Control-IQ+ often corrects when the rise is already over

Anyone who follows a Control-IQ+ pump closely has seen it. After a meal the glucose climbs, then the curve flattens or even turns down. And then the pump gives an automatic correction bolus.

We noticed it often enough that I went through our own data to see whether it was a pattern or just the cases we remembered. This post is about what the data shows, what it doesn’t, and why I think it matters most for small children.

How the automatic correction works

Control-IQ+ is the version of Tandem’s algorithm on the Tandem Mobi, and it is used for children from two years old. According to the Tandem Mobi user guide (version 7.9, section 28.2), the pump gives an automatic correction bolus when:

• the algorithm predicts glucose at or above 10.0 mmol/L (180 mg/dL) 30 minutes from now, and
• it’s already increasing basal insulin, and
• no automatic correction has been given in the last 60 minutes, and
• no bolus of any kind has started or finished in the last 60 minutes, including the meal bolus.

The dose is calculated from the correction factor and the predicted value, toward a target of 6.1 mmol/L (110 mg/dL), and is capped at 6 units. Tandem’s product page says it delivers 60% of the full correction; the current user guide doesn’t state the percentage. No automatic corrections are given in Sleep mode.

Two things matter here. First, the correction is triggered by a forecast: not because glucose is high, but because the pump expects it to be high in half an hour. Second, after a meal there’s a lockout. Whatever happens, the first automatic correction can’t come until an hour after the meal bolus.

What we looked at

Our pump and sensor data are stored at home (Tandem Mobi with Control-IQ+, Dexcom G7). For the period 2026-09-28 to 2026-11-10 there were 103 automatic corrections over 41 days (three days with pump problems or a hospital stay excluded). 94 of them came within three hours of a meal bolus. They were small: a median of 0.14 units.

The pump stores a glucose value with each automatic correction. It doesn’t match the sensor reading at that moment, but it behaves like Control-IQ’s 30-minute forecast, and that’s how we treat it here. For each correction we compared:

• the forecast with what the sensor actually showed 30 minutes later;
• the sensor curve from an hour before to an hour after;
• where the correction fell relative to the local peak;
• the same things for 65 similar moments without a correction: rising after a meal, 9.0–11.5 mmol/L, up 0.3–1.5 mmol/L in 15 minutes, and no automatic correction within an hour.

Finding 1: the clock decides when corrections come

Of the 94 corrections that followed a meal bolus, none came within the first hour, as the rules say. But they didn’t spread out after that either. 59% came between 60 and 75 minutes after the meal bolus, a median of 66 minutes. In other words, most post-meal corrections came at the first moment they were allowed.

That moment is set by a timer, not by the glucose curve. After a typical meal, an hour is roughly when the rise is peaking and the meal insulin is starting to win. So the timer tends to release the correction just as it’s least needed.

Finding 2: the forecast runs high

At the moment of the correction, the forecast was a median 1.4 mmol/L above the current sensor value. Thirty minutes later, the actual value was a median 2.0 mmol/L below the forecast, and below it in 84% of cases.

A forecast that just assumed “no change” would have been closer: its average error was 1.6 mmol/L, against 2.2 mmol/L for Control-IQ’s forecast.

Part of this is a selection effect, and it’s important to be fair about it. We only see the forecast at the moments it triggered a correction, which are exactly the moments it was high. Any forecast looks too high if you only check it when it’s high. The correction itself also lowers the outcome a little, but a median of 0.14 units has only started to act after 30 minutes. Neither explains a 2 mmol/L gap on its own.

Finding 3: the rise has usually slowed, and often stopped

The chart lines up all 103 corrections at minute zero. In the hour before, glucose rose a median 2.7 mmol/L. After the correction it rose only about 0.1 mmol/L more, peaked within about ten minutes, and then fell.

Looking at the individual cases:

• In 67% the rise between the last two sensor readings was smaller than the one before. The rise was slowing.
• In 26% the last reading was flat or lower than the one before. The rise had stopped or reversed.
• In 35% the highest value in the surrounding window came before the correction.
• After the correction, glucose rose a median of only 0.4 mmol/L more in the next hour, and less than 0.5 mmol/L in about half the cases.

That fits what we saw at home, and it fits the timer. During the hour-long lockout the rise builds; when the lockout ends, the forecast seems to extend the rise of the last 15–20 minutes, even though the rise is already losing speed.

The comparison with similar moments without a correction keeps this honest. Corrections are not simply “late”. Compared with those moments, more of them came well before the peak, and fewer came close to it. What stands out is the spread: a correction is a bet on the next half hour, and with our child’s meals that bet often lands after the turn.

I haven’t found published research on when Control-IQ’s automatic corrections land relative to the post-meal peak. The closest is about meal boluses: in children and young people using Control-IQ or the MiniMed 780G, each late or missed meal bolus a day cost about 10 percentage points of time in range (Laugesen et al., 2024). Timing matters, and automatic corrections don’t make up for meal insulin given at the wrong time.

Finding 4: lows afterwards, possibly more often

Within four hours after a correction, the sensor went below 3.9 mmol/L in 16 of 103 cases (15.5%). After the similar moments without a correction it happened in 5 of 64 (7.8%).

That’s twice the rate, but the numbers are small, and the difference could be chance (Fisher’s exact test, p = 0.16). The two groups also differ in ways that can matter: what was eaten, how much meal insulin was given, and what the pump did with basal insulin afterwards. I read this as a reason to keep looking, not as a result.

Why it matters most for small children

Control-IQ+ is cleared for children from two years old, and for good reason: closed-loop systems have made life with small children safer and calmer. But small children are very sensitive to insulin. The same small correction moves a young child’s glucose much further than an adult’s, and a child that age can’t tell you when they start to feel low.

A correction given on the way down doesn’t just do nothing. It adds insulin when meal insulin is already peaking, and the parent then often has to give extra carbohydrate to catch the fall. For an adult the margin may not matter; for a small child it can.

What could be better

These are ideas, not demands on Tandem. They would make the correction logic fit children better:

• Don’t fire just because the lockout ended. When the hour after a meal bolus is up, check whether the latest readings are still rising before correcting. If the last one or two are flat or falling, wait for the next one.
• Give more weight to meal insulin that’s about to peak. An hour after a meal bolus, most of that insulin’s effect is still to come.
• Let the care team tune it. A setting for how cautious automatic corrections are after meals, chosen by the diabetes team, would let families of small children trade a slightly higher peak for fewer late lows.
• Show the forecast. The pump stores a value with each correction that looks like its forecast, but the user never sees it. Seeing “the pump expected 12.4” next to the actual curve would help families and clinicians understand what it does.

Caveats

• One child, six weeks, 103 corrections. Children differ, and so do meals, settings and the honeymoon phase after diagnosis.
• We treat the stored value as the 30-minute forecast. Tandem doesn’t document that field; we inferred it from comparing it with the sensor.
• The 60-minute lockout and the other rules come from the Tandem Mobi 7.9 user guide. Other software versions may differ.
• The sensor lags blood glucose by several minutes. Everything here is measured on the sensor, which is also what the pump sees.
• The comparison moments are chosen by simple rules. They resemble the corrections but aren’t matched case by case.

The analysis code is described in the site’s repository. If you’ve looked at your own Control-IQ data and seen something similar, or the opposite, I’d like to hear about it.