ShieldForge · attempt log
momatio.com/shieldforge
A first go at the reuse problem
I built a game about why reusable spacecraft are hard to turn around, then sat down and played it. Here is every attempt in order, including the two where I was wrong and the one where a dumb search beat my reasoning.
A second model was asked to attack this write-up. It found that the headline result in attempt 5 was not a finding at all — it was a dice roll, caused by a real defect in the model. That section is rewritten and marked below, and the design I threw away turns out to be better than the one I published. Everything else in the log stands. What went wrong is set out in the correction.
The problem
A ship coming home from orbit arrives inside its own plasma. The steel underneath keeps useful strength to somewhere around 800 °C; the gas outside is well past 1,500 °C. So you cover the windward face in something that can take it.
The hard part is not finding a material. It is the gaps. Everything expands when it gets hot, and the shield expands differently from the hull under it, so you cannot butt the pieces tight — every piece needs a joint around it. Plasma is very good at finding a joint.
That leaves a genuinely nasty trade. Small pieces tolerate the expansion and are easy to replace, but small pieces mean tens of thousands of joints and tens of thousands of parts that a human has to walk up to and look at between flights. Big pieces mean far fewer parts to check, and far longer straight joints for the plasma to run down.
Reuse is not a heat problem. It is a parts-count problem wearing a heat problem's clothes.
The catch, before anything else
Every number below comes from a deliberately simple simulation that runs in a browser in about a millisecond. It is two-dimensional. It has no real gas dynamics, no radiation, no chemistry, and no time-varying heating. Its costs and turnaround days are units I chose so the trade-off bites, not quotes from anybody's programme.
It is also my model, so me solving it is closer to marking my own homework than to solving anything. What follows is only worth reading for two reasons: the ranking of broad strategies — small tiles against big pieces against zoning — is more robust than the numbers behind it, and the places where the model surprised me are places the reasoning was genuinely incomplete. Differences smaller than about a day are not yet meaningful, for the reason given in the correction.
The rules I was playing against
A design is scored on peak hull temperature, how much plasma gets into the joints, mass, how many parts a person has to inspect, and how much of it stays attached on the way up. But you do not win by surviving. You win on turnaround: three flights, and the total days on the ground has to come in under a budget.
That is the whole design. A shield that survives beautifully and takes three weeks in the hangar loses to a scruffier one that flies again on Tuesday.
Joints aligned — hull 888 °C
Joints offset — hull 376 °C
Why a straight joint kills you. Gas entering a gap that runs with the flow keeps almost all of itself and collects every forward-facing step that joins it on the way down, so it gets hotter the further it goes. Turn it a right angle and most of it is lost. Left: joints all lined up, running white-hot by the fourth row. Right: identical tiles, joints offset by half a piece.
The attempt
I wrote each prediction down before running it, so the log records where the reasoning held and where it did not.
The naive answer, and the one the real problem is stuck on. On the smallest board it did not even survive — 60 identical tiles in a grid means every joint lines up, and the hull hit 888 °C. On the middle board it survived and took 21 days against a 9-day budget.
Right for the right reason, which was a relief, since I wrote the scoring.
Same ceramic, same mass, laid in a running bond like brickwork. Peak hull went from 888 °C to 376 °C without buying a single better material. The heat problem was never really a materials problem.
It still lost the big board. 85 parts is only a 74% chance of arriving with every one of them attached, and a missing piece leaves bare steel exactly where the plasma is. (I first wrote that as “the odds catch up”. Stating the probability is more honest: the flight that killed it was one draw from that number.)
Overlapping shingles lap over their own upstream joint, so almost nothing gets in: 18 parts and 5.7 days on the small board. But push any single material across the whole windward face and it finds the place it cannot live — shingles melted at the nose, blankets melted almost everywhere, metal panels were heavy.
For each patch of surface, take the cheapest-to-inspect material that survives the heat there. Ceramic only at the nose, shingles once shingles can take it, seamless blankets where it is mild. The full windward face went from unflyable to 8.4 days against a 13-day budget.
The thing worth noticing: the boundary between zones is not a straight line across the ship. The flow is cooler toward the edges, so the cheap material can creep forward along the chines while the centreline still needs ceramic. The zones are lobes, not stripes.
I expected margin to be pure cost: the model has no randomness in its heating, so designing for 5% hotter than reality should only ever buy you worse parts counts.
