One success probability does not fix an entire protocol
A repeated preparation or completion attempt is often summarized by one success probability. That number determines a geometric retry law only when the state entering each attempt has the required independence and stationarity. A failed attempt may instead retain a hidden variable, alter the apparatus, or select a more difficult subensemble.
The book's future-test criterion makes this issue precise: two present outputs should not be identified if later admitted experiments can distinguish them. The published reconstruction likewise asks whether a retained record supports its own predictive evolution. [B05] [P01]
The laboratory compares two ordinary classical models. The independent model uses p̄ on every attempt. The mixture model first chooses a high-success type with probability w or a low-success type with probability 1−w, then keeps that type fixed for the run. Conditional on the type, the attempts are independent with probabilities p_H and p_L. Both models share p̄=wp_H+(1−w)p_L on the first attempt.
Keep the first attempt fixed and inspect later attempts
Change the two success probabilities, their mixture weight, and the number of attempts n. The curves show the chance of at least one success by that attempt. They are exact expectations under the two stated models, not simulated trial histories.
Interactive calculation
Play scans Number of attempts n. Step advances it. No animation starts automatically.
Enable JavaScript for adjustable values. The static figure and worked calculation remain available.
With p_H=0.8, p_L=0.2, and w=0.5, both first attempts succeed with probability 0.5. After two attempts, independent retries succeed with probability 0.75, while the persistent mixture succeeds with probability 0.66. The difference comes from the hidden type retained across failures.
Compute survival before computing success
In either case, the complement of success is failure on every attempt. In the mixture, that survival probability must be averaged after the powers are taken because the type is fixed for the run. Drawing a new type independently on every attempt would instead recover the first formula.
For n≥2, q↦qⁿ is convex on [0,1]. Jensen's inequality therefore gives P_mix≤P_ind at the same mean one-attempt probability. Equality holds when the two types coincide, when only one type has weight, and at n=1. Persistent heterogeneity slows completion in this precise comparison; it is not an arbitrary numerical accident.
After n failures, wₙ is the conditional high-type weight whenever the denominator is nonzero. The next-attempt success probability becomes wₙp_H+(1−wₙ)p_L. When p_H>p_L, repeated failures enrich the low-success type. The retained failure count therefore carries predictive information about the unseen preparation.
If both type probabilities are positive, eventual success approaches one, but the approach can be much slower than the mean-probability model suggests. If p_L=0 and p_H>0, the limiting success is w: a permanently impossible subensemble never completes. A long finite record can confuse a very small p_L with zero, so an observation window must be part of any empirical claim.
Another diagnostic is the covariance of two potential success indicators: Cov(X₁,X₂)=w(1−w)(p_H−p_L)². It is nonnegative because both attempts share the same latent type. To measure that covariance directly, the protocol must define both attempts even when the first succeeds; stopping at success changes which records are available.
Test memory without assigning it a unique ontology
A protocol should retain attempt number, failures, resets, and preparation information, rather than reporting only successful runs. Compare a genuine reset between attempts with a continuation that leaves the hidden preparation intact. Fit a model on one retry window and reserve a longer or differently reset window for prediction.
Different mechanisms can produce the same survival curve. Classical heterogeneity, detector recovery, environmental correlations, and a specified quantum instrument are distinct possibilities. A two-component fit is evidence for the usefulness of that response model only after its alternatives and uncertainty are examined. It does not uniquely establish a second temporal ontology.
The book makes the analogous point for reduced memory kernels: matching one response does not distinguish every hidden realization. Additional preparations or measurements must separate competing mechanisms. [X04] [X05]
Continue to coupling identification to see a parameter ambiguity that persists even when an entire noiseless response curve is known.

