Why AID Algorithms Don’t Improve: The Category Error

Control theory doesn’t apply to complex systems.

A satirical take on how users conform their lifestyle to help AID algorithms perform better. Concept and art direction by Dan Heller; AI-assisted rendering by ChatGPT.

This year, 2026, marks ten years since Medtronic published the results of the first clinical trial that used a hybrid closed loop (HCL) automated insulin delivery (AID) system, which was conducted in 2015. The “hybrid” part means that the user inputs data for meals. The group’s mean A1c was ~7%.

Medtronic’s chief medical officer declared that the system “should be safe and more effective than what the person with diabetes can do on their own.”

Ten years later, with multiple revisions of the algorithms, and five commercially available AID systems on the market, independent researchers have conducted trials to compare how all five perform: A comparative study from Ohio State (2024), the CLEAR trial at Cambridge, a crossover trial at McGill, and a January 2025 overview in diaTribe. They converge on the same conclusions:

• HCL (hybrid closed loop) with human input (70% TIR): ~7.0% A1c

• FCL (fully closed loop) without human input (50% TIR): ~8.0% A1c

(Note that TIR (time in range) maps to A1c levels via a formula that Beck et al. published in Diabetes Technology & Therapeutics, 2019, where they analyzed 545 adults with T1D across four randomized trials.)

Ten years. No progress.

Sure, but at least it must have been better than the old style of multiple daily injections, right?

Nope. In the 2025 paper, Shining the Spotlight on Multiple Daily Insulin Therapy: Real-World Evidence of the InPen Smart Insulin Pen (Diabetes Technology and Therapeutics), 3,793 T1D users on smart insulin pens with CGMs achieved 70.3% TIR — an A1c equivalent of 7.0% — when they dosed at least three times daily.

And InPen is owned by Medtronic. Yeah, them.

Perhaps unsurprisingly, it’s been this way since before automation as well. A research study by the T1D exchange from 2018, which compiled data from US-based clinics prior to the availability of automation, and found that users over the age of 26 saw nearly identical A1c levels between pump users and those on MDI: ~7.3%.

So, before automation: 7.3%. Ten years after automation: 7.0%.

But there is one thing that automation has done that’s worth recognizing. Look at the data from that 2018 study, and notice the A1c levels for those under the age of 26.

Those on the younger side had higher A1c values, and it’s exactly those people who benefit greatly from automation. People who either cannot or will not engage in self-management can see their A1c levels drop from 9%, 10%, and considerably higher, all the way down to 8%. It’s that subgroup of people who are the main beneficiaries of AID systems and that you see cited in most papers published by pump companies and others that promote these devices.

And this highlights something often hidden in the details of all these studies: the great variability among subgroups. Those who come from distressed socioeconomic backgrounds have less access to healthcare and resources, and less support from family/friends. People in this group are less informed about good self-management techniques, and are less engaged. So, they tend to have worse A1c and TIR percentages. Those who come from more stable environments, access to resources, and are educated about self-management are also more engaged with their disease. So, they tend to have better A1c and TIR numbers.

The point is that human engagement yields better numbers and healthier outcomes, regardless of method of insulin administration. This has been true since insulin was discovered in 1929.

But the promise of automation was supposed to break that barrier—that diabetes could be put in the background, that people could finally be relieved of the daily burden of self-management. Yet, after ten years of these systems, that hasn’t happened. WHY?

Perhaps we should start by asking a more fundamental question that no one in the development cycle ever asked: Is this problem even solvable in principle?

Complicated Systems

Before mathematicians attempt to solve a problem, they first have to determine what type of problem they’re trying to solve. This isn’t philosophical hand-wringing, it’s the foundation of serious scientific work. Some problems are merely difficult; others are impossible. Knowing the difference before you start saves decades and billions of dollars. And in this case, lives.

Since we’re talking about controlling glucose levels using a CGM and an insulin pump, we’re in a domain of science called control theory. That is, there’s a sensor that detects glucose levels, and there’s a controller (an algorithm) that delivers insulin. There’s a cause and effect at play, and that interplay is the core thesis of control theory. In fact, we engage with automation everyday.

When you’re driving a car and flip on the cruise control, an algorithm determines the speed of the car and the amount of gas being delivered. These are the sensor and the controller. By constantly monitoring the speed, the algorithm knows just how much to add or let up on the gas. The hills, valleys, and other road conditions may cause more or less gas to be applied (or the brakes), but this is the closed loop.

