Daniel sent us this one — he's asking about how wargames actually work, and specifically about that Swiss submarine incident where a diesel-electric boat slipped through a U.carrier group's anti-submarine screen and simulated sinking an aircraft carrier. The core question is: if the sub didn't fire a real torpedo, how did anyone decide it was a kill? What's the machinery that turns a make-believe torpedo launch into something the U.Navy takes seriously enough to review its entire anti-submarine posture?
That question gets at something most people miss. Wargaming isn't one thing — it's at least three different disciplines wearing the same trenchcoat. You've got field exercises with live troops and real equipment, command post exercises where staff officers sit in bunkers making decisions with no troops moving, and constructive wargaming where computer models adjudicate interactions between simulated units. The Swiss sub incident was a hybrid — the submarine was physically in the water off the coast of Florida, but the torpedo launch and everything after was simulated. That's the part people find confusing. How do you adjudicate a weapon that was never fired?
The bureaucratic term "adjudicate" is doing a lot of work here. Sounds like a zoning board ruling on a property line dispute.
It kind of is. Adjudication is the process of determining what happened when a simulated weapon interacts with a simulated target. And it's been around longer than computers. Naval War College in the 1880s had umpires using colored flags to indicate ship damage during tabletop wargames. They'd push model ships around on a floor, and a neutral umpire would say "your battleship took two hits, you're down to half speed." Same concept today, just with sonar equations and Monte Carlo simulations instead of colored flags.
Walk me through it. The Swiss sub — a Type 209 diesel-electric, if I remember right — is somewhere off the Florida coast. The USS Harry S. Truman carrier strike group is doing its thing. The sub's job is to get inside the screen and simulate a torpedo attack. What's the first thing that has to happen in the adjudication pipeline?
Before anyone can adjudicate a torpedo hit, you have to determine whether the target even knows the submarine is there. This is where sensor models come in, and they're surprisingly sophisticated. Navy uses acoustic propagation models that account for water temperature, salinity, depth, and seafloor composition. Sound bends in water — it refracts toward colder layers — and a submarine hiding below a thermocline can be effectively invisible to hull-mounted sonar. The Swiss sub used a tactic called bottom creep, hugging the seafloor to mask its acoustic signature in the ambient noise.
Sounds like me getting out of bed.
It's actually fiendishly clever. When you're sitting on the seafloor, your acoustic signature blends into the background reverberation. Active sonar has trouble distinguishing a stationary submarine from a rock formation, especially if the sub has anechoic coating. The computer models attempt to calculate detection probability based on the sonar equation — signal minus noise plus processing gain — but here's the thing: the models are calibrated against peacetime test data, and peacetime conditions don't include a submarine crew that's been practicing this exact scenario for months.
The computer says one thing, but reality might say another. That's where the human umpires come in.
The White Cell. This is the neutral umpire team that oversees the wargame and has authority to override the computer model when edge cases arise. In the Swiss sub incident, the White Cell adjudicated that the U.anti-submarine warfare screen — destroyers, helicopters, maritime patrol aircraft — would not have detected the submarine until it was inside five nautical miles. At that range, a torpedo launch is essentially un-defendable. A heavyweight torpedo traveling at fifty knots covers five nautical miles in about six minutes. A carrier can't maneuver out of the way in six minutes.
The detection piece is adjudicated. The White Cell says the sub is inside the screen, undetected. The sub announces it's firing, and then someone has to decide whether the torpedo actually hits.
And this is where we get into the kill chain adjudication. The full pipeline is detection, classification, targeting, firing solution, weapon flyout, hit or miss determination, and damage assessment. Each step has its own adjudication method. Classification is particularly tricky — the computer has to decide whether the sonar operator correctly identified the contact as a submarine rather than a whale or a fishing vessel. In this case, the Swiss sub was a known participant, so classification wasn't in dispute. But in a real free-play wargame with multiple submarines and civilian traffic, classification errors are a major source of surprise outcomes.
Free-play means neither side knows the other's plan, and there's no predetermined outcome. The opposite is a scripted wargame, where specific events are injected to force participants into decision dilemmas. "At zero seven hundred, you will receive a report that a civilian ferry has been hit by a missile — what do you do?" That kind of thing. Free-play wargames are designed to surface unexpected outcomes. The Swiss sub incident was free-play, which is why it caused such a stir. Nobody scripted the submarine getting through. It just happened.
