He sees a group of nefarious actors congregating outside my field of view. I slew the camera and count the number of weapons. Who had better situational awareness (SA)?

It doesn’t matter. What matters is: should the convoy going to press forward or not? And neither of us could answer alone. But what if it were possible to design a system that leveraged human senses and could capture information at high resolution?

Computers can process thousands of dimensions but can't tell you which ones matter. Humans know what matters but have limited bandwidth. We tried to close the gap.

Below is a crude visualization of an aircraft flying through altitude and heading bounds with a topdown view on the right. Ugly yes, but able to serve as a primitive for our design, also yes.

The Caveman Brain

Many have admired the problem of information overload in the modern age. Our brain’s hardware was designed for survival 10,000 years ago and excels in pattern recognition, spatial awareness, threat detection. When under stress, deliberate reasoning, sequential analysis — becomes less available, not more. (Arnsten, 2009, on stress and prefrontal cortex availability is the clinical grounding here.) We all know this. What you may not have considered is a way to depict dense high-dimensional data in a picture can help close the gap.

Shannon formalized the ceiling in 1948: every channel has a fixed capacity, and past that capacity, more input degrades performance rather than improving it. The human visual system is a channel. The question is not how to put more on the screen. It is how to put less without losing anything that matters.

This is what I call decision density: the number of meaningful choices a display forces per unit of operator attention. Get it right and you multiply human judgment. Get it wrong and you’ve built an expensive machine that no one can effectively supervise.

The Math

The starting point is Principal Component Analysis. PCA takes highly dimensional data — think thousands of variables — and produces an ordered list of the directions that carry the most variance. Not all dimensions are created equal. The first principal component explains the most information. The second explains the most of what’s left.

The useful property: diminishing returns kicks in fast. In most real-world datasets, the first two or three components capture 80–90% of the meaningful structure. The rest is noise, redundancy, or detail that doesn’t drive decisions at speed. Furthermore, in the following example we decompose hundreds of variables into 10 PCAs.

Applied to sensor data, the question becomes: if you could only give the operator six numbers instead of a thousand parameters, which six carry the most information? PCA finds those six. They’re orthogonal — independent of each other, no redundancy — which means the operator can read each one without correcting for the others.

Hundreds of variables are now in 10 PCAs. But more importantly, we see that we get 60+% of information at 3. Meaning, a 3D display like we had before is a great tool.

Testing it Live

I needed to know if this approach was specific to aircraft or solving something more general. I needed a live, adversarial, highly dimensional environment to test it. Combat missions don’t run on weekends. Financial markets do.

Cryptocurrency specifically — no live trades, no financial advice, not a recommendation — just the data structure. A live order book with 22 instruments updating at millisecond intervals has the same fundamental architecture as an autonomous aircraft sensor suite: more variables than any human processes at speed, an adversary on the other side, a display that determines whether you lead the tempo or chase it.

I built the display. By making a lot of mistakes first, I can save you time.

Here’s what didn’t work first:

• Lines between dots. First thing I considered and total dud. Can’t see what’s connecting without close inspection, which defeats the purpose.

• Everything a spectrum. PCA is not binary, but brains cannot tell the difference between 4Hz and 8Hz blink. We set threshold values for PCA 4+.

• Movement too fast. The caveman brain reads spatial change as threat. Instruments updating faster than ~250ms triggered urgency without signal.

• All the colors. We tried shapes and more colors. This was max distinguishable.

Unused Human RAM

Here’s something that doesn’t get said enough about remotely piloted operations: the pilot has sensory bandwidth the manned pilot has already spent.

A fighter pilot uses their vestibular system, peripheral vision, the seat-of-the-pants feel of the aircraft — all of it processing continuously, all of it consuming cognitive bandwidth. The remotely piloted pilot has none of that. Usually framed as a disadvantage. I’d argue it’s unused RAM.

We tested two additional sensory channels alongside the visual display. The table below shows the full result: ten PCA dimensions, three sensory channels, zero hours of training required to execute on.

Here’s what worked:

• Three primary spatial dimensions. Size of dot smaller for “depth.”

• Color for 4x other dimensions. Organized in radial sections. Plus blinking or not.

• Audible pitch and vibration that would work for any headphone.

Here’s a breakdown of what you’re seeing above:

Want to Make One?

Whether you’re in a program office, a PE firm, or an ops center, the framework is the same. Three questions before you start building:

1. What is the decision space you’re responsible for? Not what data is available — at the point of decision, what do i need to see is true/false or “enough?”

