Friday Fun: The Unit Was Agreed in Principle

In September 1999, a spacecraft approached Mars after travelling hundreds of millions of kilometres. It carried instruments intended to study the planet’s atmosphere and provide a communications relay for another mission. Its navigation depended on a succession of small corrections, calculated from measurements and the known effects of its own manoeuvres. Nothing about the journey suggested that the most dangerous distance lay between Colorado and California.

The Mars Climate Orbiter did not need anyone to mistake Mars for Venus. Its trouble required a smaller disagreement. A ground software file supplied data about the impulse produced by the spacecraft’s thrusters in pound-force seconds, while the navigation software expected newton seconds. The numbers crossed an organisational boundary with their meaning detached. Each application could process the values it received. The equations did not object, because equations rarely demand to see the minutes of the meeting in which someone defined their inputs.

Navigation works by maintaining an estimate of where a vehicle is and where it will go, then adjusting that estimate as observations arrive. Thrusters complicate the calculation: even modest forces accumulate when a craft has months in which to drift. The physical spacecraft responds to the force actually applied. The mathematical spacecraft responds to the force described in the data. When those two forces differ, both trajectories can look internally coherent while gradually separating. At the next observation, the difference should invite investigation. If the organisation treats the model as the authoritative witness and the observation as an awkward guest, the invitation may never reach anyone empowered to accept it.

The orbiter was lost during its attempt to enter orbit around Mars. NASA’s investigation identified the unit mismatch as the root cause, alongside failures to detect and correct it through systems engineering and operational checks. This matters because the familiar retelling, in which one engineer forgot to convert a number and ruined a spacecraft, offers the comfort of a replaceable culprit. The actual problem survived many opportunities to notice a growing disagreement between prediction and reality. Blaming one conversion would also have allowed everyone else to keep the procedures, the reporting lines and perhaps even the commemorative slide about cross-functional excellence.

An interface with excellent manners

A number alone does not describe a physical quantity. Twenty may mean metres, seconds, kilograms, degrees or the number of people invited to approve a decision. Software can enforce a field type while remaining indifferent to the unit that gives its contents meaning. A value arrives in the expected position, the parser accepts it, the calculation runs, and the output acquires several decimal places. Precision then lends an undeserved air of authority to a misunderstanding that began before the calculation.

Engineers deal with this by defining interfaces, validating units, checking independent measurements and investigating residuals: the differences between what a model predicts and what instruments observe. No single practice guarantees safety. An interface specification can say the right thing while an implementation does something else. A test can reproduce the same mistaken assumption on both sides of a boundary. A review can verify that the arithmetic works while overlooking what the numbers represent. The system needs a way for reality to contest the model, and people who can interrupt the work when that contest produces an uncomfortable result.

There is a particular comedy in calling the data transfer successful. The file arrived. Its structure passed. The receiving system consumed it. Had the exchange belonged to a quarterly programme, someone could have marked the integration complete, recorded the remaining discrepancy as a monitoring enhancement and thanked both teams for their collaboration.

The integration would have closed on Friday, the residual risk would have entered a register on Monday, and the register would have acquired an owner just in time for that owner to leave. Mars, regrettably, declined to attend the steering committee.

The joke depends on a serious distinction. Transporting information does not establish shared meaning. The more elaborate the delivery mechanism, the easier it becomes to overlook that distinction. A number can pass through an API gateway, a validation service, an event bus, a data warehouse and a dashboard without any of those systems asking whether its unit changed at the first handover.

Each stage can add a little more polish to the error. By the time an executive sees the chart, the mistake has a colour scheme, a trend line and an owner. The owner, naturally, owns the chart.

The reassuring geometry of green

Organisations generate similar quantities without writing units beside them. A team reports velocity. Another reports throughput. A portfolio office reports capacity. Finance reports savings. All four reports may count different things over different periods, with different exclusions. The labels still travel cleanly into a common presentation. Nobody has confused pounds with newtons, but the resulting calculation may require an equally imaginative interpretation of physics.

