You cannot patch a telescope.

On 21 August 2026, in the Payload Hazardous Servicing Facility at Kennedy Space Center, teams closed a 43-foot SpaceX fairing around the Nancy Grace Roman Space Telescope. Launch is no earlier than 07:26 EDT on Sunday 30 August, on a Falcon Heavy from Launch Complex 39A, and the destination is the second Sun-Earth Lagrange point, roughly a million miles out (NASA, 2026).

The engineering headline is the survey power. Roman carries a Wide Field Instrument with a field of view at least a hundred times larger than Hubble's, enough to measure light from a billion galaxies across the mission, plus a Coronagraph Instrument for direct imaging of planets and debris discs (NASA, 2026). That is the part the science press will run with. The part worth borrowing is the fairing.

The moment the design stops being editable.

When that fairing closed, the argument ended. Every trade made over a decade of design became permanent, and no amount of money or cleverness reopens it. L2 is not a place anyone visits. Hubble was different, and that difference is the whole lesson. Hubble flew in low Earth orbit with a spherical aberration in its primary mirror, and in December 1993 the crew of STS-61 flew up, performed five spacewalks over roughly 35 hours outside the orbiter, and installed COSTAR, a box of ten coin-sized mirrors that refocused the light before it reached the instruments (Smithsonian NASM; ESA/Hubble). A famously flawed telescope was corrected because it was reachable.

Roman gets no such visit. So the correctness has to be spent up front, in test campaigns and margin, rather than held back as a repair option. That is a different way to spend a budget, and it produces a different culture.

Every system we ship has a fairing, and most teams never name the date.

This is not a space note. The reason it matters to a studio that builds installations and operational software is that our work has the same structure and we are worse at admitting it. A web app feels infinitely patchable, so nobody draws the line. Then the line turns up anyway, uninvited.

• 01The enclosure goes to the fabricator. From that morning the cutout positions, the cable entry and the service access are fixed, whatever the software wants later.
• 02The cable run is chased into a listed wall. There will not be a second chase.
• 03A tenant fit-out completes and the ceiling closes over the mounting points.
• 04The client signs off artwork that goes to print for a nationwide rollout.
• 05A firmware image ships to devices with no reliable route back, which is most devices in most public buildings.

Each of those is a fairing. The distinguishing feature is not that the decision is important, it is that the cost of reversing it jumps by an order of magnitude overnight and stays there. Anything on the wrong side of that date has to be right rather than fixable.

What we actually do about it.

The practice is simple and slightly boring, which is usually a good sign. At the start of a build we write down the dates on which parts of the system stop being patchable, and we work backwards from those instead of from the launch date. It changes the order of work. Physical interfaces get resolved first, even when the software behind them is still a sketch, because the mounting plate is the thing that cannot be redeployed on a Tuesday afternoon.

The second habit is to push decisions across that line deliberately rather than by accident. If a screen size, a sensor position or a power budget can stay soft for another three weeks at no cost, keep it soft. If it cannot, force the decision early and take the argument while it is still cheap. The failure mode we see most often in other people's projects is a team that never noticed the line existed and discovered it on install day.

The clever bit is what NASA did with the risky instrument.

Roman also shows the other half of the discipline, which is that not everything on a locked platform has to carry the same bar. The Coronagraph is flown explicitly as a technology demonstration. Its threshold requirement is to detect a point source between roughly 6 and 9 lambda over D from a bright star at a flux ratio of about one part in ten million, and it is allocated 90 days of observing in the first 18 months of operations (SPIE / JATIS, 2025). That is a real target with a fixed time box, and it is deliberately below what a future mission would need. If the Coronagraph underperforms, the survey science is untouched.

That is the pattern worth stealing. Separate the part that must be right from the part you are trying to learn from, hold them to different standards, and say which is which in writing before anyone builds. On an installation that usually means a rock-solid core experience plus one genuinely experimental layer that can be switched off from a config file without a site visit. The experiment is allowed to be disappointing. The core is not.

A fairing closed in Florida last week and a decade of decisions became history. Most of our deadlines are softer than that, which sounds like an advantage and mostly is not, because a soft deadline is one nobody plans around. Find the hard ones in your own build. They exist whether or not you have written them down.