The Math That Humiliated Allied Spies

Between June 1940 and September 1942, Allied intelligence put German tank output at roughly 1,400 a month [S1][S3]. A small team working for the Economic Warfare Division of the American Embassy in London read serial numbers off captured Panzers and said 256 [S1][S3]. After the war, German factory records showed 255 [S3]. The spies were off by more than a factor of five; the statisticians were off by four-tenths of one percent [S3].

The economists Richard Ruggles and Henry Brodie ran the operation and eventually wrote it up in a 1947 Journal of the American Statistical Association paper that the CIA later declassified [S1][S2]. Their first target wasn't tanks. It was tires [S1]. The Economic Warfare Division started serial-number analysis in early 1943 on rubber, then worked through trucks, guns, flying bombs, and rockets before anyone ever pointed the technique at Panzers [S1].

The method rests on a single observation that sounds wrong until you sit with it. If a factory numbers its output 1, 2, 3, and so on, and you grab a handful at random, the largest serial in your sample is a near-perfect proxy for the total. Say you spot five tanks and the highest chassis number is 60. The minimum-variance unbiased estimator hands you N ≈ m + m/k − 1, where m is your maximum and k is your sample size [S3]. Plug in: 60 + 60/5 − 1 = 71 [S3]. Five tanks, one line of arithmetic, and the entire fleet is pinned to within a rounding error.

The intuition is that the maximum carries almost all the information about the population. If the true total is 300 and you've already seen #295, the gap left to discover is tiny; if you've only seen #60, the expected gap above your max scales roughly with m/k [S3]. Human spy networks are systematically blind to this. The agent watching the rail yard sees whatever rolls past, not the highest-numbered Panzer in the Reich. The maximum has to be hunted with paperwork, not binoculars.

The single-month data is sharper still. For August 1942, conventional American and British intelligence put German tank output at 1,550 a month; the serial-number method said 327; German records later showed 342 [S3]. The spies missed by a factor of 4.5. The basement team missed by four percent [S3].

What made the London estimates uncanny was triangulation. German subcontractors stamped their own sequential serials on chassis, gearboxes, engines, and road wheels — four independent streams, each from a different supplier [S6]. Each stream gave its own estimate. When all four converged, Ruggles and Brodie had a level of confirmation no human network could touch [S6].

The road-wheel analysis was cleverer than the popular telling suggests. Analysts didn't just read the largest wheel serial off captured tanks; they counted how many wheel molds had been in use, then asked British road-wheel manufacturers how many wheels a single mold could spit out in a month, and multiplied [S3]. It was a two-stage industrial-engineering estimate dressed up as a statistics problem. From two captured Panthers — 96 wheels in total — they pinned February 1944 Panther production at 270; German records later showed 276 [S3].

Statisticians have been arguing about what the formula actually estimates for seventy years. Leo Goodman's MVUE is the minimum-variance unbiased estimator for a discrete uniform distribution, which is not the same thing as the expected number of tanks [S4]. Cory Simon's 2023 Bayesian treatment calls the problem a "weird case" in frequentist estimation and notes that there is no neutral prior for the Bayesian alternative either [S4]. The elegant one-line formula sits on top of a methodological argument that popular accounts cheerfully ignore [S4].

The Wehrmacht could have defeated the entire technique by randomizing serials or skipping blocks. They didn't, because no one on their side imagined the Allies were reading the stamps on gearboxes. Modern manufacturers haven't absorbed the lesson either.

Beginning in August 2008, a London investor posting on The Mac Observer's Apple Finance Board under the handle Tommo_UK asked iPhone owners to submit their device serials, IMEIs, and purchase dates [S5]. By early October the registry held data on nearly 150 iPhone 3Gs [S5]. He fed the numbers into the same formula Ruggles and Brodie had used on Panthers and concluded Apple had sold 9,190,680 iPhones by the end of September 2008 [S5].

Hobbyist statisticians later ran the same arithmetic on Commodore 64s. Reading serials off surviving machines, they arrived at roughly 12.5 million units [S7] — a figure consistent with internal Commodore numbers through 1993 (which showed roughly 10.6 million C64s alone and 12.35 million once the Commodore 128 is folded in), and well below the 17 million figure Commodore officially claimed and the 22-to-30 million Jack Tramiel asserted for decades [S7].

The technique works because the structure of the problem hasn't changed since the first Panther rolled off the line. A manufacturer numbers its output sequentially, the maximum leaks every time a unit leaves the building, and a one-line formula extracts the total. Ruggles and Brodie worked from offices the Wehrmacht didn't know existed and beat the combined output of MI6, the OSS, and the SOE [S1]. They did it because the answer was sitting on the side of every gearbox the Wehrmacht had stamped [S6]. Every numbered serial on every numbered object you own is doing the same job for someone else, right now.