim survives.

Established The first ethical point is arithmetical and it is the core of this brief. The three canonical projects were delivered in a labour-safety regime running at 150–200 construction deaths per 100,000 workers against 13–15 today. Frontier Praising historical delivery performance without that clause is praising a trade that was made on other people's behalf — and the people it was made on behalf of were, at Panama, overwhelmingly West Indian labourers, of whom roughly 5,250 of the American era's 5,600 dead were not American.

Established Second, the classification question is an ethical one and not only a statistical one. Forty-two workers at Hoover Dam died and were recorded as pneumonia cases outside the official total, against workforce allegations at the time of carbon-monoxide poisoning in the diversion tunnels. Frontier Whether those deaths count is the difference between 112 and 154, and the decision was made by the party paying for the project. Reporting the headline figure without the exclusion is repeating a contractor's accounting as history.

Frontier Third, the ethics of the comparison itself. An argument that the past delivered better, made without the fatality column, is not merely incomplete — it recommends a production function whose principal input has been made illegal. Speculative Nobody making the argument intends that, which is exactly why the omission matters: the recommendation is carried by the omission rather than by anything anyone asserts.

Established Fourth, the honest counterweight, because this brief owes it. Modern scope — environmental review, accessibility, consultation, archaeology — is not overhead. It is the cost of obligations that societies chose to take on, and much of it protects the same people the old fatality rates killed. Handwave Whether it is efficiently discharged is a separate question that no source fetched for this brief measures, and the “scope creep” claim remains the least evidenced assertion in the whole argument.

Speculative And fifth, a point about benefits and who bears their absence. The Empire State Building's twenty-year benefit lag fell on its owners, which is why it is a footnote. Frontier Benefit shortfall on public infrastructure falls on taxpayers and on the users who were promised the service, and the historical record's near-total absence of benefit forecasts means that side of the ledger is invisible for exactly the projects where it would have mattered most.

Established The durable finding is a null, and nulls of this kind are rare and valuable. Across 258 projects completed between 1927 and 1998, year of decision has no detectable effect on cost-estimate error, p=0.22. Seventy-one years of construction, war, computing, project management as a profession, and two generations of reform, and the error did not move. Frontier That is a stronger statement about the stability of a human institutional failure mode than almost anything else on this map.

Established The second durable point is that the argument is symmetric and both sides lose. Nostalgia loses because the trend is not there. Progress loses for the same reason. Frontier A civilization that wants to claim it has learned to build has the same evidentiary problem as one that wants to claim it has forgotten, and neither claim has been demonstrated on the data that exists.

Established Third, the largest measured change between the eras is not delivery competence. It is the price of a worker's life. More than 90% since 1940, driven by liability moving onto employers through forty-four state compensation statutes in a decade. Speculative Read at civilizational scale, the twentieth century's most consequential change to how large things get built was a legal one, and it happened in the 1910s — which is not where anyone looks for it.

Frontier Fourth, the survivorship lesson generalises well beyond infrastructure. The Panama Canal is the same objective at the same site attempted twice, and only the successful attempt is in the reference class the argument uses. Speculative Any “the past was different” argument is a claim about a sample, and the sample is almost always what survived to be remembered. This brief is one worked instance; whether the pattern is general is not established here.

Speculative And fifth, the uncomfortable implication for ambition. If the overrun rate has been time-invariant for a century, then large projects have always cost substantially more than promised and always will, and every civilization that built anything large did so while being wrong about the price. Handwave That could be read as a reason for caution or as a reason to stop pretending the estimate is a forecast and budget accordingly — and the record supports the second reading no better than the first.