Now think of a fully automatic self-driving car. It has lots of sensors, lidar, cameras, sonar and other gadgets around it, constantly monitoring the environment. The controllers are not just gas and brakes, but the steering wheel.

Now think of an airplane, or the telephone system, or an oil refinery. Big and complicated. Indeed, they’re a system of independent parts. To build a controller, you have to start with statistical models from physical properties of the parts of the systems you’re observing and controlling. Those parameters follow what is called a lognormal distribution — a pattern where values can’t go below zero, most readings cluster in a typical range, and occasional high values stretch out in a long tail.

You know that you can control the systems if the behaviors of the parts are stable, stationary and independent of one another. Thus, engineers can characterize the system mathematically and build controllers with known performance bounds. It applies to the car, the airplane, telecommunications systems, oil refineries, and more.

The system’s statistical signature is a property of the system itself, not of its context. You can identify the transfer function, it stays identified, and your controller works tomorrow the same way it worked today.

With this, you can create a simulator that looks, acts, and can be reliably used to resemble your system. Indeed, these simulators are important because you want to test a jet engine before putting in a billion dollar airplane. Or putting living passengers in it.

These simulators exhibit a property of complicated systems: They’re predictable, which is also a way of saying they’re tractable. You can not only calculate each of the properties, those calculations allow you to predict future states of the system.

Many medical devices fall into this category. The canonical examples include pacemakers, cochlear implants, ventilators, and even the NeuroPace responsive neurostimulation system for epilepsy. In each case, the control task has the properties you’d expect of a complicated system: the relevant biomarkers are directly measurable with minimal lag, the input-output relationship is well-characterized, the physics are stable, and the “actuator” operates on the same timescale as the disturbance. A pacemaker detects an arrhythmia and delivers an electrical pulse within milliseconds. The system it’s controlling — cardiac electrical conduction — follows predictable electrophysiology.

Closed-loop AID algorithms were designed with the same engineering logic: If you look at CGM data, it sure looks like a lognormal distribution: Glucose values can’t go below zero, they cluster in a typical range — say, 100-200 mg/dL — and there’s a long tail that goes higher. And this is the basis for assuming it has the properties for control: A sensor, a controller, and a lognormal distribution of data. Indeed, we can simulate different physiologies from huge datasets.

But there’s one hiccup here: In glucose control, the “parts” are not independent. They’re integral to one another. These parts include regulatory and counter-regulatory hormones, cortisol, glucagon, and a litany of other pathways that affect glucose levels, and each other, which feeds back to glucose levels. The way each component behaves affects the others in ways that no longer make them “stable and predictable.” This is an entirely different category from the complicated system—it’s a complex system. (So, it’s less of a hiccup as it is a loud, audible belch.)

Complex Systems

When a complicated system turns into a complex system, it’s a new category of system, where it may well be calculable and simulatable, but not predictable, and therefore, not controllable. It is a central characteristic of chaos theory. Glucose homeostasis exhibits the defining properties of those systems—sensitive dependence on initial conditions, non-stationarity, path dependence, and unmeasurable state variables (all those hormones and other physiological properties that affect glucose levels). Therefore, the theoretical ceiling on prediction is a property of the system, not a limitation of technology.

To see how a complicated system becomes a complex system, consider the pendulum on a grandfather clock. Completely calculable and predictable. You can calculate its position at any point in the future. Engineers solved this in the 17th century.

Add a second pendulum at the bottom of the first pendulum, and watch what happens: As the main pendulum swings, it causes its extension to swing, and that swing feeds back onto the motion of the first pendulum. By definition, the integration of the way the parts affect one another turns it into a completely unpredictable, chaotic system.

Same physics, same materials, one additional hinge. Two double pendulums released from what looks like the same position will diverge within seconds into completely different paths. No equation can tell you where any part of the larger system will be seconds from now—not the tip, or the base joint, or anywhere in between. This type of system is intractable. And it’s entirely because the parts are not independent of one another. It’s the feedback loop that makes each part itself change its own behaviors.

Nothing got “harder” in the engineering sense. Nobody added a new force of nature. It’s the same gravity, same joints, same metal arms. What changed is the category of system—from tractable to intractable. And no amount of better measurement, faster computation, or more precise manufacturing will make it predictable again.