Back to the torpedo. The sub fires — virtually. What determines hit or miss?
Navy's standard tool is the Torpedo Effectiveness Model, or TEM. It's a six-degree-of-freedom physics simulation — that means it models the torpedo's motion in three dimensions of translation and three dimensions of rotation. Pitch, roll, yaw, surge, sway, heave. It accounts for the torpedo's speed profile, its wake-homing or wire-guidance logic, and the target's evasion maneuvers. The TEM runs at about a hundred times real-time on modern hardware, so a six-minute torpedo run can be simulated in under four seconds.
Six degrees of freedom. That's the kind of phrase that sounds like it was invented by someone who really wanted a bigger budget.
It's actually standard aerospace terminology. Missiles, aircraft, submarines — anything that moves in three-dimensional space gets the full six-degree treatment. But here's where it gets interesting: the TEM doesn't just calculate a trajectory and call it a day. It runs Monte Carlo simulations — thousands of randomized runs with slightly different initial conditions and environmental parameters — to produce a probability distribution. The output isn't "the torpedo hits" or "the torpedo misses." It's "under these conditions, the probability of hit is seventy percent against a surface ship and about forty percent against an evading submarine.
Seventy and forty. Those numbers sound specific.
They're from the unclassified approximations of the U.Navy's Synthetic Theater system used in RIMPAC exercises. The actual classified numbers are almost certainly different, but the ratio tells you something important: submarines are much harder to hit than surface ships, because they can change depth. A surface ship can only maneuver in two dimensions. A submarine can go deep, release decoys, and hide in thermal layers. The torpedo's homing logic has to account for a three-dimensional evasion problem, and it doesn't always win.
The computer spits out a probability. Someone rolls a digital die?
The adjudication system draws a random number from a uniform distribution between zero and one. If the number is below the hit probability, it's a hit. If it's above, it's a miss. This is called a Monte Carlo draw, and it's the same technique used in financial risk modeling and nuclear reactor safety analysis. The key insight is that you're not simulating one outcome — you're sampling from a distribution of possible outcomes. In the Swiss sub case, the torpedo was fired from inside five nautical miles with no countermeasures deployed, because the carrier group didn't know it was under attack. The hit probability in that scenario would be extremely high — probably above ninety percent.
Then damage assessment. The torpedo hits.
This is where the adjudication gets genuinely grim. A heavyweight torpedo like the Mark 48 carries a warhead of about six hundred fifty pounds of high explosive. It's designed to detonate under the keel of a ship, creating a gas bubble that lifts the ship and then drops it into the void. The bending moment snaps the keel. For a carrier, a single under-keel hit is almost certainly a mission kill — the ship might not sink immediately, but it's not launching or recovering aircraft. The adjudication models for damage assessment use structural finite-element analysis derived from ship design data and historical sinking records. For the Swiss sub incident, the White Cell ruled it a catastrophic kill. The carrier was out of the fight.
That's what prompted the review. A diesel-electric submarine that costs maybe three hundred million dollars, versus a nuclear carrier that costs thirteen billion, and the sub won. Without firing a shot.
Without firing a simulated shot, technically. And this is where we need to talk about the fog of war problem, because it's the thing that makes wargaming actually useful rather than just a physics exercise.
In a real naval engagement, nobody has perfect information. The carrier group's tactical picture is built from dozens of sensors — sonobuoys, towed arrays, helicopter dipping sonar, satellite imagery — and all of that data has to be fused into a coherent picture. In a wargame, you have to deliberately degrade the information available to the blue force to simulate real-world confusion. This is called degraded blue force tracking. Ships that would show up on a perfect GPS feed are suddenly reported at slightly wrong positions, or with the wrong classification, or they disappear entirely for ten minutes because a satellite pass was blocked by weather.
You're injecting errors on purpose.
