The claim
Redundancy. A10: “If a current epistemic state is genuinely sufficient for the entire family of future decisions under consideration, then preserving its history cannot improve ideal decision performance. That’s not an empirical conjecture. It follows from what sufficiency means.” CASE-082: under a correctly specified model with observations conditionally independent given Θ, p(Θ | E₁:t) is all a Bayesian updater needs.
Relativity. CASE-087, from A10 §2. Two readings of μ, X₁, X₂ ~ N(μ, σ²) independent with σ known. History A has (0, 10), history B (4, 6). Both have mean 5, and since σ is known both likelihoods for μ are proportional to exp(−(μ − 5)²/σ²). The mean is sufficient and the histories are indistinguishable for inference about μ. Now instrument 1 is found miscalibrated and X₁ must be dropped. History A leaves 10 and history B leaves 6, and the mean cannot say which. Introduce a two-component mixture and the mean cannot test it either, since the histories differ in spread. A18: “The earlier statistic has not somehow become mathematically incorrect. Its sufficiency was conditional on a specification that has changed.”
Why it is Active
Both halves follow from the definition of sufficiency and the stated constructions, which were checked for this entry. They hold whatever any line concludes (ADR-040).
What it bears on
- CLAIM-152 makes sufficiency relative to its target. This adds the model and the decision family, and change over time.
- CLAIM-294 is not a special case. A marginal is not sufficient even under the original model; here the state is sufficient and stops being so.
- A10 drew the consequence for history: “justificatory history is not extra information beyond a sufficient state.” CLAIM-295 takes it up.
- A10 §2 noted what CASE-087 needs is “a source-indexed data record”, not the researcher’s argument. That is an objection to reading this as a case for keeping justifications.