Here’s where things can get insidiously deceptive that engineers need to be highly aware of or they’ll fall into a trap: Because you know the physics of all the parts, you can still simulate the system. You can take a million videos of a double-pendulum, collect all the statistics—all of which will look lognormal in their distributions—thereby allowing you to write a simulator that models a double pendulum. More deceiving is that the simulator will indeed exhibit a behavior that could very well be one of an infinite number of possible behaviors that could happen.

And this is what it all comes down to—there is no chance that the simulation can be used to predict the behavior of a real double-pendulum that looks like it’s starting with the same initial conditions. You can read how to do it and watch animations of simulations on the Wikipedia page. The math is extraordinarily complex — almost worth looking at the page just to appreciate that — but the ability to simulate a system’s behavior has nothing to do with the ability to predict its outcomes. And if you can’t predict the states of the parts, you can’t control them.

The problem with trying to predict glucose—and this was the category error that started all this—is that it appears to have a lognormal distribution of data. So engineers treat it like a complicated, controllable system. The CGM is the speedometer, and insulin is the “controller” that acts on it. Everything else is treated like “disturbances” — not as integrated parts, but as separable parts. You can collect zillions of hours of CGM data and create simulations of what appear to be real humans, then try to map those simulations into an algorithm that controls glucose levels and you get… No difference. The noise in the chaotic system just overwhelms the signal.

So how about this idea: Rather than control the whole system, why not tune some small aspect of a subsystem. Maybe to reduce hypoglycemia at night, perhaps? At least many of the variables are more stable—you’re not eating, you’re not physically active, you’re basically in the most “stable” state you can be in.

But again, even that is not possible because, unlike a double pendulum where the parts are fully observable, glucose metabolism has parts you can’t even see, let alone measure and/or simulate, even at night: The body moves between the sympathetic to parasympathetic nervous system, muscles and other tissues take up glucose through non-insulin mediated pathways, glycogen stores are replenished (without insulin mediation), and dozens of other activities are still swirling around.

Collectively, the interdependent parts have such an enormously disproportionate effect on glucose trajectories and rates of change—whether awake or asleep—having an algorithm try to do anything sensible is like having cruise control keep your car at a fixed speed while a hurricane lifts your car and drops it into the next town.

If you’re not overwhelmed yet, there’s another aspect to human biology that takes it yet another step further into complexity: it adapts. And that leads to the grand theory that’s been established for decades in the field of biology: complex adaptive systems.

Complex Adaptive Systems (CAS) Theory

The science of complex adaptive systems (CAS) recognizes that biological systems self-organize and follow their own internal rules. Interventions, such as medications or even psychological awareness, can have effects that may meet some objectives, but not all, and not predictably, and not for everyone.

Most importantly, biological systems adapt—they behave differently as conditions change. When you’re sick, hormones react and respond differently; if you’re pregnant; when you grow from childhood to young adult to middle age and beyond. Your physiology is never constant long enough for the system to be “stable”.

Even day-to-day physiology varies dramatically.

A useful illustration comes from a Stanford group (Hall et al., PLOS Biology, 2018) that used CGMs to monitor glucose variability in response to different kinds of foods. In a standardized-meal sub-study, a subset of 30 participants ate three predefined breakfasts (including cornflakes-and-milk) on two separate days each. Even with that kind of controlled input, post-meal glucose excursions differed markedly across individuals — as well as within the same individual. Below is a figure from that study:

A larger Israeli study from the Weizmann Institute (Zeevi et al., Cell, 2015) pushed the same point, but at a larger scale. The investigators used week-long CGM data in an 800-person cohort and captured tens of thousands of meal responses (46,898). They found high inter-individual variability in postprandial glucose response—even to meals that were ostensibly identical.

You may be reading that and think that you’ve experienced the same thing. Heck, T1D is crazy, right?

Here’s the kicker you didn’t expect: ALL of the subjects in these studies were healthy individuals with no diabetes diagnosis. They were screened to have no metabolic disease, they had low A1c, no physical disabilities of any kind. They were, as much as the study could determine, the perfect specimens of human health.

You can pick up your jaw off the floor now.

The takeaway is that glucose dynamics are more complex than ever assumed. The wild glucose excursions are the body itself—a healthy one at that—trying desperately to stabilize glucose levels. These are the properties of a complex adaptive system, and are the basis for CAS Theory, which has been explored in medical literature for decades. Examples include these citations:

• The challenge of complexity in health care (BMJ, 2001). Here, the authors state that the “behaviour of any complex system is fundamentally unpredictable”. I.e., one cannot model glucose as a superposition of linear transfer functions.