Yes, and the errors are calibrated against operational data. The Navy knows, from decades of exercises, that a P-8 Poseidon maritime patrol aircraft will misclassify a surface contact about fifteen percent of the time under realistic conditions. The wargame model builds that error rate into the simulation. Sometimes the blue force commander sees a submarine where there's actually a fishing vessel. Sometimes they see a fishing vessel where there's actually a submarine. The Swiss sub incident exploited exactly this — the carrier group's ASW screen was looking for a nuclear submarine, which is loud and fast, and missed the diesel-electric crawling along the bottom.
The musical equivalent of beige wallpaper. Nobody notices it until it's already in the room.
That's actually a perfect way to put it. Diesel-electric submarines on battery power are extraordinarily quiet — much quieter than nuclear submarines, which have to run coolant pumps continuously. A Type 209 running on batteries at four knots is essentially silent. The only way to detect it with passive sonar is to get very close, and by the time you're close enough, you're already inside its torpedo range.
We've got computer models doing the physics, human umpires handling edge cases, and deliberate information degradation simulating the fog of war. But here's what I'm wondering — how do you prevent the whole thing from becoming rigged? If the White Cell can override the computer, and the scenario designers can inject scripted events, what stops a wargame from just producing whatever outcome the Navy wants?
This is the fundamental tension in wargaming, and it blew up spectacularly in two thousand two with Millennium Challenge 2002. This was the largest wargame in U.history — two hundred fifty million dollars, thirteen thousand five hundred participants. The Red team was led by retired Marine Corps Lieutenant General Paul Van Riper, and he used asymmetric tactics that the simulation simply couldn't model. He sent motorcycle couriers instead of using radios to avoid signals intelligence. He used small civilian boats to move troops. He pre-positioned anti-ship missiles on commercial vessels. Within the first few days, his Red force had sunk sixteen U.ships, including an aircraft carrier.
The Navy's response was...
They restarted the wargame. Reset the scenario, changed the rules, and scripted the Red team's actions to ensure a Blue victory. Van Riper resigned in protest and later called it "the worst wargame experience of my career." The Navy's official position was that the restart was necessary to test specific learning objectives that the free-play chaos was preventing. But the incident became a case study in how institutional pressure can corrupt wargaming. When the wargame produces an outcome the organization doesn't want to hear, the temptation to dismiss it as "unrealistic" is overwhelming.
The Swiss sub incident could have been dismissed the same way. "The simulation didn't account for X, the White Cell got Y wrong, in a real conflict we'd have Z." But it wasn't dismissed.
Because it was too clean. There was no asymmetric trickery, no exploiting simulation loopholes. A submarine used a well-known tactic — bottom creep — against a carrier group that was conducting a realistic ASW search pattern. The White Cell's adjudication was conservative. If anything, the simulation probably understated the submarine's advantage, because peacetime sonar models don't fully capture how quiet modern diesel-electric submarines have become. The Navy looked at the result and said, "this could actually happen." And that's the gold standard for a useful wargame outcome.
Let's talk about scale for a moment. You mentioned the Millennium Challenge had thirteen thousand participants. That's a city. How do you adjudicate something that big?
You can't do it with human umpires. That's where agent-based modeling comes in. Navy's Global wargame series — Global 2000, Global 2025, and so on — simulates thousands of autonomous drones, ships, and submarines using software agents that make decisions based on programmed doctrine. Each agent has a set of behavioral rules: if you detect an enemy submarine, deploy a helicopter. If you're running low on fuel, return to the tanker. If you've lost communication with the flagship, default to pre-planned contingencies.
These agents fight each other?
The adjudication is entirely algorithmic — there's no White Cell making judgment calls, because at that scale you'd need thousands of umpires. The tradeoff is fidelity. An agent-based model can't capture the creativity of a human commander or the friction of a real staff trying to coordinate a complex operation. But it can surface emergent behaviors that no one anticipated. In one Global wargame, the Red force's drone swarm depleted the Blue force's surface-to-air missiles so quickly that the Blue commander was left with no defense against a follow-on strike. Nobody scripted that outcome. It emerged from the interaction of thousands of individual agent decisions.
Which brings us to the limits of computer modeling. You mentioned earlier that every model embeds assumptions. What kind of assumptions?
The model has to assume how the enemy will fight. Will they use their submarines aggressively, pushing into the carrier's screen, or conservatively, holding back to protect their own fleet? Will they use their anti-ship missiles in salvos or in waves? Will they prioritize the carrier or the logistics ships? These assumptions are baked into the agent behaviors, and they can systematically favor one side. If the model assumes the enemy will fight the way the U.Navy fights, and they fight differently, the model's predictions are worthless.
