The case
A6’s first version: two groups both put P(H) = 0.9, one from a direct experiment, one from a simulation that relies on an untested calibration C. Then C is invalidated. A8 made it a construction. H, C ∈ {0, 1}, P(C = 1) = 0.5.
| 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 |
On learning C = 1: 1.0 in A, 0.8 in B.
What it can show
That a memory keeping P(H) cannot make both updates. Its facts are its definition.
What it cannot show
A6: “That an accountability graph is the best way to preserve the distinction, or that this occurs when the agent instead retains a complete sufficient posterior”. Keeping the joint over (H, C) is enough. CASE-082 is the matched case where the state suffices; A6: “Together, they force us to articulate what is meant by current belief state before making claims about its limitations.”