A recorded value and an explanation of that value are different achievements. Inference connects them through a model, assumptions about uncertainty, and comparisons that may leave more than one explanation possible.
Two observations, two explanations
Imagine a detector recording a decaying signal at equally spaced times. Its first two exact readings are z(0) = 10 and z(1) = 6, in arbitrary units. Consider the response family z(t) = A exp(−kt) + B, where A is the initial signal amplitude, k is its decay rate and B is a constant background.
One model sets A = 10, B = 0 and k = ln(5/3). Another sets A = 8, B = 2 and k = ln 2. Both reproduce the two observations exactly. At t = 2, however, the first predicts 3.6 while the second predicts 4.0.
The recorded pair is not faulty because it permits both models. It simply does not identify all three parameters. A background measurement, a third time point or a justified restriction on the model family can add information. Assuming B = 0 without checking it would convert an unresolved question into an invisible premise.
Ask how well each model predicts new data
Suppose the parameters are fixed from earlier work and a reserved measurement at t = 2 returns 4.0. Assume its error is independent, Gaussian, and has a known standard deviation of 0.2. Under these assumptions the likelihood ratio favouring the second model over the first is exp[(4.0 − 3.6)²/(2 × 0.2²)] = exp(2), approximately 7.39.
This ratio compares the probability densities assigned to the same observation by two specified models. It is not a 7.39-to-one posterior odds statement unless prior odds and the relevant model alternatives have also been supplied. If the observed value were 3.8 instead, these two fixed models would assign equal densities.
The calculation also depends on the uncertainty model. An unrecognized background drift or uncertainty in the earlier parameter estimates would change the comparison. Those qualifications do not make inference useless; they identify what must be measured or propagated for the comparison to answer the intended question.
Keep the levels of the claim visible
A careful account can distinguish the stored detector numbers, their conversion into a calibrated signal, an estimated decay rate and an interpretation of the underlying process. A correction at one level need not invalidate every earlier level. Discovering an overlooked background may alter the inferred rate while leaving the archived numbers intact.
Likewise, two candidate mechanisms can generate the same response function. Establishing exponential decay would then constrain their observable behaviour without selecting a unique hidden realization. The previous collection's source-model example gives an explicit case of hidden-coordinate ambiguity.
It is useful to record which conclusions survive alternative processing choices and which depend on a narrow modelling decision. A stable inference has earned its stability through those comparisons; it is not stable merely because the final number is printed to many digits.
Let disagreement suggest the next observation
The two decay models disagree at later times, so a new time point is informative. But an extremely late reading might fall below the detector's useful range. The experiment should compare predicted separation with expected uncertainty and the available dynamic range before selecting the test.
A direct measurement of the background is a different option. It targets the variable responsible for the ambiguity. A stronger research design may combine both approaches: measure the background independently, then reserve later decay readings to test whether it remains constant.
This is a practical reason to publish alternatives alongside a preferred interpretation. Readers can see which observation would distinguish them and which additional data would merely repeat the same unresolved comparison.
From records to an SMM/DTT inference
A proposed inner-time model should supply a route to observable probabilities or response curves. The expanded book's tests with explicit alternatives explains why agreeing with a formula already used to construct a model is weaker than predicting a reserved response.
The quantitative interference record offers another useful pattern: specify what coherence remains accessible and calculate the probabilities of later tests. A measured probability curve then supports statements about that readout model. Establishing the unique physical reality of a proposed inner history requires a further distinguishing result.
Sources and further reading
- G. Cowan, Statistics, in the Particle Data Group's 2024 Review of Particle Physics, especially likelihood and parameter estimation. The decay comparison is independently worked here.
- From a signal to knowledge and laws and initial conditions.

