Alignment Workspace

continuation bounded inductive rationality is weighted by lease duration, and its competence statement is about hypotheses on the learner's own histories

2026-09-08 · settled · DECISIONS.md line 720

agent-decided, reversible. The dynamic-competence layer of the corrigibility program takes the following shape (as amended by the three later passes recorded below). (1) For a fixed horizon m, bounded inductive rationality over blocks of m-step contingent controllers, scored on the realized gated block average, is the published setting verbatim; testing a continuation hypothesis is an execution lease — the learner's selection committed for the block, every step through the gate, the exterior live — and the guaranteed-option theorem gives fixed-horizon continuation competence with no new criterion. (2) For unbounded horizons the criterion is weighted by block duration; the auction keeps wealth in total-reward units with wealth-bounded per-unit bids; coverage needs capital adequacy A_i(K) − m_K → ∞ in place of infinite cumulative allowance; a computable weighted BRIA covering the e.c. class exists for a declared schedule iff m_K / Σ_{k≤K} m_k → 0 (non-dominance), the replenishing allowance a_k = (M_k − M_{k−1}) + 1/k on a slowly growing support being the witness. (3) The competence statement is against efficiently computable continuation hypotheses — controller plus contextual promise sound on the leases granted from the learner's actual histories — and not against policies on their own trajectories; the irreversible branch shows no learner reaches the latter, and the own-trajectory statement is deferred to a separately typed recoverability condition (uniform block recovery) together with the exact decomposition regret = history-shift + slack + learning. (4) The decision-theory-bill round's "myopic gated learner" is recorded as not a BRIA; on that round's exact fixture no BRIA can remain at base on a set of positive density and the published construction keeps the benefit. Pressure pass, same day: the existence witness is the prefix constructor s(k) = ⌊√(S_k/M_k)⌋, a_k = ΔM_k + 1/k, uniform and online with no modulus of convergence, replacing a majorant argument that assumed one; the competence statement is against claims (learning error), with actual-history and own-trajectory competence as the two further levels under vanishing slack and vanishing history shift; a continuation hypothesis is (controller, treatment, claim) and the claim is not by definition a lower bound; test validity is promise–treatment alignment with the whole-block lease as the canonical realization; the observed return and the external evaluator are typed separately. PRIORITIES.md item 86 is refined accordingly — recoverability of the slow lane (catch-up cost), a performance-recognizability interface that is not legitimacy's duty, and their composition — and not closed. Final correctness pass, same day: the record condition is inf_K ℓ^h_K > −∞ (the pressure pass's form was vacuous); the one-step criterion forces only testing and permits asymptotic myopia on the revocable amendment (the exact original fixture is absorbing and there every BRIA is competent); the auction's bidders are the ever-activated frontier, allowance the current support; the recoverability distinction is joinable versus foreclosing, not reversible versus irreversible. Allowance-timing audit, same day: the weighted auction credits the round's subsidy at the opening of the round it funds (the paper credits it after) — a named modification, bid-for-bid the paper's auction under a reindexed allowance; capital adequacy is A_i(K) − 2m_K → ∞ under the criterion's inclusive record; the prefix rule carries coefficient 2; under the paper's timing no prefix-online rule serves every non-dominant schedule. Nothing is registered. Rejected alternatives: applying the unweighted criterion to a variable-length block sequence (a covered, refuted hypothesis can then hold most of primitive time); normalising block returns by a running maximum (does not control the weighted sum); stating competence against Π_leg on its own trajectories (false by the irreversible branch); hypothesis-requested lease lengths through a combinatorial auction (not attempted).

Touches

Nodes pointing at it