A closed-form specimen of symmetry-breaking: Ballpark Figures enumerated all 30,093,975,536 legal Battleship placements and showed that uniform random placement on the asymmetric 10×10 substrate crystallizes a sharply non-uniform structure (center 21% vs corner 8%) — and that non-arbitrary optimal play emerges from it, all the way down into the one choice that looks purely free. Guessing is a best-response against that fixed structure (deep skill gradient: 63 → 45 turns, beats a human 86%). Placement is a minimax against an adversary, and he proves in extensive form the optimum is ~uniform random — i.e. even the apparently-arbitrary layer is uniquely forced (be maximum-entropy or get exploited). The portable claim: structure creates norms — given a structure + a telos, optimal solutions emerge from what looked like complete randomness, and there is no escape into arbitrariness even where the structure seems to permit it.
Links: Symmetry Breaking (the mechanism this is a closed-form specimen of), The Four Trunks, The Constitutive/Elective Distinction, The Dominance-Frontier Lens, Randomness as the Termination Mechanism (N≥3), Capability Without Leverage, NA1 — A Game-Design Crucible (the epistemic-axis twin + the binding-constraint-determines-depth thesis this sharpens), The Hollow Opponent (NA1’s AI = the flat-opponent end of this gradient), Catan — 47k Empirical (placement-equilibrium flatness rhymes with opening-barely-above-random), Yahtzee — 259 Trillion → 405 Million (the sequel specimen, same creator — he calls back to this video explicitly; where Battleship’s randomness is adversarial (placement = minimax vs. an opponent), Yahtzee’s is exogenous, which is why its whole interaction surface collapses to one risk-posture knob — the “solo-together” thread), Hangman — Solving Both Sides (the third specimen, and the one that inverts half of this page: same best-response/minimax split, but on a lumpy substrate the minimax optimum is a hard concentration rather than max-entropy — because the indifference set of an equilibrium is bounded by the substrate’s structure. Where this page shows structure emerging from apparent nothing, hangman shows structure that cannot be optimized away)
A specimen page (dated, source-anchored). Source: Ballpark Figures — “I Analyzed All 30,093,975,536 Battleship Boards So You Don’t Have To” (18:51, 2026-06-13). Transcript in
raw/battleship-ZBdajiTz48k.en.vtt. Its portable home is symmetry-breaking (structure→norms); the best-response-vs-minimax gradient below is the game-theoretic mechanism, and the constitutive/elective reading is the philosophical payoff.
Boards are drawn uniformly at random — every legal placement equally likely. Yet the per-square occupancy heat map is wildly non-uniform: a center square is ~21% likely to hold a ship, a corner only ~8%.
The cause is pure counting. A length-k ship has more ways to pass through an interior square than a corner one (the carrier: 10 placements through the center vs. 2 in a corner). So a uniform distribution over the joint (whole boards) induces a sharply non-uniform distribution over the marginal (single squares). That gap — uniform joint, peaked marginal — is the entire strategy. It’s worth holding onto as a standalone trap: “everything’s equally likely” almost never implies “every cell is equally likely.”
You are deciding against a distribution you don’t control and that doesn’t fight back. So you exploit it:
This is a decision problem, not a game: the prior is fixed, so there’s enormous room to be smart. The heat map is the dominance ordering over guesses.
Now flip sides. You’re choosing a placement distribution against an opponent who will best-respond to whatever you do. The author hunts for the “holy grail” — a placement so good that even an opponent who knows it can’t sink you faster. That’s a minimax / Nash strategy.
The punchline: optimal placement is barely better than uniform random. The best he found nudges corner-placement to ~2× the random rate, slows the opponent ~½ turn/game, and wins 51.7%. His own closing line: “doing tons of work to build a fancy algorithm only to find out that randomness works almost as well.”
Why flat? The minimax equilibrium against an adaptive opponent is near-maximal-entropy — any structure you add is structure the opponent can exploit, so the un-exploitable strategy is the one that looks like noise. Max-entropy being near-optimal is the flatness.
A sub-problem’s skill gradient tracks the exploitability of the opponent model.
- Best-response against a fixed distribution → exploitable → deep skill gradient (guessing).
- Minimax against an adaptive adversary → equilibrium is high-entropy → flat skill gradient (placement).