These horizons track whether the historical comparison becomes measurable, not whether construction improves — the delivery question has been flat for a century:

• 10 yr: Frontier The cheap archival work either happens or does not: a defensible Empire State Building fatality figure or a formal statement that none exists, and a resolution of the Hoover Dam count against Bureau of Reclamation and contractor records. Frontier Expect the 2002 F-test to remain the only direct test of the era claim, because extending it outside transport requires assembling a series nobody is funded to assemble. Speculative Expect the “we have forgotten how to build” literature to continue at volume regardless, because its claim is about a mood and its refutation is in a planning journal.
• 25 yr: Speculative The plausible split is that modern outturn data improves where a jurisdiction mandates post-completion evaluation, and the historical half stays exactly where it is, because the documents either survived or did not and no future policy reaches backwards. Speculative If any government does mandate comparable outturn reporting, the interesting consequence for this brief is that the era comparison becomes asymmetric in the opposite direction — a well-measured present against an unmeasurable past. Handwave Forecasting which jurisdictions do that is forecasting legislation.
• 50 yr: Speculative At this range the only thing that would genuinely change the answer is a historical reference class that includes the abandoned projects, and building one is a documentary programme rather than a statistical one. Speculative The safety baseline is the one variable with a clear direction: it has fallen 3.2% a year for eighty years and there is no evident floor in the series. Handwave Extrapolating that rate is extrapolating a trend, which is the move this brief spends most of its length declining to make about everything else.
• 100 / 250+ yr: Handwave Beyond useful forecasting, and the only honest content at this horizon is a warning drawn from the brief's own subject. Handwave The Suez Canal opened in 1869 at a 1,900% overrun and is remembered as a triumph; whatever is built this century will be remembered by whichever documents survive, and the surviving documents will be the promotional ones.

• Depends on Nothing on this map produces a result this brief is waiting for. The direct test of its framing has already been run and published, and the remaining questions are archival rather than scientific. Megaproject Governance supplies the modern half of the comparison and is cited rather than depended on: a finding already in print is not a pending edge. No typed depends-on edge is claimed.
• Requires (not on this map) Two things only an institution could assemble, and neither is a research result. First, megaproject outturn data recorded in a schema comparable across projects, jurisdictions and eras: date of decision, the estimate as it stood then, what kind of artefact that estimate was, outturn cost in comparable terms, schedule at decision and at completion, forecast and realised benefits, and a fatality count with its classification rules stated. Megaproject Governance records that such collection is mostly absent today; it was more absent in 1931, and the largest existing series — 258 transport projects completed 1927–1998 — excludes buildings, dams and canals, which is to say it excludes all three of the examples the popular argument runs on. Until that schema exists the era comparison is being made on a sample that was never built to support it. Second, a historical reference class that includes the projects that were abandoned, went bankrupt or were never finished. The French Panama attempt is the model case: the same objective at the same site, over 22,000 dead, a complete failure, and absent from every list of what the nineteenth century built well. A base rate assembled only from what survived to be admired is a selection on the dependent variable, and correcting it is a documentary programme — finding the projects that left no monument — that no body is funded to run.
• Enables This brief supplies the historical half of an argument that recurs wherever scale does — a tested null on the era claim, a worked account of the columns the canonical examples omit, and an explicit statement of what the record does and does not permit to be measured. No typed enabling edge is claimed, because an epistemic correction is an input to how other briefs are read rather than a prerequisite for anything being built.
• Adjacent Economic and labour history, which supplies the fatality series and the compensation-statute mechanism; project management and cost engineering, which supplies the outturn method; the historiography of technology, which supplies the survivorship problem; and within this map Megaproject Governance, Automated Construction Systems, Continental Transportation Systems, High Speed Transit Networks and Infrastructure Resilience.

Handwave “Megaprojects used to be delivered better than they are now.” This is the framing under test and it has been tested directly, in the field's foundational paper, on a sample that begins in 1927. Flyvbjerg, Holm and Buhl regressed cost-estimate error on year of decision across 258 projects completed between 1927 and 1998 and could not reject the null: “The null hypothesis that year of decision has no effect on the difference between actual and estimated costs cannot be rejected (p=0.22, F-test)… We therefore conclude that cost underestimation has not decreased over time.” Established The framing predicts a decline in delivery quality. The data show no trend in either direction. And the worst overrun in the canonical catalogue — the Suez Canal at 1,900% — is a nineteenth-century project.