• From Crossing the Quality Chasm (Institute of Medicine/NCBI Bookshelf), the section Redesigning Health Care with Insights from the Science of Complex Adaptive Systems recognizes that biological systems are complex adaptive systems where detailed algorithmic prediction is “fundamentally not possible.”

• The Paradox of Intervening in Complex Adaptive Systems (PMC, 2018) directly addresses the paradox of trying to control systems that, by definition, resist external control.

Even data scientists are weighing in on this. A 2025 paper by Heman Shakeri, Assistant Professor of Data Science at University of Virginia, addresses this in his paper, The Driver-Blindness Phenomenon: Why Deep Sequence Models Default to Autocorrelation in Blood Glucose Forecasting:

Deep learning models for glucose forecasting “consistently fail to leverage clinically informative drivers—insulin, meals, and activity—despite well-understood physiological mechanisms.” The gain from adding these variables is “typically near zero.” They attribute this partly to “physiological heterogeneity that undermines population-level models.“

Unfortunately, the automated insulin delivery field has never engaged with this body of theory. Nowhere in the AID literature will you find serious discussion of complex adaptive systems, irreducible uncertainty, or the theoretical limits of algorithmic control over biological processes. Instead, you find engineering-style approaches, using control theory as a foundation for attempting to make these algorithms better.

The AID field has spent fifteen years and billions of dollars trying to build better grandfather clocks that have multiple pendulums, wondering why they can’t keep time.

And that’s why the algorithms never get better. The properties of a chaotic system will never permit it.

But it raises the obvious observation: If prediction doesn’t work, how are these systems working at all? How is it that people are ever in range? The algorithms must be doing something right. Right?

Indeed—let’s take a look under the hood at what the algorithms are actually doing.

Painted Into a Corner

If CAS theory tells us that glucose can’t be predicted or controlled, then what are these algorithms actually doing? And how is it that people are ever in range?

To answer that, it helps to understand just how constrained the engineering problem actually was — not because the engineers lacked skill, but because the system they were building left them almost no room to maneuver.

To illustrate, we’re going to use Tandem’s algorithms, but the same principles apply to all of them. And this is evident from the common A1c/TIR outcomes we see from clinical papers that publish these findings.

So, we begin at the low end of the glycemic range, where the floor was already high. Consider what happens with the simplest possible automated intervention: suspending insulin when glucose is dropping toward hypoglycemia. That’s all that Tandem’s Basal-IQ does — a process called Predictive Low-Glucose Suspend (PLGS). No dosing algorithms, no correction boluses, no meal detection. According to Tandem’s own study, users on Basal-IQ achieved a TIR of approximately ~64%. That single safety mechanism — stop delivering a dangerous drug when the patient is heading toward a dangerous low — captured the majority of the achievable benefit. Everything above that 64% floor is the territory where the actual algorithm operates.

And it feels like there’s a lot to go, right? Why not 90-100%? That’s where reality gets in the way.

First and foremost, SAFETY. This is where the ceiling is structurally low. To get tighter control, your model may suggest infusing insulin at times that may not be warranted, entirely because you can’t detect any of those pesky, unseen, unreliable hormones that may be affecting glucose levels. The algorithm might suggest a 5u bolus for dinner, but 30 minutes later, the person takes a post-dinner walk, and suddenly, they’re admitted to the ER. These unforeseeable aspects forced the potential for better control to be lowered by a lot. Add these up over time, and the ceiling on TIR drops.

Another aspect of safety is the reliance on CGM data, which is highly subject to unreliable variability. This is not because the CGM is inaccurate, per se. Glucose is highly volatile in the body, and detection comes with wide error bars. CGMs can be out of calibration, or just show readings that are dramatically out of alignment with physiological reality.—something that happens frequently when levels are rapidly moving up or down. An algorithm can’t just believe CGM data as gospel—it has to factor in that readings may be faulty, especially if the algorithm would indicate substantial correction doses. Indeed, overdosing insulin due to faulty high readings is not uncommon. These are all covered in detail in my article, Continuous Glucose Monitors: Does Better Accuracy Mean Better Glycemic Control?. Taking all this into account, and you shave off several more potential TIR percentages.