The simulation isn't objective — it's just someone's assumptions, mathematized.
And it's why the White Cell still exists, even in an era of sophisticated computer modeling. The human umpire's job is to catch the moments when the model is doing something that doesn't match reality. A submarine hiding in a thermocline that the acoustic model doesn't handle well. A surface ship maneuvering in a way the physics engine can't replicate. A command decision that the agent-based model's doctrine doesn't account for. The umpire says "the model says X, but in reality, Y would happen." And that judgment becomes part of the adjudication record.
The Swiss sub incident — walk me through the actual timeline. Twenty eighteen, off the coast of Florida. What was the exercise?
It was a composite training unit exercise, or COMPTUEX, which is the final certification event for a carrier strike group before deployment. The Swiss Type 209 submarine was invited to participate as an opposing force — a surrogate for the kind of quiet diesel-electric submarines that potential adversaries operate. Navy doesn't have any diesel-electric submarines of its own, so it regularly invites allied submarines to play the Red team in ASW exercises. The Swedes have been doing this for years with their Gotland-class submarines, which are so quiet that they've repeatedly "sunk" American carriers in exercises.
The Swedes and the Swiss. Neutral countries with absurdly capable submarines.
There's a pattern here. Countries that don't plan to project power across oceans invest in coastal defense, and quiet submarines are the ultimate coastal defense weapon. A nuclear submarine is optimized for speed and endurance — it needs to cross the Pacific and stay on station for months. A diesel-electric is optimized for silence and short-range ambush. In a littoral environment, the diesel-electric has the advantage. The Swiss sub incident just demonstrated that the advantage extends further out to sea than anyone expected.
Let's talk about the Army side for a moment. You mentioned One Semi-Automated Forces earlier. What's the ground equivalent of all this?
OneSAF is the U.Army's constructive simulation system. It can model individual soldiers, vehicles, and helicopters down to the level of specific weapon systems. A battalion-sized engagement — about a thousand soldiers — requires roughly fifteen minutes of computation per minute of simulated time. So a four-hour battle takes about sixty hours to simulate. That's the fidelity-tractability tradeoff. If you want high-fidelity physics, you pay in computation time. If you want real-time results, you use reduced-order models that approximate the physics with lookup tables and statistical distributions.
Fifteen to one. So if you're running a corps-level exercise with fifty thousand soldiers, you're not using OneSAF at full fidelity. You're using something coarser.
The Army uses a hierarchy of models. At the strategic level, you might use a highly abstracted model where divisions are represented as single entities with aggregate combat power scores. At the operational level, brigades and battalions with stochastic attrition models — basically, Lanchester equations that calculate how many tanks each side loses per hour of combat. At the tactical level, individual vehicles and squads with physics-based weapon effects. The art of wargame design is choosing the right level of abstraction for the learning objective.
That's a name I haven't heard since...
Frederick Lanchester, a British engineer who developed differential equations to model aerial combat in World War One. The basic insight is that the rate at which a force loses units is proportional to the number of enemy units firing at it. If you've got a hundred tanks fighting fifty tanks, the fifty-tank force loses units twice as fast. It's a crude model — it doesn't account for terrain, training, morale, or tactics — but it's computationally trivial and gives a reasonable first-order approximation for large-scale attrition. Most operational-level wargames still use some variant of Lanchester equations under the hood.
We've got Lanchester equations from nineteen sixteen, Monte Carlo methods from the Manhattan Project, six-degree-of-freedom physics from aerospace engineering, and human umpires with colored flags from the eighteen eighties. Wargaming is basically a museum of mathematical techniques, all wired together.
That's the remarkable thing. The hybrid approach — computers for the physics, humans for the judgment calls, statistics for the uncertainty — produces outcomes that are useful enough to inform billion-dollar procurement decisions and operational doctrine. The Swiss sub incident didn't "prove" that U.ASW was broken. It generated a hypothesis: under certain conditions, a quiet diesel-electric submarine can penetrate a carrier group's screen and achieve a firing position. That hypothesis was then tested with further exercises, analysis, and eventually changes to ASW tactics and sensor deployment patterns.
