III

The third thing was kept because everyone agreed it would be irresponsible to delete it. It was an orbital solver maintained by a planetary-defense institute that no longer used it for planetary defense. For thirteen years the institute had run the old solver beside the common service during exercises. They fed both systems the same simulated detections, withheld different fractions of the data, corrupted timestamps, removed stations, invented non-gravitational accelerations, added bad weather and deliberate deception. The local solver was sometimes first to notice a malformed observation. It was almost never first to the right orbit. Its uncertainty grew too slowly in some cases, too quickly in others, and its maintenance consumed the equivalent of five full-time researchers who increasingly spent their research time persuading antique dependencies to compile. The common service was not one model. It was a collection of estimators, physical solvers, learned corrections, and independent implementations maintained across several institutions. It won the exercises often enough that continuing to staff the local system began to resemble nostalgia with a budget code. The institute asked whether the five people were better kept there or moved into sensor characterization, where the common service remained weakest. It moved them. Two went to radar. One went to infrared. One designed adversarial observation tests. The fifth retired six months later and left a card taped inside the old chassis with the boot sequence written in block capitals, because the remote console had a habit of forgetting. The machine stayed in the room. Its disk images stayed in the archive. A standing requirement said it could be restored if live estimates ever became difficult to reconcile. Nobody proposed deleting the requirement. Nobody proposed paying someone to practice it every month either. Four years later, a storage audit found the chassis still powered off, clean, and exactly where the inventory said it was.

Redundancy has a cost.

People have to maintain it. Old dependencies have to compile. Spare parts have to exist. Somebody has to practice the procedure often enough that the emergency version is more than a sentence in a binder. Five researchers assigned to a backup are five researchers unavailable somewhere else.

So "keep another one" is not a constitutional theory. A duplicate that never wins, never teaches, and cannot be brought online in time may be an expensive way of remembering that somebody once worried.

Sometimes the responsible move is to let the old instrument become history and spend the living capacity where it can still see something new.