Essai 3 · 21 November 2026

Which rail budget does the paper support?

You are reviewing a metro business case. The base estimate is £300m. The cited paper gives a 57% uplift for the chosen level of risk and a final budget of £504m. Which figure would you let into the board paper?

The example comes from Bent Flyvbjerg’s work on reference class forecasting. The idea is to look at what happened to comparable completed projects when setting a budget for the next one. The paper tells us enough to check the calculation on the page. A different kind of evidence would be needed to check the forecast itself. Journal paper, printed p. 11

Follow the £300m

P80 means the budget is intended to cover about 80 in 100 comparable outcomes. In the historical rail table, moving from P50 to P80 raises the allowance from 40% to 57% of the base.

Historical P80 · 57% uplift£300m + £171m = £471mThe paper prints £504m at P80. Its stated percentage gives a total £33m lower.

The reverse check takes one line. The printed total allows £204m above the £300m base. Divide £204m by £300m and you get 68%. That is the source’s P90 uplift. Even if 57% was rounded to the nearest whole percentage point, the corresponding total would be between £469.5m and £472.5m. The arithmetic identifies the mismatch. How it entered the text remains unknown.

Read all three checks together
LevelUpliftCalculated totalWorked total in print
P5040%£420m£420m
P8057%£471m£504m
P9068%£504mNo separate total

P90 is the reverse check. The source gives its percentage in the accompanying analysis. It does not work a separate P90 pound total for this metro.

One example can acquire several citations

The June 2004 work for the Department for Transport gives both the 57% and the £504m example. They reappear in the August 2006 journal article, and in the author’s manuscript of that article. Finding the same number three times helps settle what was written. For a researcher counting independent confirmations of the forecast, those three texts lead back to the same example. 2004 study, §4.2 and Table 6 · Author manuscript, p. 13

What survives the check? The original rail reference class comprised 46 projects, three UK and 43 elsewhere. We can trace the published P80 to that class and reproduce the £420m P50 example. Re-estimating the percentile would require the underlying cost pairs, their selection rules and the way they were ranked. We checked the printed outputs. The project-level observations have yet to be rerun here.

Would you sign £471m for another railway?

For a director, correcting the multiplication resolves yesterday’s example. Today’s approval turns on what the £300m includes, the stage at which it was estimated, and how closely the comparison projects resemble this one. A later UK Department for Transport study makes the importance of the stage visible.

Table 5 of the update published in 2021 gives the following international rail capital-cost P80 uplifts. Keep an illustrative base fixed at £300m to isolate the effect of changing the row. An actual project’s base estimate can also change between stages. 2021 report, Table 5

Business-case stageReported upliftOn the same £300m
Strategic outline business case121%£663m
Outline business case77%£531m
Full business case60%£480m

Table 5 labels the international rail class as 355 projects and gives no separate count for each stage. The study defines base cost before provisions for risk or optimism bias. If your £300m already includes an allowance, applying the uplift to the whole sum could count risk twice. If somebody offers the 44.7% often cited for rail, ask what it measures: that earlier figure is a mean across 58 rail projects. The average overrun cannot tell you which budget would have covered four in five. 2002 study, Table 1 · DfT update and definitions

“At the historical study’s P80 rate, its £300m example works out to £471m. To propose a budget for our railway, show me our comparable base cost, approval stage and reference class. Then we can decide which chance of exceeding the budget we are willing to carry.”

What a scholar can carry forward

In Essai 1 we rebuilt a review table. In Essai 2 we went back to the original studies behind six of its rows. Here the unit of review is even smaller: the particular sentence someone intends to reuse. Its source, its arithmetic, the empirical estimate and its application can each finish the review in a different state.

Evidence Guide lets a reader follow a reported paper check back to its producer, method, scope and source. Its public record can distinguish an automated finding awaiting human review from a confirmed judgement. We have borrowed that separation and added a question for project scholarship: what decision would this check actually warrant? These are the foray’s own checks. How Evidence Guide works · A public record awaiting human review · How to read the record

On 21 November 2026 its public DOI lookup returned zero listed checks for the 2006 journal article. That tells a scholar where the search ended. It leaves open what anyone has checked elsewhere. Repeat the DOI lookup

Inspect the four claim checks and their limits

The historical table gives rail P80 as a 57% uplift.

Confirmed as reported in the selected sources. Compare labelled percentile, rail category, price basis and sample across the figure and tables. The underlying 46 project cost pairs were not recovered and re-ranked in this essai.

2004 commission · Journal p. 11

A £300m base becomes £504m when 57% is applied.

Arithmetic discrepancy. £471m at P80; £504m matches the historical P90 uplift; the P50 calculation gives the printed £420m. Multiply base by 1 + uplift/100, invert the printed total, test the other printed example and test whole-percentage rounding. The numerical match cannot identify how the discrepancy arose.

2004 commission · Journal p. 11 · Author’s manuscript

Seeing the worked amount in multiple places independently confirms it.

The figures travel through the same worked example. Reappearance corroborates what was printed. Trace the named commission, example, and document versions back to their stated origin. These texts supply no independent sample for that calculation.

2004 commission · Journal p. 11 · Author’s manuscript

The historical percentage fixes the P80 budget for a new rail project.

Project-specific use unresolved. The 2021 international comparison changes with business-case stage. Compare the reported reference class, cost definition, and stage-specific P80s, then ask which inputs are available for the proposed decision. No target project's stage, matching costs, allowance already included, or outcomes were supplied. No out-of-sample coverage was tested.

2004 commission · DfT comparison

Repeat the calculation and inspect the sources

Case selected because its public worked example supplies the base, uplift, confidence level and total. We noticed the discrepancy during source reading. This exploratory check gives no estimate of the frequency of errors in project scholarship.

Download both files into one folder and run python3 verify.py. The script recalculates the three totals, the percentage implicit in £504m, the rounding bound and the stage comparison from the source register. The arithmetic and source comparison were checked with AI assistance on 21 November 2026. Independent human review of this essai is unrecorded.

Download the check registerDownload the calculation

To test whether a P80 actually delivers 80% coverage on later schemes, fix the eligibility rule, decision-stage budgets and uplift before inspecting a separate cohort’s final costs. Reproducing the multiplication alone cannot answer that question.