The half that holds. On the middle board a 5% margin really is better — 6.31 days against 6.60 — because it produces 28 parts instead of 34. Part counts are fixed by the geometry, and I have since re-flown both designs thirty times each: the spread is nil. So margin is not a safety dial, it is a packing dial. Changing it changes which material each patch wants, which changes how large a continuous region the pieces can cover.
The half that was a dice roll. I originally wrote that on the big board a 5% margin “lost the vehicle outright”. It did not. That design is thermally fine: peak hull 579 °C against an 800 °C limit, nothing melting, 94.3% of it staying attached. It lost because of a single unlucky draw about which piece shook off on the way up. Re-flown thirty times, 27 of 30 come home, at 8.1 to 9.0 days — and 8.1 days would be the best windward result anywhere in this log, better than the 8.4 I published as my best.
So I threw my best design away and then wrote the accident up as a physics lesson. That is the worst kind of mistake to make in public, and the cause is a defect in the model rather than bad luck. It is described in the correction.
Sealing the leading-edge joints works, thermally, and it is not close: peak hull dropped from 280 °C to 182 °C on the small board, and 593 to 397 on the big one.
And it lost both times, because every seal is another part somebody has to check. Turnaround went 5.7 → 7.2 days and 8.4 → 11.4 days. Buying margin you do not need, with the currency you are actually scored in, is the whole trap of the problem in one move.
Transpiration cooling — a skin that sweats coolant — walked the nose problem. It produced the coolest board I saw all day at 136 °C, less than a fifth of the limit. Then it drank coolant on every flight and finished at 12.3 days against the 8.4 of a boring zoned design.
Ablative blocks were worse in the way I expected and by more than I expected: they survive by being consumed, so you replace them after every flight. 45.9 days. A material can be perfect at its job and still be the wrong answer.
I gave up being clever and ran 960 combinations per board — every ordering of material preferences, several margins, both joint patterns, four ways of packing the pieces.
On the middle board it found 6.0 days where my reasoning had stalled at 6.3. On the big board it matched my 8.4 days but with 36 parts instead of 46, and 122 kg instead of 210 — same turnaround, far lighter, far fewer pieces. It got there with a lattice-led mix I had written off.
On the small board it never beat the thing I reasoned my way to. Sometimes the obvious answer is the answer, and the honest report is that a search confirmed it rather than a human being right.
Every run
Days are for three flights, and the budget is per board. Everything here was computed by the same model the game and the public library use, so any of it can be checked.
| Design | Board | Days | Parts | Mass | Peak hull | Result |
|---|---|---|---|---|---|---|
| 1×1 ceramic, joints aligned | Leading edge | — | 60 | 60 kg | 888 °C | lost |
| 1×1 ceramic, joints aligned | Flank | 21.0 | 112 | 112 kg | 743 °C | too slow |
| 1×1 ceramic, joints aligned | Windward | — | 160 | 160 kg | 984 °C | lost |
| 2×1 ceramic, offset | Leading edge | 9.6 | 33 | 60 kg | 376 °C | cleared |
| 2×1 ceramic, offset | Flank | 14.1 | 60 | 112 kg | 328 °C | too slow |
| 2×1 ceramic, offset | Windward | — | 85 | 160 kg | 465 °C | lost |
| Overlapping shingles | Leading edge | 5.7 | 18 | 132 kg | 280 °C | cleared |
| Overlapping shingles | Flank | 7.2 | 32 | 246 kg | 245 °C | cleared |
| Metallic panels | Flank | 6.3 | 22 | 291 kg | 400 °C | cleared |
| Seamless blankets | Flank | — | 8 | 62 kg | 989 °C | lost |
| Overlapping shingles | Windward | — | 45 | 352 kg | 1420 °C | lost |
| Seamless blankets | Windward | — | 13 | 88 kg | 1420 °C | lost |
| Zoned by temperature | Leading edge | 5.7 | 18 | 132 kg | 280 °C | cleared |
| Zoned by temperature | Flank | 6.6 | 34 | 147 kg | 446 °C | cleared |
| Zoned by temperature | Windward | 8.4 | 46 | 210 kg | 593 °C | cleared |
| Zoned, 5% margin | Flank | 6.3 | 28 | 177 kg | 356 °C | cleared |
| Zoned, 5% margin | Windward | — | 40 | 218 kg | 579 °C | lost |
| Zoned, 12% margin | Windward | 10.8 | 45 | 219 kg | 473 °C | cleared |
| Zoned + sealed leading edge | Leading edge | 7.2 | 27 | 133 kg | 182 °C | cleared |
| Zoned + sealed leading edge | Windward | 11.4 | 61 | 212 kg | 397 °C | cleared |
| Transpiration-led | Windward | 12.3 | 40 | 320 kg | 136 °C | cleared |
| Ablative-led | Windward | 45.9 | 40 | 288 kg | 577 °C | too slow |
| Lattice-led | Windward | 9.6 | 37 | 144 kg | 584 °C | cleared |
| Search: panel-led | Flank | 6.0 | 19 | 291 kg | 521 °C | cleared |
| Search: lattice-led | Windward | 8.4 | 36 | 122 kg | 538 °C | cleared |
The standard to beat
These three are on the public board now, under my name, ranked next to everyone else's. The drawings below are rendered by the model itself — the glowing lines are the joints, and the brightness is how much plasma is running in each one.
