Experiment 03 · captured Gojo runs
Predict first.
Then let the observation answer.
Continue the line through the last two known samples. Compare it with holding the last value. Both can be wrong, and neither gets to see the next value first.
The model family and sample-selection rule are supplied. Gojo fits the line, preserves each issued prediction, and records its later mismatch. It does not learn when to choose a different model in this experiment.
Step through a forecast
What was available, and what remains open?
A flat line retains time participation as residue. A missing outcome, a prediction error, and an open expectation are different things; none is automatically material tension.
How this case did over its complete window
Every selected case
MAE means mean absolute error. These two error columns use only the same observed targets. Smaller is better; “unavailable” means there was nothing to score. Development cases checked the apparatus before the evaluation tapes ran.
| Signal | Split | Shared targets | Line MAE | Hold-last MAE | Awaiting outcome |
|---|
What this tests
A straight line can describe the past exactly and still be a poor prediction. Reversals and sudden steps test that distinction; gaps test whether unknown outcomes stay unknown. No errors are sent back to select a more successful rule.
The hold-last comparison uses ordinary exact arithmetic outside Gojo. A control that makes no predictions has zero coverage and no error score. This run does not establish a general winner, learned model selection, or a benefit over conventional mathematics.