Whether to adopt an evaluation-timing option in place of the single evaluation of the BRIA design, and which.

queued · open · DECISIONS.md line 101

Whether to adopt an evaluation-timing option in place of the single evaluation of the BRIA design, and which. The BRIA-corrigibility round (Part F) compares a deterministic growing lookahead (a known time leaves a free delay at every day), a mixture over evaluation times with per-time gates (settled in pieces) and a random time T ~ ρ_n (one settlement, a hidden draw); its first follow-up recommended the mixture on the escrow released as weight settles; its second follow-up shows the expected escrow is the same for the mixture and the random time at every time (BRIAFollowup2.expected_escrow_eq) and adds the hybrid — k hidden draws per decision scored by their average, the same expected weight on every delay for every k, score variance σ²/k (average_mean, average_variance) — and the criterion the earlier passes missed, her evaluation load: one evaluation per decision (options 1 and 3), one per time in the support or a truncation cost bounded by the tail weight times D (option 2, truncation_cost), k per decision (option 4). No option is recommended; Part A's single evaluation stays the default. Turns on: her evaluation load against the variance of the score and of the escrow (zero for the mixture, σ² for one hidden draw, σ²/k for k), and whether a hidden draw is acceptable; secondarily the weights' tail against the agent's delay reach.

Touches

Nothing named.

None.