Established “We have learned to build better than our grandparents.” The same F-test kills this one, and it should be stated with equal force because half of the audience for this brief arrives holding it. Nothing in this evidence licenses a progress narrative. Frontier The honest terminal position is a tie, and the tie is the finding. Seventy-one years of data, a null result, and both sides of a loud argument refuted by the same statistic.

Established “The canonical examples are myths.” They are not, and this brief refuses the easy version of the correction. The Empire State Building really did go up in about thirteen and a half months within a $50 million budget. Hoover Dam really was handed over more than two years early at a bid price of $48,890,955. The American phase at Panama really did finish two years ahead of its 1916 target. Frontier The framing does not fail because its examples are false. It fails on what has been left out of them.

Established “The Empire State Building is the proof that fast and cheap delivery was normal.” On the field's own three criteria it is a partial success. On budget, dramatically early, and roughly twenty years late on benefits — unprofitable until the early 1950s, its upper half unrented, nicknamed “The Empty State Building.” Frontier The strongest defence is real and is stated here at full strength: blaming the developers for not forecasting the Great Depression “is moving the goalpost,” and the building has been consistently profitable for most of its ninety-year history at about 5.3% average annual growth. Both halves belong on the page, and the reason the benefit column matters is that benefit shortfall is the modern literature's central finding.

Established “Hoover Dam killed 112 people.” That is the recorded figure and it is incomplete by an amount the record itself states. A further 42 workers were recorded as having died from pneumonia and were not included in the total, with workers alleging at the time that these were carbon-monoxide deaths from the diversion tunnels; an independent modern account puts 37 tunnel workers dead from carbon monoxide released by idling gasoline engines. Frontier The death toll of the framing's best example depends on a classification made by the contractor being paid, and disputed by the workforce at the time.

Established “The Panama Canal shows what a confident century could do.” The same canal, at the same place, was attempted twice. The French effort of 1881–89 cost $287 million, killed an estimated over 22,000 people, and failed completely. The American effort of 1904–14 killed about 5,600, of whom roughly 350 were Americans and the great majority West Indian labourers, and finished two years early. Frontier The remembered project is the second attempt, and the reference class has been silently pruned to n=1.

Frontier “They built faster then.” Partly true and radically incomplete. US construction fatality rates ran roughly 150–200 per 100,000 workers in the 1930s and 1940s against 13–15 today. Established Any honest version of this sentence has to carry the clause “at roughly ten to fifteen times the death rate per worker”, and once it does, the comparison stops being about competence. Handwave What this does not establish is that safety regulation explains historical speed. No source fetched for this brief decomposes historical schedule performance into safety, regulation and anything else. The mechanism is a strong hypothesis with a large measured input and an unmeasured contribution, and it is flagged accordingly.

Handwave “Modern projects are slow because of regulatory scope creep.” This is the most confidently repeated claim in the entire “we have forgotten how to build” literature and it is quantified in none of the sources fetched for this brief. Speculative Scope that was absent in 1931 — environmental review, seismic standards, accessibility, archaeology, public inquiry — is now inside the project and inside the cost, and that much is not in doubt. The magnitude is entirely unmeasured, and producing it requires constructing a counterfactual project rather than observing one.

Established Two numbers this brief deliberately does not print, so that their absence is not read as an oversight. Any Empire State Building construction fatality figure — the commonly cited official number is five, higher figures circulate, and no reliable source was obtained. And the Hoover Dam death toll as a settled number — 112 is a contractor-influenced count with 42 disputed exclusions, and printing it alone would repeat the accounting rather than report it. Frontier A third restriction applies to the whole brief: the 2002 result covers rail, roads and fixed links, so extending it to buildings, dams and canals is not licensed by the evidence, which means the three examples this argument runs on are outside the sample that tests it.