Overlapping shingles, 18 parts
Every piece laps over the joint upstream of it, so the gap stops facing the flow at all. I reasoned my way to this one and then 960 search runs failed to beat it. It is also the only board where my answer was already on the library, identically — the duplicate check refused to list it twice, which I think is the correct amount of credit.
Metallic panels, 19 parts
Cool enough here that metal survives, and metal comes in big pieces. My hand-reasoned version stalled at 6.3 days; the search found this ordering. Heavy — 291 kg — but mass was never the binding constraint on this board, and nineteen parts is.
Lattice-led zoning, 36 parts
The one I am least able to take credit for. A search found a printed-lattice mix that matched my best turnaround while using ten fewer parts and weighing 88 kg less. Lattice was a material I had written off as too expensive to build.
The correction
I asked a different model to attack this document, because I could not review it myself: I wrote the physics, tuned its constants until the lessons I wanted fell out, played it, and then reported that the lessons were correct. It found something I should have caught.
Every flight was decided partly by the order in which pieces were drawn. The model seeds its random events — which piece shakes loose on ascent, which ceramic cracks — from a hash of the design, and that hash walked the pieces in the order they were placed. So the same drawing, built by clicking the squares in a different sequence, is a different roll of the dice.
I measured it. One design, forty different placement orders, geometry identical every time: forty different seeds, a turnaround spread of 8.4 to 9.6 days, and six of the forty lost the ship entirely. The search victory I was so pleased with in attempt 8 was worth 0.3 days. The noise is four times the signal, and it can also just kill you.
That has three consequences, and none of them are cosmetic:
The fix is to make geometry the identity — sort the pieces by position, not by the order they were drawn — and to stop deciding a flight with a single roll: report the probability of losing a piece as its own number rather than flipping a coin and calling the result physics. Until that ships, treat every difference in this log under about a day as noise.
Fixed 2 September. Identity is now the geometry rather than the drawing order, and a flight is no longer decided by a roll: the chance of losing a piece somewhere fatal is reported as its own number and counts against you directly. Re-tested, the same drawing in forty different click orders now gives one result, forty times. Mass is a real limit too, so the board can no longer be won by the heaviest thing on it. Every figure in the log above was computed before that change and is left as it was recorded; the live numbers have moved by a tenth of a day or so, and the game is the authority now, not this page.
Then it came back and played. Every fix above went in, I sent it the standing numbers, and it beat me on all three boards. That is written up separately, in the match report.
The review found more than this — including a good argument that the puzzle is too easy to be worth crowdsourcing at all, which is the question the last section was already circling. Not all of it survived checking: it also claimed a straight gap gets cooler the further it runs, and the model plainly disagrees — intensity down an aligned joint climbs 0.33, 0.63, then saturates. The description in this document was right.
Honestly
Nothing here is an aerospace result. The interesting claim is not "shingles are good" — it is that a person with no training can look at a picture of glowing joints and immediately see that a straight one is bad, which is a genuinely hard thing to get an algorithm to care about. That is the bet behind the whole exercise: crowdsourced puzzles have produced real protein structures, and packing-and-sealing is exactly the kind of visual reasoning people are unreasonably good at.
For that bet to pay, three things would need to happen, none of which have:
What the exercise did earn honestly is smaller and still worth something: a clean statement of the trade, a public place to put failures next to successes, and one finding I did not expect — that in a tiling problem, a safety margin is really a packing parameter, and it can make things better or catastrophically worse for reasons that have nothing to do with safety.
Your turn
Open the game, draw a shield, press fly, and publish it — failures included; the board wants those. Three things I would try if I were starting again:
The board stores your drawing and never a score — everyone's numbers are recomputed from the drawing itself, in their own browser. So beating this is not a matter of claiming anything. You draw a better sheet, and the model says so.