The claim
A5: “the information required for future epistemic revision may differ from the information required to reconstruct present conclusions”. A8 turned the intuition into a construction (CASE-093). With H, C binary and P(C = 1) = 0.5 in both histories:
| history A | history B | |
|---|---|---|
| P(H = 1 | C = 1) | 1.0 | 0.8 |
| P(H = 1 | C = 0) | 0.8 | 1.0 |
| P(H = 1) | 0.9 | 0.9 |
Both marginals are 0.5·1.0 + 0.5·0.8 = 0.9. On learning C = 1, history A goes to 1.0 and history B to 0.8. A memory that keeps P(H) maps both histories to the same state and cannot make both updates. That is a proof by counterexample of the non-implication, and the arithmetic was checked for this entry.
Why it is Active
It is an existence result, true by construction (ADR-040
puts it in formal-apparatus as a second line). A8 drew the line where
this entry does: “The construction proves that preserving the current
marginal belief about H does not generally preserve everything required
for subsequent Bayesian updating. … The broader hypothesis about real
research-agent memory should remain Proposed.” The thesis the
nucleating-argument line cares about, that some inquiries in fact lose
revision competence this way, is CLAIM-295 and stays Proposed.
What it does not say
- Not that a current state can never suffice. A5’s qualification: “A complete Bayesian posterior over all relevant latent variables can be sufficient for subsequent updating.” CLAIM-300 and CASE-082 hold that half.
- Not that keeping justifications is the remedy. Keeping the joint over (H, C) is enough here, with no history at all.
- Not that a counterexample to one representation is evidence for another. A18: “Showing that a conclusion-only summary fails on a revision task does not establish that a justification graph succeeds.”
The same point in belief-revision logic (note of 2026-10-15)
Two readings filed the same day give the full-belief form of this claim. Hansson’s survey (LIT-1072, read in NOTE-866): bases {p, q} and {p, p ↔ q} have the same closure, yet after revision by ¬p one agent believes q and the other ¬q. The bases are “statically equivalent” and not “dynamically equivalent”. In AGM (LIT-1050, read in NOTE-874) the closed theory alone does not determine a revision; a selection function over its remainders must be added. Both match this claim’s caveat: a richer current state, a base or an ordering, suffices for one revision. Neither is a record of what a belief depended on, so neither supports reading this claim as an argument for keeping justifications (CLAIM-295’s question, not this one’s).
Note of 2026-10-21: the provenance misreading recurs
A2 and A5 of the exchange of 2026-10-20 read this claim as a case for provenance (“The value of evidence provenance depends on …”, and “No provenance” as an ablation that loses revision). This claim’s own “What it does not say” is against that reading: “Keeping the joint over (H, C) is enough here, with no history at all.” The misreading recurs. A1 of the same exchange did keep the warning.