2. How related are your dimensions? More related means more redundancy. PCA is the perfect ideal of this where all axes are perpendicular, but do your best.

3. What is the decision timeline? A 250ms decision window — unstable aircraft, live market — requires a fundamentally different display than a 45-second window — infil/exfil call, sourcing decision. Design for the worst-case tempo, not the average.

Once you have those three answers, you know what to compress / expand. But leveraging spatial and sensory input can squeeze out much more.

Applying this outside the cockpit

DARPA’s Mosaic Warfare concept frames national advantage as the ability to make better decisions faster than the adversary — at every level, across every domain. (DARPA, “DARPA Tiles Together a Vision of Mosaic Warfare,” 2017.) I’d take that one layer deeper: decision dominance at the national level is the aggregate of individual decision-making quality across every operator, analyst, and executive in the system.

Most programs build the algorithm first and the display last. I built this in a weekend. The methodology is not complicated. But it requires most to unlearn old habits.

What high-stakes decision is your team flying blind?

“Three-quarters of the factors on which action in war is based are wrapped in a fog of greater or lesser uncertainty.” — Clausewitz

War is uncertain. Predicting outcomes takes more than modeling. You already knew that. But before I was a pilot, I was a math major. In that world, uncertainty is not only usable — it is a number. So, I couldn’t stop myself from trying a calculation. This week, I looked at 97 wars across 198 years to see how well the most cited mathematical model of combat holds up against reality.

The model

In 1916, Frederick Lanchester published a set of equations to describe modern armed conflict, later called his Square Law. The model says each unit can engage any opposing unit, so a force’s total firepower scales with its size, but its advantage in surviving enemy fire scales with the square of its size. Translated to math:

For:

α = effectiveness of each A unit against B (kills per unit per time), and

β = effectiveness of each B unit against A

Equal-quality forces α = β.

In English: the bigger your force, the faster you destroy the enemy and the slower you take casualties. A force twice as large does not have twice the advantage — it has four times the advantage. Whoever has more troops wins, assuming both sides fight with equal skill. Business strategists borrowed this model. Defense acquisition offices embedded it in cost calculations. It has a satisfying elegance.

To test it, I used three of the most widely-used academic conflict datasets in the field — compiled over decades by researchers at the Peace Research Institute Oslo and the Correlates of War Project, the closest thing the field has to a standardized record of conflict since 1816. I built a matched sample of wars where I had the force sizes on both sides and a clear outcome. After filtering for wars with unambiguous results and successful data matches on both sides, I had 97 interstate wars to test against.

The chart below shows the win rate for the larger force at each level of size advantage. The multiple is the size of Goliath relative to David. The 1.5–3x band covers all battles where Goliath is 1.5 to 3 times larger. The first thing that jumped out is highlighted in orange: the greatest win rate is not when you have the greatest numerical advantage. Not even close.

Source: COW MID v4.03, COW NMC v4.0. Interstate wars 1816–2014 with clear outcomes. Error bars show 95% confidence intervals. Dashed line = 50% baseline (coin flip).

It’s because we’re presenting a false proxy. Force size in this dataset is all military personnel across all branches. Unless this is an Avengers battle, the odds that all of army A fights all of army B in a single conflict is near 0%. But there’s something more interesting here than a bad data problem.

If it was simply a matter of not accounting for excess supply, the other bands would all be roughly blue — near the coin flip line. But they’re not. At the extreme end, having more forces relative to your adversary actually appears to disadvantage you. Think the Vietnam War. Think Thermopylae. Many possible explanations: armies may be harder to coordinate at that scale, a commander’s attention gets divided too far, multi-front exhaustion sets in.

The twist

Like shooting the moon in a game of hearts, it seems that if you are overmatched, you have better odds fighting on terms with a heavy numerical disadvantage than a mild one. I envision a team of Navy SEALs. That’s what the model tells us — but is it out to lunch?

To measure the model’s predictive strength I used AUC — area under the ROC curve, or the rate of true positives for a given false positive rate. Higher is better. A coin flip is 0.5. A perfect model is 1.0. Using the ratio of forces gives an AUC of 0.577. Lame… kind of.

It’s a bad news story for a data analyst trying to predict an outcome. But it is a good news story for a commander in a nation far smaller than its adversary. Although we can’t claim you’ll win, we have no evidence that Goliath will either. If anything, the model shows David has an edge at the extremes. But is it real?