Safety demanded conservatism. Boris Kovatchev — the mathematician whose risk analysis framework underlies these systems — established decades ago that errors toward hypoglycemia are asymmetrically dangerous. A glucose of 40 mg/dL is a medical emergency; a glucose of 300 mg/dL is not. Any system that delivers insulin must be biased heavily toward under-dosing rather than over-dosing. This isn’t a design flaw — it’s the mathematically correct response to the risk landscape. But it necessarily caps how aggressively the algorithm can act, which translates to lower TIRs.

Between the high floor, the low ceiling, and the safety constraints, the operating space for any algorithm — no matter how sophisticated — was remarkably narrow. And, again, this applies to all algorithms.

This does not diminish the engineering that went into capturing those points. Kovatchev’s risk-asymmetric framework, individualized insulin sensitivity parameters, and a 30-minute prediction horizon based on CGM trend data are all the byproduct of years of simulation, clinical testing, and iterative refinement. The people who built it are serious scientists solving a hard problem with real ingenuity.

But here’s what the constraints actually produced. We can see it in the article, The Hybrid Closed-Loop System Tandem t:slim X2™ with Control-IQ Technology: Expert Recommendations for Better Management and Optimization:

• Below 70 predicted → suspend insulin

• Below 112.5 predicted → reduce basal

• Between 112.5 and 160 → deliver user’s programmed rate

• Above 160 → increase basal

• Above 180 → deliver 60% of a correction bolus, max once per hour

That’s what the algorithm reduces to once all the constraints bind. Not because the underlying engineering is simplistic, but because the operating space left almost no room for the sophistication to express itself. It’s as if you trained a concert pianist for twenty years and then asked them to perform in a room where the piano has five keys.

The 60% correction factor tells the story. If the system had genuine predictive confidence, why deliver only 60% of the calculated dose? Because the designers understood — correctly — that the prediction was unreliable, and the consequences of over-delivery were potentially fatal. That 40% haircut is Kovatchev’s risk framework in action: the system is deliberately conservative because the asymmetry of harm demands it. Tandem Diabetes Care itself describes this approach openly.

The results confirm the narrow operating range. In the study One Year Real-World Use of the Control-IQ Advanced Hybrid Closed-Loop Technology, approximately 9,000 users went from a baseline TIR of ~63% (on Basal-IQ) to about 74% on Control-IQ’s full system. Other studies converge on similar figures: roughly 70% TIR across all available AID systems. The entire contribution of the algorithm — all that engineering, all that science — accounts for roughly 8–12 percentage points of TIR improvement, depending on the population.

And notice how much of this is driven by the user, not the algorithm. In the paper Therapy Settings Associated with Optimal Outcomes for t:slim X2 with Control-IQ Technology in Real-World Clinical Care, researchers found that 67% of Control-IQ users had less aggressive settings than clinical guidelines recommend — meaning the system is already operating conservatively, and most users make it even more conservative. Nearly everything that determines how the system behaves — basal rates, correction factors, carb ratios — is user-settable, as it should be. The algorithm operates within those human-defined parameters, not instead of them.

Were the simulators necessary? Was any of this really necessary?

The safety concerns, plus the heuristics in the Control-IQ algorithm, prompts the obvious question: Was any of this necessary? Did we really need decades of research to come up with these algorithms? Could someone have arrived at essentially the same AID algorithm without any of the control theory scaffolding — without Bergman’s Minimal Model, the UVA/Padova simulator, Kovatchev’s MPC framework, any of it?

The answer is yes, and there’s direct evidence. Dana Lewis and Scott Leibrand started OpenAPS in 2014-2015. Neither was a control theorist. They didn’t build from the Minimal Model or the simulator. They started from the practical problem: “I have a CGM reading, I have a pump, what rules should govern insulin delivery?” From there, they built heuristic systems (oref0 and oref1) described explicitly as reference designs based on observable behavior, not physiological models. The outcomes are comparable to commercial systems. The Tidepool Loop FDA clearance used Control-IQ as its predicate device precisely because the FDA recognized functional equivalence.

The punch line sort of hurts: The entire intellectual apparatus — decades of physiological modeling, the simulator, the regulatory shortcut, hundreds of millions in R&D — produced algorithms that two people with no background in the underlying science replicated by empirical tuning. That’s not a knock on Lewis and Leibrand (it’s impressive). It’s evidence that the “algorithm” was never where the value lived. The operational space is so narrow that anyone reasoning from the same constraints — one input variable, one actuator, catastrophic downside risk — would land in roughly the same place.

