A NASA craft reached Mars 57 km above the surface instead of 226.

•  Two teams built one trajectory in different units.

•  One side worked in pounds, the other in newtons.

•  Nine months of small errors added up unseen.

•  The orbiter arrived 170 km too low.

The Mars Climate Orbiter went up from Cape Canaveral on 11 December 1998 on a Delta II. It was to circle Mars as a weather satellite and act as the radio relay for the lander arriving ten weeks behind it. NASA now records the whole mission: unsuccessful, on arrival, 23 September 1999.

Two organisations flew it. Lockheed Martin Astronautics in Denver built the spacecraft and ran it day to day. The Jet Propulsion Laboratory in California did the navigation. Between them sat a written specification setting out exactly what each file passing from one to the other would contain, and in which units.

Nothing about the craft itself was wrong. The software on board computed in metric and computed correctly. The navigators in California read their files as metric, which is what the specification required. Every person involved was doing their own work properly, in their own units, to their own standard.

Every row below comes from one document: the board's first report, signed on 10 November 1999 by Arthur Stephenson, director of the Marshall Space Flight Center. The board had seven weeks and one instruction: find what mattered in time to help the lander that was still flying. It runs to forty-eight pages and it names one root cause.

Read the rows as altitudes. Each is where the navigators believed the craft would pass closest to Mars, and each new reading sits lower than the one before it. The number at the foot of the table is not an altitude. It is the factor by which one file was wrong, every time it was read, for nine months.

Four things sat between the mistake and the planet, and none of them held.

•  The specification was written and not followed. The interface document named the units for the file: newton-seconds. The ground program wrote pound-seconds into it instead. One team used English units and the other metric, and the copy computed on board, in metric, was right throughout.

•  The thrusters fired far more often than planned. The solar array was mounted off to one side, so sunlight kept pushing the craft out of true. The firings that corrected it ran 10 to 14 times more often than the navigators had allowed for, and each carried the error.

•  A disagreement stayed open for months. Three methods were used to work out where the craft was. The one using Doppler alone kept putting it closer to the planet than the others. That difference was noticed, reported informally, never settled.

•  The last chance was discussed and not taken. A fifth correction burn had been kept as a contingency. It was raised out loud shortly before the engine lit. Nobody had written the procedure or the criteria, so it was never attempted.

The orbiter went into the atmosphere of Mars on 23 September 1999, nine and a half months after leaving Florida. The engine lit at 09:00:46 and the signal went at 09:04:52, forty-nine seconds earlier than predicted. It never came back out from behind the planet.

Ten weeks later, on 3 December, the Mars Polar Lander reached the south pole and was never heard from either. The orbiter was to have been its radio relay. The same board's second report, signed in March 2000, looked past the units at how the work had been organised, and found schedule and spending pressure running ahead of the checking on more than one project.

The first report put it plainly. The laboratory had been sending craft to other planets for thirty years, and there was a widespread perception that orbiting Mars was routine.

•  Correct and correct can still be wrong. Both sides did their arithmetic properly. The thing that belonged to nobody was the join between them.

•  Small and constant beats large and rare. No single firing moved the craft far. There were thousands over nine months, and not one was ever checked against the ground.

•  A disagreement left open is itself an answer. Three methods gave two answers for most of a year. The one that kept coming out worse was the one that was right.

•  A contingency nobody rehearsed does not exist. The correction burn sat on paper as an option. When the moment came there was no procedure, no criteria and no time.

I watched two desks in 1996 carry the same position in opposite signs for a fortnight, because each sheet was read on its own and never beside the other.

Of anything that passes between two parties, ask who owns the join and in what units the thing is counted. Gold is counted in one unit everywhere, by everybody, and the number does not change when it crosses a desk.