Tommy Flowers

Tommy Flowers

Electronic engineer, 1905–1998. Study focus: statistical counting, synchronous input, and engineering a reliable specialized computer.

Flowers led the engineering of Colossus at the British Post Office research station at Dollis Hill. The machines helped Bletchley Park attack Lorenz-encrypted teleprinter traffic, called Tunny by the British. They combined fast tape input, electronic logic, counters, and configurable connections. They were specialized cryptanalytic machines, not stored-program general-purpose computers.

A collaboration with an institutional base

Flowers brought experience with telephone systems and large electronic installations. Colossus also depended on cryptanalytic methods, management, engineers, manufacturing, maintenance, and operators. Max Newman's section at Bletchley Park developed machine-assisted statistical attacks; William Tutte's work on the cipher was crucial to the underlying problem. Describing Flowers as the person who single-handedly broke Lorenz erases those dependencies. Museum history

A widely repeated story says he built the machine with his own savings after officials refused support. Archival research by Thomas Haigh and Mark Priestley shows an authorized collaboration between Bletchley Park and the Post Office, supported by management. Personal initiative mattered, but it operated inside institutions with staff, facilities, and wartime resources. The unsupported financial-sacrifice story is therefore not used here. Haigh and Priestley, author preprint

The first Colossus was tested at Dollis Hill in December 1943 and entered operational use at Bletchley Park in early 1944. Mark 2 followed in June 1944, and ten machines were operating by the end of the war. The distinctions between a workshop test, installation, and operational codebreaking explain why attaching one date to the word “completed” can mislead. Wartime secrecy limited public knowledge for decades, but it does not establish that the project had no influence through people who later worked on computing. Colossus history

From a cipher to a counting problem

Lorenz encryption combined teleprinter symbols with a generated key stream. XOR is a useful teaching model: for each bit, c=p XOR k, where p is plaintext, k the key bit, and c ciphertext. XORing by the same key again recovers the plaintext because k XOR k=0. This identity alone does not break encryption. The difficulty is recovering useful information about an unknown key.

Colossus evaluated statistical conditions associated with candidate settings. Its output supported further cryptanalytic work; it was not a box that automatically translated every intercepted message into readable text. The Boolean expression below is an original simplified example of scoring a candidate sequence, not the actual complete Tunny attack.

Take the eight-bit observations c=10110100. Candidate A is 10100110 and candidate B is 01011001. Count positions where the observation and candidate agree:

c       1 0 1 1 0 1 0 0
A       1 0 1 0 0 1 1 0
equal   1 1 1 0 1 1 0 1   score 6

B       0 1 0 1 1 0 0 1
equal   0 0 0 1 0 0 1 0   score 2

The first candidate scores higher under this particular rule. That is evidence only if an independently justified model predicts elevated agreement for a correct setting. A candidate can score well by chance, especially when many candidates are tested. A real attack needs the cipher's structure, a valid statistical bias, and procedures for confirming promising results.

Timing is part of the algorithm

Earlier approaches compared streams on separate tapes, making synchronization difficult. Colossus generated key-related patterns electronically while reading the message tape. Optical sensing and timing tied to the moving tape let the logic evaluate the corresponding positions together. The machine's fast tape reader is commonly described as processing 5,000 characters per second. That is an input rate, not a modern CPU clock-frequency specification. Machine and reconstruction descriptions

For an original capacity calculation, a 20,000-character tape takes four seconds for one pass at that rate, ignoring leaders, acceleration, setup, and output. Testing 100 settings with one setting per pass would require at least 400 seconds of tape time. If a design evaluates five independent conditions during a pass, the useful count rate may increase without making the tape move faster. This illustrates why parallel function units and counters matter; it is not a performance claim for a specific historical job.

Suppose one stream arrives a position late. The circuit then compares c[i] with candidate bit k[i-1] rather than k[i]. Every component could be individually accurate while the combined statistical result is wrong. Synchronization is therefore a correctness condition, not merely an optimization.

Reliable systems from imperfect components

Flowers's experience challenged pessimism about using large numbers of valves. Reliability depended on circuit margins, power and thermal behavior, construction quality, testing, and operating practice. Continuous operation could avoid stresses associated with repeated power cycling, but it did not make components immortal.

Consider a constructed system of 1,500 independently failing components, each with failure probability 10^-5 in an hour. If any single failure stops the system, the probability of a failure-free hour is (1-10^-5)^1500, approximately 0.985. Even excellent component reliability accumulates at scale. The independence and every-failure-stops assumptions may be unrealistic; the calculation shows why system reliability must be analyzed at the system level. Historical component counts alone cannot prove that a design was reliable or unreliable.

What reconstruction can establish

Tony Sale's reconstruction account identifies surviving photographs, diagram fragments, engineers' notes, and interviews. These are concrete evidence sources rather than a claim that a perfect blueprint survived. A working reconstruction tests whether an interpreted design can function. Its choices and substitutions must remain visible when it is used to infer details about wartime machines. Sale's account

Avoid three conflations: configurable is not automatically general-purpose; electronic is not automatically stored-program; and an early successful machine is not automatically the sole ancestor of all later computing. Flowers's achievement is substantial with those distinctions intact.

Exercises

1. For observation 11001010, score candidates 11011000 and 00110111 using the agreement rule. Why is selecting the higher score not sufficient to claim successful decryption?
2. A 30,000-character tape runs at 5,000 characters per second. There are 120 settings and four can be tested per pass. Compute the ideal tape time. Explain two reasons actual elapsed time will be longer.
3. You find a late interview claiming that Colossus was privately financed, but contemporaneous correspondence authorizes institutional work. How should a biography handle the disagreement? What can a reconstruction contribute to this question?

Compare your reasoning with the worked solutions

1. The first candidate differs at positions 4 and 7, giving six agreements. The second agrees only at position 7, giving one. Those scores rank the candidates under a deliberately chosen rule. They do not demonstrate that either is a valid key, that plaintext is linguistically plausible, or that the score is statistically unusual after searching many alternatives. The relation between the score and the real cipher must be established separately.
2. Each pass takes 30000/5000=6 seconds. With four settings per pass, 120 settings require 30 passes, giving 180 seconds. Loading and aligning tape, choosing configurations, starting and stopping, reading results, or repeating a faulty run add time. The ideal estimate assumes all four settings really can be evaluated simultaneously and that tape speed remains constant throughout each counted pass.
3. Attribute the recollection and evaluate it against dated records rather than converting memory into certainty. The records may distinguish personal expenditure on small items from funding an entire staffed project. A biography should state the supported institutional role and avoid the contradicted lone-financier narrative. Reconstruction can test technical feasibility, but it cannot establish who paid wartime salaries or authorized work; those are archival questions.