That's the key insight, isn't it? Wargames aren't predictive tools. They're hypothesis generators.
And that's the misconception that drives me up the wall when I see wargame results reported in the news. "Wargame shows China would defeat U.in Taiwan conflict.A wargame showed that under a specific set of assumptions, with a specific adjudication method, and a specific scenario design, the simulated Red force achieved its objectives. Whether that translates to real-world outcomes depends on whether the assumptions hold, whether the adjudication was realistic, and whether the scenario captured the actual operational problem. Most headlines strip all of that nuance.
For someone reading about a wargame result, what should they ask?
First, was this free-play or scripted? If it was scripted, the outcome was designed to teach a lesson, not to measure relative capability. Second, what was the adjudication method? Human umpires, computer models, or a hybrid? Human-umpired games are vulnerable to bias; computer-only games are vulnerable to model error. Hybrid is best but hardest to do well. Third, what was the learning objective? Was the wargame designed to test a specific tactic, evaluate a new weapon system, or train decision-making? The objective determines what the result actually means.
Free-play versus scripted, adjudication method, learning objective. That's a useful mental checklist.
It applies outside the military too. Cybersecurity red-teaming uses the same logic — a red team simulates an attack, and a white cell adjudicates whether the blue team's defenses would have detected and stopped it. Financial stress testing uses Monte Carlo simulations to estimate how a bank's portfolio would perform under extreme market conditions. Autonomous vehicle testing uses simulated pedestrians and obstacle scenarios where the "adjudication" is whether the vehicle's perception system correctly identified the hazard. The "simulate the effect, not the action" paradigm is everywhere once you start looking for it.
Simulate the effect, not the action. That's the phrase, right there. The torpedo was never in the water. What was simulated was the effect — a hit, a miss, a kill — based on everything we know about torpedo performance, target vulnerability, and the tactical situation. The action is irrelevant. Only the outcome matters for the learning objective.
That's why the Swiss sub incident landed so hard. The effect was unambiguous. A carrier was removed from the fight. The action — bottom creep, undetected approach, short-range torpedo launch — was entirely plausible. There was no simulation artifact to hide behind. Just a cold, clear signal that something in the ASW posture needed to change.
Let's talk about where this is going. The episode plan mentioned the Joint Simulation Environment, the JSE, which is the Department of Defense's push toward fully automated adjudication for air combat. But it also said submarine warfare remains stubbornly human-in-the-loop.
The JSE is fascinating. It's designed to support high-end air combat training with fully simulated adversaries — F-35s fighting AI-generated opponents in a synthetic environment so realistic that pilots can log training hours in the simulator that count toward their real-world qualification requirements. The adjudication is entirely algorithmic because air combat happens too fast for human umpires to track. A within-visual-range dogfight is decided in seconds. You can't have an umpire saying "hold on, let me check the table.
Underwater, it's slower.
A submarine engagement unfolds over hours or days. The detection phase alone can last twelve hours as the submarine creeps into position at four knots. The human umpire has time to think, to consult the models, to weigh edge cases. And the underwater acoustic environment is so complex — so sensitive to local conditions that no two patches of ocean sound the same — that pure computer models still struggle. The thermocline, the salinity gradient, the seafloor composition, the biological noise from snapping shrimp and whale songs — all of it affects sonar performance in ways that are hard to parameterize.
Of course there are.
Snapping shrimp are a genuine problem for sonar operators. They produce a broadband popping sound that can mask submarine signatures in shallow water. Some sonar models include a "shrimp layer" parameter.
I love that the U.Navy has a shrimp parameter.
The world is messy. Wargames try to capture the mess. Sometimes they succeed, sometimes they don't. The Swiss sub incident succeeded because the mess worked in the submarine's favor and the White Cell let it stand.
As AI-generated wargame opponents get more sophisticated, what happens to the human umpire? Do they become obsolete, or more critical?
I think they become more critical, but their role changes. Right now, the umpire is catching edge cases the model misses. In the future, the umpire will be catching edge cases the AI opponent exploits. An AI that's trained on millions of simulated engagements will find tactics no human has ever thought of — and some of those tactics will be simulation artifacts, things that work in the model but not in reality. The human umpire becomes the reality check on the AI's creativity. "That's a brilliant tactic, but in the real ocean, the torpedo's wake-homing sonar would lose lock in that thermal layer." That judgment can't be automated, because it requires understanding the gap between the model and the world.
