Module predict 

Predicting the next reflection.

F(M(t)) -> M_hat(t+1)

Why the baselines matter more than the model

Predicting a mirror snapshot is easy to do impressively badly. Most cells barely move between consecutive reflections, so copying the last value scores extremely well on almost every channel. Any model that does not clearly beat that has learned nothing, however small its absolute error.

So every prediction here is scored as skill relative to the best naive baseline, and the baselines are chosen to be genuinely hard to beat:

• Persistence. x(t+1) = x(t). Very strong for slow-moving readings.
• Drift. x(t+1) = x(t) + mean recent delta. Near-perfect for accumulators, which otherwise make a model look brilliant for free.

The drift baseline is the reason the accumulator question in corescout_represent::discover is not academic: an observer that has not worked out which variables accumulate cannot construct this baseline, and will mistake trivially predictable counters for evidence that it understands the machine.

The model

A per-cell autoregressive step fitted by least squares:

delta(t+1) = a * delta(t) + b

Deliberately the simplest thing that can express momentum. The point of this milestone is not a good predictor; it is to find out whether the reflection contains enough signal for any predictor to beat copying the last value. A weak model that clearly beats the baseline is a stronger result than a complicated one whose advantage cannot be attributed.

Structs

CellScore

How well something predicted one cell.

PredictionReport

Prediction performance over a whole machine.

Functions

evaluate

Fit and evaluate a one-step predictor over remembered series.