Consider a programme that asks several teams to estimate delivery in story points, then adds their numbers into a single forecast. Each team may use its scale consistently. That consistency does not make the scales commensurable. One team’s eight may describe a familiar change that takes a few days. Another team’s eight may describe a risky integration that occupies a fortnight. Adding the figures produces a sum, but the sum does not acquire meaning through the act of addition. Once the total reaches a dashboard, however, disagreement sounds like resistance to measurement. The people closest to the work must explain why a number calculated correctly describes nothing in particular.

Their objection can then be captured as a dependency, which adds another number to the dashboard.

Now attach a target to that total. Teams learn which numbers preserve the appearance of progress. Estimates become negotiations about acceptable reporting, work splits into shapes that fit the measure, and uncertainty migrates into categories the dashboard excludes. The organisation then observes an improvement in the metric it has trained people to produce. This outcome need not involve dishonesty.

People adapt rationally to a reporting system that rewards the legibility of work more reliably than its consequences. Eventually the forecast becomes an input to performance assessment, which makes correcting it personally expensive. The spacecraft version has thrusters. The corporate version has a spreadsheet that can change the trajectory without leaving the room.

The same problem appears when an incident ends in one system and continues in another. Operations restores a service and closes the incident. Support still handles customers whose transactions failed. Finance later reconciles the losses. Each function reports accurately within its boundary. The organisation reports recovery because the metric attached to the original alarm has returned to green. Customers, who neglected to organise their experience around departmental boundaries, may hold a different view.

Their inconvenience has no compatible field in the incident record. It may, however, appear three months later as a mysterious decline in retention, prompting a separate initiative to improve customer listening.

Small forces, long journeys

Why do these discrepancies persist? The size of the initial error often makes escalation look disproportionate. A slightly optimistic completion estimate, a small exclusion from an availability calculation, a customer outcome represented by an internal milestone: each can appear manageable alone. Over a long journey, their effects compound. Plans take subsequent plans as inputs. Staffing decisions follow forecasts. Commitments harden around dates. Teams optimise local work for a destination that the underlying measures no longer describe.

This is where the Mars story becomes less amusing. The unit error mattered because it affected a series of navigational estimates. Organisational errors also accumulate through repeated decisions, although their paths prove harder to reconstruct. A company can spend months optimising handovers between teams while the customer waits through all of them. Each handover meets its service agreement. The end-to-end delay expands. Everyone can point to a compliant local process while the product approaches the atmosphere at an unfortunate angle.

The response usually starts with more reporting. More fields promise more clarity, and a consolidated dashboard promises one version of the truth. Yet a new reporting layer cannot repair incompatible definitions that remain unresolved beneath it. It may hide them more efficiently. The useful question concerns the journey a quantity takes: who created it, what it meant there, what changed when it crossed a boundary, and which observation could prove that the interpretation no longer fits reality. That investigation takes longer than adding a red threshold. It also has the discouraging habit of revealing that several green thresholds worked exactly as designed. A reporting programme can therefore announce success while the original question moves to the backlog for a later phase.

The more distant an organisation grows from its work, the more valuable representation becomes and the more dangerous it feels to question one. Senior groups cannot personally inspect every deployment, customer interaction or invoice. They need abstractions. But an abstraction earns trust through contact with the thing it represents, especially when the two disagree. If the dashboard wins every dispute by definition, the organisation has stopped measuring the journey and begun navigating by the drawing of one.

Somewhere in a modern enterprise, a programme has a signed interface agreement, a fully automated data pipeline and a weekly assurance meeting. Its teams have aligned on the shape of the number, the owner of the number and the date on which the number must turn green. The unit remains a matter for a later workshop. The journey continues on schedule. When the planet fills the window, the programme will commission an independent review to establish why the dashboard failed to mention it.

Historical basis: NASA’s Mars Climate Orbiter investigation and JPL’s contemporary account.