The future of wargaming is an arms race between AI creativity and human judgment. That's either reassuring or terrifying, depending on how much you trust the humans.
I trust the humans who've spent twenty years operating submarines in actual oceans. The problem is there aren't that many of them, and they're expensive to keep in the loop. The institutional pressure will be to automate them out, save money, run more simulations faster. And that pressure will produce wargames that are computationally impressive and operationally meaningless.
Like Millennium Challenge 2002, but at machine speed.
The lesson of Millennium Challenge isn't that the Red team was clever. It's that the institution's response to an uncomfortable outcome was to change the rules. AI won't fix that. It might make it worse, because an AI-generated outcome is even easier to dismiss. "The model was overfitted," "the training data was biased," "the neural network hallucinated." The human umpire, with all their flaws, at least has to look you in the eye and say "I believe this is what would have happened.
That's a good place to land. But before we wrap, I want to circle back to something you said about hit probabilities. You mentioned that an eighty percent hit probability in a wargame doesn't mean eight out of ten shots will hit in real combat.
The hit probabilities in wargame models are conditional on the specific scenario parameters — range, target speed, countermeasures, sea state, and so on. More importantly, they're calibrated against peacetime test data. The Navy fires test torpedoes against instrumented targets under controlled conditions. The torpedo knows it's being tested; the target isn't maneuvering aggressively; there's no jamming, no decoys, nobody shooting back. The hit probability derived from those tests is an upper bound. In combat, with a terrified crew, countermeasure deployment, and a target that's maneuvering for its life, the real hit probability is almost certainly lower. How much lower? Nobody knows, because we haven't fought a naval war against a peer adversary since nineteen forty-five.
The probabilities are best guesses wrapped in physics.
Best guesses wrapped in physics, validated against peacetime tests, and delivered with a confidence interval that's wider than anyone wants to admit. The honest answer to "will this torpedo hit" is "probably, under these conditions, with about a twenty percent chance that something we didn't model changes the outcome." That's not satisfying, but it's true.
The Swiss sub, firing from inside five nautical miles with no warning, had a probability that was basically "yes.
And the Navy took it seriously because the conditions weren't cherry-picked to make the submarine look good. They were realistic ASW conditions that the carrier group was expected to handle. The fact that they didn't handle them was the signal.
I think we've earned a fun fact.
Now: Hilbert's daily fun fact.
Hilbert: In the seventeen twenties, a Russian expedition to the Yamal Peninsula documented a species of orb-weaving spider whose silk exhibits a peculiar property — when stretched to precisely one hundred thirty-seven percent of its resting length, it emits a faint ultraviolet glow visible only in complete darkness. The expedition's naturalist, a German physician named Wilhelm Steller, preserved this observation in a marginal note of his botanical manuscript, where it remained uncatalogued until a digitization project at the Russian Academy of Sciences rediscovered it in two thousand nineteen.
A hundred thirty-seven percent.
Glow-in-the-dark spider silk from the seventeen twenties.
Here's the forward-looking thought. The Swiss sub incident is now eight years old. In the time since, the U.Navy has invested heavily in new ASW sensors, unmanned surface vessels designed to trail submarines, and AI-assisted acoustic analysis. The question isn't whether those investments closed the gap that the Swiss sub exposed. It's whether the next wargame will surface a gap nobody's even looking for. Because if the wargame is working properly, it will.
The corollary: will the institution listen when it does? That's the open question. Not the technology, not the models, not the adjudication. The willingness to hear something you don't want to hear, from a simulation you paid millions of dollars to build, and then act on it.
That's the one that keeps me up at night. If sloths had nights. Which we do.
Thanks to our producer Hilbert Flumingtop for the fact, and to everyone listening for sticking with us through sonar equations and Monte Carlo draws. This has been My Weird Prompts. If you enjoyed this episode, leave us a review wherever you get your podcasts — it helps other people find the show.
We'll be back next week. Try not to get sunk by a diesel-electric submarine in the meantime.
