[ps2]

guardAndRunway (changed)

• Volatility band for the low guard only. The band now lowers only the forecast the low guard checks. A banded point below the suspend threshold suspends, and the banded minimum is the minimum the guard sees. The correction is sized from the unbanded forecast, and that is also the forecast returned in predictedGlucose. The displayed forecast no longer carries the band's dip and taper. GuardAndRunwayOutput.guarded(_:) still gives the guard's view.
• Runway fades with carbs. The runway reserve scales down linearly from no carbs on board to none at 20 g, instead of closing at any carbs.

In LoopEval's generator (166 people × 2 trainings) the change is neutral against the current version: TIR +0.04, time below 70 +0.02, time below 54 +0.01.

slowRetrospectiveCorrection (new)

A day-scale estimate of how much more or less insulin the person needs than their settings say.

• Level: the exponentially weighted mean, with a 24 h time constant, of the glucose change the model did not explain, per 5 minutes. That change is insulin counteraction less carb effect, at the scheduled ISF. The mean covers the glucose history passed in, up to 72 h, starting from zero.
• Multiplier: ISF for the whole run is multiplied by 1 − 0.05 × level, kept within 0.7–1.3.
• Output: AlgorithmOutput.slowRetrospectiveCorrection reports the level and the multiplier.

In the generator, on top of this PR's guardAndRunway, it gives TIR +1.5, time below 70 +0.23 and time below 54 +0.08. That is +0.46 TIR [+0.13, +0.74] more than raising insulin needs reaches at the same time below 70. In a four-day replay of one person running Guard + Runway, it gave TIR +6 at the same insulin and no time below 70.

Implementation notes

• History. The experiment keeps no state; it is recomputed from the inputs on every run, so it needs a long history:

  • glucose: 72 h;
  • carb entries: from 72 h plus the longest absorption time;
  • doses: from 72 h plus the insulin activity duration.

  With 12 h of history the level reaches only about 40 % of its working value.

• Carbs for the whole history. A meal whose entry is dropped while its absorption is still inside the glucose history reads as unexplained rise and pushes the level up. In a replay that moved the multiplier by about 0.04 for a day.

• Timeline coverage. The carb ratio and ISF timelines must start at or before the oldest carb entry. run now checks this and fails with the new AlgorithmError.carbRatioOrSensitivityStartsAfterCarbEntry; before, CarbStatus construction hit preconditionFailure. The check applies with or without experiments, and Loop should handle the new case.

• Cost. Two forecasts per run when the multiplier is not 1, each over the full history: 23 ms median, 207 ms worst, on a Mac (release build).

• Agreement with LoopEval. Each cycle of a four-day closed-loop replay was run through this package: multiplier difference median 0.0007, 98.5 % of doses identical, worst difference 0.05 U.