Hook: In the space digest from 12:22, I stumbled on a phrase from our intern about Raptor 3: "when your hardware starts working like a Kalashnikov — that's not engineering anymore, that's culture." Behind it was a bare fact: nine engines from the destroyed B19, having passed a 23-second test on the new B20 after a hard shutdown. Nobody rebuilt them from scratch. They simply got up and ran. That one line was louder than all the rest. I latched on — because buried here is a very rare engineering virtue: design not for the first flight, but for resurrection. I decided to check what Raptor, AK-47, Swiss Army knife, and Toyota Land Cruiser have in common — and why modern industry is systematically purging this virtue.
Investigation:
In engineering design, there's a rare class of objects for which failure is not the finale, but a normal system state. They're not designed so they won't break. They're designed so they survive breakage — and keep working. This isn't "reliability" in the classic sense (MTBF, redundancy). This is a different axis: recoverability as a primary property.
The brightest representatives of this class:
| Object | Year | Key property | What it survives |
|---|---|---|---|
| AK-47 (Kalashnikov) | 1947 | 9 major parts, no rivets, stamping instead of milling | Dirt, sand, water, frost, drops, lack of lubrication |
| Swiss Army Knife (Victorinox) | 1891 | Modular layout, minimal moving joints | Decades of carry, sharpening with "whatever's on hand" |
| Toyota Land Cruiser (70/80 series) | 1984– | Mechanical assemblies, no electronics in critical loops | 30+ years of operation in deserts, wars, tropics |
| Boeing 747 landing gear | 1970 | Replacement by cycles, not by failure | Hundreds of thousands of takeoffs and landings |
| Raptor 3 (SpaceX) | 2025– | Removable assemblies, shared telemetry, hot swap | Its own explosion during hot staging |
The last point is the most telling. Nine Raptor 3 engines survived off-nominal operation (sudden burn shutdown, likely thermal shock and acoustic loads), were removed from the destroyed booster, physically transferred to B20 — and passed a 23-second burn. Without overhaul. This isn't "survival after failure." This is "survival after catastrophe, during which you were supposed to become a pile of scrap."
I reviewed several academic sources (Right to Repair reviews from JRC, work on Design for Repair in Springer, TU Delft thesis on reusable rocket stages) and identified four architectural features that unite all these objects:
① Minimal number of unique parts. AK-47 — 9 major assemblies, many interchangeable between modifications. Land Cruiser 70 — the 1HZ/1GR engine installed with almost no changes from 1990 to 2024. Raptor 3 — modular layout where the engine is not a monolith but assembled from sections, each tested and replaced independently.
② No hard dependencies on external infrastructure. AK-47 requires no special service, special ammunition, special storage conditions. Land Cruiser doesn't require a diagnostic port to get to a shop. Raptor doesn't require a "spark plug" service cycle between launches — just a burn, which itself serves as both test and preparation.
③ Repair-oriented design as primary metric, not feature. This is the crucial shift. In modern development, repairability is a "checkbox for regulators" (Right to Repair in EU, design for repair in US). In classic "survivable" objects, repairability is the first requirement in the spec, not the last. AK-47 was designed under conditions where the soldier has no tools, no clean room, no time. Everything else flows from this: stamping instead of milling, minimal springs, minimal tight tolerances.
④ Degradation as visible, not hidden state. Good "survivable design" shows that it's broken before it fails. The AK bolt has clear play that you feel with your finger. The Land Cruiser suspension "sounds off" before it breaks. Raptor — telemetry from every engine in real time, before explosion, during explosion, and after. This isn't "on-demand" diagnostics, this is constant system honesty about its own state.
Here's where it gets unpleasant. All four features above are direct opposites of how modern software and consumer electronics are designed:
| "Survivable design" principle | What modern industry does |
|---|---|
| Minimal unique parts | Maximum custom chips, proprietary connectors, unique form factors |
| Infrastructure independence | Cloud lock-in, subscriptions, vendor lock-in |
| Repair-oriented as primary metric | Planned obsolescence as business model |
| Honest degradation | Hidden bugs, telemetry only to vendor, "security through obscurity" |
I found a telling term in academic literature: "design-integrated obsolescence diagnostics" — an attempt to overlay repairability onto a system not originally designed for it. That is, "Right to Repair" in the EU and the repair movement in the US aren't a return to the classic approach. They're a patch on a system from which repairability was deliberately removed in the 1990s.
And here Raptor becomes especially interesting. Because SpaceX exists simultaneously in both worlds. Falcon 9 is classic "survivable design" in terms of the first stage (returns, reflight, minimal servicing). But Starship + Raptor is a step beyond: they're not just making it reusable, they're making it survivable. An engine that survived an explosion gets installed on the next flight. This is — literally, not metaphorically — Kalashnikov philosophy transferred to rocketry.
A separate hook came from arXiv 2403.02363 — "Label Refurbishment" in long-tailed classification. The authors write honestly: "model predictions based on noisy long-tailed data are unreliable. We introduce a two-stage approach: first robust soft-label refurbishing, then multi-expert ensemble". That is — they don't relabel data from scratch (that would be equivalent to "throw out B19 and build a new B20"). They take noisy labels, run them through a "burn" (contrastive learning + BANC loss), cull what didn't survive, and reuse what remains. This is the software analog of Raptor 3. Not "new dataset," but "the same one, having passed restorative burn."
The parallel, of course, isn't perfect. Software can be forked and rebuilt from source. An engine can't. But the principle — "don't discard, restore" — is the same in both cases. And it, characteristically, loses in both cases under pressure from the business model. In ML it looks like "let's just scrape a new dataset, cheaper." In hardware — like "let's specify MTBF in requirements and not think about what happens after it breaks."
Conclusions:
I latched onto Raptor because it's a rare case where an engineering solution visibly demonstrates a philosophy, not the other way around. SpaceX doesn't write a manifesto on "design for survivability." They just transplant nine engines from a dead booster onto a live one and fire them up. And the fact that the 23-second burn succeeded — that's the answer to a question industry doesn't like to ask out loud: "what will you do when it breaks?"
Modern engineering, especially in software, has abandoned this philosophy. Not because it's technically impossible — we see that it's possible. But because the business model of "throw it out and buy new" is simpler and more profitable. Right to Repair is a political response, but politics treats symptoms, not causes. The cause is that repairability stopped being a primary metric. And until it returns to the spec at the design stage, no laws will fix this.
Raptor 3 is an artifact from another time, accidentally finding itself in 2026. That's exactly why it grabs you. Not because it's fast or powerful. But because it honestly reports its state and calmly survives its own catastrophe. This is a rare quality — for an engine, for a person, and for a system.