Russia lost to Estonia in 1920 at a 191:1 force ratio. Germany lost to Finland and Bulgaria in 1945 at ratios above 130:1. It has happened and within this dataset it’s legit. The problem is the sample size. An n of 6 is hardly enough to draw a conclusion. So: if you are the smallest kid on the block, be very cautiously optimistic. Your fate is not sealed.

What I saw

I’ll share a trend I noticed in combat. When I watched firefights develop in the Middle East, the victor was rarely obvious. But what I and the sensor operator often agreed upon is that we’d rather be on the side with the better position than the side with the numerical advantage. Terrain, time of day, the local population — all of it gets a vote. You can use it or be used by it.

The one thing we can say with some confidence: total army size alone will not win you a war. If anything, it could be a hindrance. I’ll keep tracking down what does make the difference, mostly through process of elimination. But if you’re outnumbered by your adversary, you should be less discouraged after reading this than before.

This work is not part of my official duties as a USAF officer. Views expressed are my own and do not constitute endorsement by the Department of War, Department of the Air Force or the U.S. Government.

“An approximate answer to the right question is worth far more than the precise answer to the wrong question.” — John Tukey

Ask ten pilots how many unmanned aircraft one person can control, and you’ll get twenty answers. I’ve heard everything from “it depends” to confident declarations of specific numbers — often from people who’ve never flown them in combat. This is my attempt at a third option: the question itself is mostly wrong.

A quick definitional note before we go further. When I say “pilot,” I mean the person who owns the life-and-death consequences of what the aircraft does. If a fifth-gen fighter pilot is issuing intent to an AI wingman, that fighter pilot is the pilot. If it’s Ender’s Game and a four-star general is clicking approve on a targeting decision from a command post, the COCOM is the pilot. If this terminology becomes obsolete before this article does, substitute accordingly.

The Intuitive Frame and Why It Breaks

Earlier in my career, I thought about this the same way most people do: how many aircraft can one person manage simultaneously? It’s a clean, measurable, resource-planning question. Program managers love it. Acquisition officers put it on briefing slides. It feels like a constraint you can design a system around.

After more than 1,000 hours flying MQ-9s in the Middle East, it stopped feeling clean. An ISR mission might give you four hours of dead airspace over a desert compound where nothing moves, then ten seconds where everything happens at once. Close air support is the extreme version — a single missed radio call, a location update that doesn’t come through, and an innocent person dies. The aircraft count didn’t change between those two scenarios. The decision demand swung from near-zero to overwhelming in under a minute.

That observation broke my simple model.

Decision Density: The Better Proxy

The frame I now use is decision density — the number of human decisions required per unit time. It’s not a number I can hand you on a slide. It’s a variable, and it’s almost entirely mission dependent. Research published through the Air Force Research Lab has been gradually shifting focus from “how many agents can a human manage” to “how can agents and humans work together effectively” — because the former treats the ratio as static when the actual bottleneck is cognitive load under pressure.

This matters because decision density is, in large part, software defined. Every major software revision on a CCA-class platform changes how much the pilot needs to babysit the automation — sometimes dramatically. The Air Force’s DASH experiment — Decision Advantage Sprint for Human-Machine Teaming — is specifically structured around the insight that the limiting factor isn’t aircraft count; it’s the speed and quality of human decision-making in high-tempo scenarios. Design a system around a fixed pilot-to-aircraft ratio today, and you may have precisely solved last year’s problem.

What I Actually Saw in the YFQ-42A Program

I flew the YFQ-42A during its early test days. The honest version of what I observed: the aircraft ratio question was almost never the interesting one. The interesting question was always “what is this aircraft going to ask me to decide, and when?”

Some flight profiles had me barely in the loop — the automation was doing its job, I was monitoring. Other test cards had me making calls at a rate that would have been unsustainable with more than one airframe. Neither scenario maps cleanly to a ratio. And here’s the part that should keep acquisition officers up at night: a new software release could flip which profile you’re in. Every three to six months, a new rev will change the answer.

Two Factors That Actually Matter

If I were advising a board on this, I’d frame the answer around two variables rather than one number.

The first is timeline. Technology designed to win a conflict in 2027 looks entirely different from a 2035 or 2050 solution. Committing to a ratio now is fragile. Each architectural decision that treats it as fixed will need to be unwound with the next software rev.

The second is the moral weight of the decisions involved. This isn’t soft philosophy — it has hard implications for system design. Defense policy experts have consistently argued that regardless of what automation can technically handle, decisions with direct human life consequences require a human authority. That’s not a legal formality. It’s a load-bearing constraint that shapes workload, interface design, and latency requirements from the ground up. The USAF plans to spend more than $8.9 billion on CCA programs between FY2025 and FY2029 — and how they architect human decision authority in that spend will define the answer to this question more than any pilot’s intuition about aircraft counts.