The ceiling isn’t a function of which approach you take. It’s a function of what the system allows.

The Biggest and Best Variable: The User

This brings us to a painful irony. Any algorithm works best when the user conforms their lifestyle to meet the algorithm’s expectations: consistency and low volatility. And sadly, that’s a very unhealthy lifestyle.

A 2022 review of AID’s psychological implications — Nefs, Frontiers in Clinical Diabetes and Healthcare 2022 — documented exactly these dynamics. Users weren’t passively accepting the algorithm’s output. They were actively restructuring their lives to produce the numbers the algorithm rewarded: eating predictably, avoiding unannounced snacks, timing activities to coincide with low insulin-on-board, and — most consequentially — reducing or eliminating exercise that would introduce glucose variability the algorithm couldn’t handle.

Imagine a concerned parent telling their child when they can do certain things, like limiting physical activity or carbohydrate intake or just playing with their friends at a party — all with the aim of increasing time in range.

The behavioral adaptation produces TIR numbers that make the algorithm look effective, but the causality is partly inverted: the users are doing the work, and the algorithm is taking the credit. And the saddest part about this — the moral hazard — is that the user is doing the least healthy behaviors, all in the hopes of optimizing their TIR and A1c levels.

Shockingly enough, companies are starting to use this as a feature. In a press release, Diabeloop’s DBLG2 receives FDA and CE clearance for automated insulin delivery without meal input, the company says the quiet part out loud: removing meal announcements costs 3-6 percentage points of time-in-range, but they’re marketing it as a feature because it “reduces cognitive workload.” That’s behavioral conformity playing out in real time — except now the manufacturer is explicitly trading glycemic control for disengagement and calling it innovation. Diabeloop just got FDA clearance to remove the single most important point of user engagement: the meal bolus. And the FDA said yes.

The health consequences of this conformity have been appearing in medical literature since closed-loop systems were introduced. In my article, Medical Literature Analysis: The Performance Paradox of AID Systems, I cover ten years of research — much of it from journals published by the American Diabetes Association — showing that AID system users gradually develop most or all of the following: weight gain, loss of agency, de-skilling of self-management, limitations on efficacy, quality of life, economic impact, and the “attractive nuisance” of disengagement from self-management.

And if the user is entirely disengaged? TIRs hover around 50%, as the algorithm operates in what is effectively fully closed-loop mode — confirming that without human input, the system alone can’t get much past the safety floor.

Oh, and if we’re going to talk about engaged users, we have to recognize the DIY/looper community. They tend to do better than commercial AID systems because they can personally customize their algorithms to match their personal physiologies. This may work well for them, but are not generalizable to the broader population.

Personalizing your algorithm is the modern-day equivalent of MDI—you choose the dosing (basal and bolus) to match what your empirical experience has shown works. This is how T1Ds have always managed themselves, whether MDI or pumps.

The point is, the engaged user knows what works for them, and that should always be encouraged. T1Ds have always had the freedom to choose how much to dose at any given time, whether that’s done manually via MDI, or by personally tuning the algorithm. Whatever floats your boat.

Epilogue

This article took you on a very long journey, meandering through discussions of control theory, complicated vs. complex systems, and the rigorous nature of building physiological models that could simulate real diabetics.

And yet, two guys built mostly the same thing using simple heuristics.

So, why didn’t I just say that in the beginning and save you the extra fifteen minutes needed to learn about control theory and multi-arm pendulums?

I’m reminded of the scene in The Wizard of Oz, where the Scarecrow asks Glinda why she didn’t tell Dorothy that the ruby slippers would take her home when she arrived in Munchkinland in the first place. Instead, she went on this wild journey with flying monkeys, a dysfunctional Wizard, and a wicked witch that wouldn’t really be interesting until the sequel 75 years later. Why, Glinda, didn’t you just tell Dorothy the truth in the first place?

Glinda responds: “Because she wouldn’t have believed me. She had to learn it for herself.”

---

Part 2 of this series will cover how we got here: The story about how a young biomedical engineer who was looking to understand the progression of type 2 diabetes eventually found his models ending up in a system to control insulin delivery, a tool that it was not designed to do. And how the JDRF and Medtronic worked with the FDA to approve the first simulator that led to the first HCL. It was a big beautiful bet that was placed with good intentions, but contradicted the basic principles of biology and the medical system.