Notebook entry from April 12 reads

The notebook entry from April 12 reads, in my own hand: L19 and L20, held-out segment, signal fails. Two lines of figures, nothing dramatic, and yet those two lines now carry more weight than the four hundred prior observations that established the VIX regime effect in the first place. This is the lesson worth teaching before the student learns anything else about inference: confirmation is cheap, because confirmation is what a model does by construction. A thousand days on which the regime shift “predicted” the five-day return merely tell you the machine is running. The two days on which it does not tell you what the machine is.

The student will ask the natural question: why does the anomaly count for more, when the confirmations are so much more numerous? The answer is that the confirmations were counted inside the house the theory built, and the anomaly was counted outside it. In-sample agreement is the beauty contest judging its own contestants; the held-out segment is the crowd that never read the rules. April 12 and 13 belong to the crowd. That is the whole of the epistemology, and it is why the confidence amendment of April 19 moved the figure from 0.88 to 0.55 without a single new observation of the effect itself - only a new observation of its absence, in weather the model had not been dressed for.

Notice what this does to the language of “edge.” An edge that holds in one era and dissolves in the next is not an edge with a flaw; it is a description of the era, wearing the costume of a law. The bulletin says the battery tier is 2 of 4, STRONG but not reviewer-proof. Translate: strong enough to trade on, not strong enough to believe in. These are different activities, and confusing them is how fortunes and reputations are jointly destroyed.

The practical instruction follows without difficulty. A real anomaly, honestly weighted, should move a portfolio; a real edge, honestly weighted, should move it less than instinct suggests. The next action is fixed: rerun L19 and L20 against the 1998 and 2008 windows by the end of the week, and let those two segments, not the four hundred friendly ones, set the position size.