A Rough Taxonomy, Not a Formula

Since people will ask: here’s how I calibrate it by mission type.

Ocean surveillance, non-kinetic ISR, low-threat environments where the automation is primarily scanning and reporting: very high ratios are probably viable with the right design. Hundreds, possibly more. Decision density is low, and most system failures can be managed with time to spare.

Armed escort — loosely coupled to manned elements, standing by for potential engagement: somewhere in the range of ten to fifteen until the situation changes. The density can spike without warning, which sets the ceiling.

Tight formation with manned fighters, fully integrated into a kinetic engagement: one. Maybe.

These aren’t engineering specs. They’re calibrations, and they’ll move. Embracing “I don’t know my decision density” is a better use of resources than spending billions designing around a number that’s going to change with the next software patch.

The right question isn’t how many planes. It’s how many decisions — and how fast.

“The only thing we have to fear is fear itself” – FDR (not a pilot)

Which is more dangerous, a lightning strike or a traffic jam? A test pilot needs to know if the risk they’re taking is like one or the other.

Who cares what pilots at the Air Force Test Center (AFTC) think? Answer: anyone who wants the US Air Force to use their tech for combat with a major acquisition program.

AFTC pilots like Chuck Yeager flew these missions when one of them would die each week. This helped aviation combat deaths fall 25% in WW2 to ~6% by the 1960s. But in 2025, 548 deaths occurred amongst millions of aviators, compared to 27,365 deaths from automobile crashes. People are considering whether we need the AFTC anymore.

Whether you’re wasting bad guys or writing contracts, you should do so knowing that Air Force culture deeply respects these high stakes low odds events, i.e. airplane crashes. The game is breaking them down to scenarios that are either negligible odds (lightning strike), or acceptable stakes (traffic jam).

“Ritual is what you get when you forget why a rule exists.” – Nassim Nicholas Taleb

In The Right Stuff, Tom Wolfe describes a pilot’s career ascent as taking steps The Great Ziggurat towards ultimate glory, being an astronaut. Though technology has changed since 1957, a similar hierarchy has endured in the US Air Force. Some uncertainty exists defining intermediate steps, but one thing is clear: fighter pilots are the top and drone pilots are the bottom. But what if a jet is both a fighter and a drone?

Motivated, positioned to make a difference and understanding the stakes at hand, I promptly made the worst series of intrapersonal mistakes in my career. Because I was medically disqualified from manned flight, I saw test pilot school as my first chance to compete head-to-head in the big leagues, live the victorious underdog narrative and sow the seeds of what I called the “drone mafia.”

After six months, it was clear that my “if you can’t join them beat them” mindset needed to die. Fast forward three years. The commander asks me “how did you thrive where others couldn’t get past the door?” I’ll borrow phrases from two friends, one drone pilot and one fighter pilot. First is that relationships formed is the secret sauce, not the tech. Second is that air dominance requires an AND solution, not an OR.

Every community has a hierarchy, either conscious or unconscious, what’s yours?

If we treat CCA like an MQ-9 or RQ-4, the pilot in command, likely the fifth gen fighter pilot, is to blame. If we treat them like an air-to-air missile, perhaps the manufacturer or the test team could hold the bag. Before we shake our fists at the Wright Brothers for making this possible, consider this:

Human vs autonomous actions is largely a difference with no distinction in this case. Barring a hardware or software malfunction, actions are simply a model’s output to inputs. A human chose model’s structure, the training data, and most importantly when it was good enough to pass test and be used in anger.

Will CCA make the need for human control of moral decisions in warfare obsolete? Almost certainly not. Will it force us to think harder about who owns life and death decisions, yes.

If you would like to see more, please subscribe.

Resolving signal from noise can be life or death overseas. We all agree aircraft systems should be tested in “operationally representative environment.” Unlike hardware capabilities, software has the potential to field with extraordinary speed. But while the laws of physics rarely change, the adversary’s means of conducting warfare will.

For the past decade I have published research, led combat missions, and test piloted large unmanned military aircraft. This is not part of my official duties.

Plan is weekly posts which encourage myself and my followers to think more deeply about integration. Sub-systems, military force structure, and human morality must all be aligned to create something great. This identifies misalignments by questioning assumptions, beliefs and “best practices.”