This sharpens binding-constraint-determines-depth: the binding constraint on Battleship skill isn’t a resource — it’s which half you’re in. All the depth lives in the best-response half; the minimax half is a near-solved coin flip. Same game, two sub-problems, opposite verdict. It also explains the Catan 47k finding that opening placement predicts win rate barely above random (27% vs 25%) — opening placement is the minimax half; in-game play is the best-response half.
This is why the video matters beyond games: it is a closed-form proof of the symmetry-breaking mechanism, the vault’s account of how constitutive features arise. Map it onto that page’s central move (equal forces on an asymmetric substrate produce asymmetric outcomes):
| Symmetry-breaking | Battleship |
|---|---|
| Symmetric forcing | uniform random placement (no bias injected) |
| Asymmetric substrate | the bounded 10×10 grid (edges truncate rigid ships) |
| Crystallized residue | the heat map (center 21% / corner 8%) |
| Load-bearing optimum | center-out greedy play, forced by the residue |
Every cosmological example on that page is gestural — we can’t re-run the universe. Battleship is the version where the substrate is fully enumerable (30B boards) and the frozen structure is computed exactly. It is the page’s best answer to its own Open Question #1 (“is there a quantitative measure of load-bearing?”).
The minimax half is the sharpest “conditional ≠ arbitrary” case in the vault. Placement is where relativism plants its flag — “place your ships wherever, it’s a free choice.” The video disproves this by theorem: the structure reaches into the one spot that looks purely arbitrary and dictates a unique optimal form of the apparent randomness — maximal entropy. Randomness here is not the absence of structure; it’s the optimal response to it. That snaps the relativist’s slide (§140 of symmetry-breaking: contingent → conditional → optional → arbitrary) at “arbitrary” — the freest-looking layer turns out to be the most tightly determined.
The constitutive/elective distinction falls straight out (constitutive-elective):
Both are real; neither is arbitrary in the sense the relativist needs. It is bounded freedom inside a forced distribution — the whole picture in a 10×10 box.
NA1 is the same proof on the epistemic axis. Battleship is structure→norms (optimal play emerges); the Nobunaga RE work is structure→meaning (what something is emerges): raw bytes → impose the 6502’s structure → reset/NMI/BRK vectors → opcode semantics → the VM → how the game works → how to play it best. The bits looked arbitrary; the structure was latent and discovered, not invented. Discovery and norm-emergence are one propagation story on two axes — the relationship symmetry-breaking’s Open Question #4 anticipates.
The honest boundary. Battleship proves the mechanism in closed form, but it hands you the telos (“win”) for free — and real ethics is the fight over whether the goal is given. So this is a clean proof of the conditional (structure + telos → emergent non-arbitrary norms), not of the telos itself. The vault pays for the telos elsewhere: symmetry-breaking’s stability-relative-to-timescale (§129) plus the convergence/selection trunk — a structure that ignores its emergent optimum gets selected out (sunk in Battleship, conquered in NA1, the Wilson force-doctrine’s “fit → durable might”). Battleship proves the clean half; selection supplies the goal Battleship gets for free. Stated that way the claim survives the obvious objection (“but you stipulated the goal”).
feedback_randomness_axis_lens): Battleship’s randomness/information lives on the config axis (the board setup), like Catan — but hidden and static. You can’t read the config (Catan you can), so you accumulate information about it over time. That makes it hidden-trackable (feedback_capability_value_and_information_design) — the legitimate-soft-flaw category, not hidden-random. The whole guessing half is config-axis information accumulation.First-move advantage under mutual-greedy play = 51.6%. Saving a single turn per game ≈ +3 percentage points ≈ one extra win per 30 games. The post-hit direction heuristic alone saves ~1 guess/ship. Small mechanical edges compound — the same “slots are the binding resource” intuition as the NA1 turn-economy work.
The author teases a “slightly better way to win” in a future video and admits in his write-up to many failed attempts at a placement strategy beating random. The interesting unsolved question: how much of the ½-turn placement edge survives against a greedy opponent that adapts its prior mid-game (he retunes the heat map for known-pattern opponents)? If placement edge ≈ guessing-prior-misspecification, the two halves aren’t independent — and the “flat” verdict is only flat against a non-adapting greedy searcher.