The recovery domain of the updateless extension.
queued · open · DECISIONS.md line 212
- The recovery domain of the updateless extension. An updateless candidate cannot be a BRIA at unit level on the environments P1′ covers, so its BRIA-recovery obligation needs a domain on which the local and policy-level demands coincide. The round rejects two candidates with witnesses — "rewards independent of the agent's choices" (too small) and "rewards a function of the history and the current choice" (too large: per-round counterfactual mugging with a frequency predictor is in it and P1′ forces refusal there while paying is policy-optimal) — and defines none. Turns on: what property of the reward process, stated without reference to the candidate, makes myopic optimality and policy optimality agree; until it exists no updateless candidate can be certified as having recovered BRIA where it should (
projects/decision-theory/rounds/2026-09-29-li-bria-synthesis-spec/PROBLEM_STATEMENT.md§2, M3 domain 4).
Touches
Nothing named.
Nodes pointing at it
None.