Smarter Autopilot: Why We Now Cross-Check Two Signals Before Adjusting Your Calories

The old model, and what was wrong with it

When Autopilot first shipped, it made decisions off of a single signal: the trend slope across your comparison window. Roll up the last 5 days, fit a line, look at where it's pointing. If the trend was flat — stagnation. If down — losing. If up — gaining.

Clean idea. Worked most of the time. But there was a failure mode.

Weight is noisy. Water, salt, sleep, the bathroom you visited (or didn't), the day of your cycle — all of it moves the scale by 0.5 to 1.5 kg with no relationship to actual fat mass. When that noise lined up against the slope just right, Autopilot would see a trend that wasn't real, and recommend a calorie change it didn't need to.

Worst case: a Cut phase, two weeks of solid adherence, one bad sleep night spikes a single weigh-in, the trend turns "down" → Autopilot suggests another -200 kcal. You're already in a deficit. Now it's deeper, for no actual reason.

That's a confidence killer. So we fixed it.

The new model: two signals, both must agree

Autopilot now looks at two things before suggesting anything:

1. The trend slope — same as before. Fit a line across the comparison window. Where is it pointing?

2. The endpoint delta — what is your actual weight today vs. your actual weight at the start of the window? A raw, single-number comparison.

The system only acts when both signals point the same direction. If the trend says "down" but your actual first-to-latest weight is flat, no change. If the endpoint moved down but the trend is messy, no change. Both must agree.

When they do agree, you can trust it. The trend confirms the direction, and the endpoint confirms it's real — not an artifact of where the line happened to land.

Why this isn't just "more conservative"

It would be easy to dismiss this as "the app is just doing less now, of course it's wrong less often." That's not quite right. The change is directional patience, not blanket hesitation.

When both signals agree clearly — a clean cut, a steady gain, a flat plateau — Autopilot acts exactly as fast as before. The full ±200 kcal step. The same comparison window cadence.

What changed is the edge cases. Specifically: the cases where the trend slope was the only thing arguing for a change, and a single weigh-in had outsized influence on it. Those used to trigger an adjustment. Now they don't.

In testing, this cut spurious recommendations by a meaningful margin during noisy weeks (travel, cycle, illness) without slowing down recommendations during clean ones.

The mental model

The old way: "the trend is what it is."

The new way: "the trend has to mean something."

A line through 5 noisy points can point in any direction. A line that points down and ends meaningfully below where it started — that's a signal. That's what Autopilot waits for now before nudging your calories.

You'll see fewer Briefs that feel premature. The ones you do get will feel earned.

If you haven't tried Autopilot yet — Phase 1 (Find Peak Maintenance) is free, and now it's smarter than ever about when to act. Read the full Autopilot guide →