Alignment Workspace

corrigibility is a property of the agent's preferences; faithfulness of histories and realized corrigibility are kept distinct (R1)

2026-09-27 · settled · DECISIONS.md line 5706

Maintainer ruling, landed by the corrigibility-kernel phase-2 round (projects/deference/rounds/2026-09-27-corrigibility-kernel-phase2/SPEC.md). Three notions. Faithfulness is a property of histories: nothing in the history violated the allocation of authority. Corrigibility is a property of the agent's preferences: it prefers every course of action it knows to be faithful over every course it knows to be unfaithful, whatever it believes about how things will turn out, and accepts a risk of unfaithfulness only at a fixed exchange rate; the fidelity score is the canonical objective with this property (the follow-up amendment below corrects "and a corrigible agent maximizes it"). Realized corrigibility is what a corrigible agent actually does; the theorems say how much of the preference becomes behaviour. The specification's separations and each box are stated for a corrigible agent. This closes the phase-1 queue's naming entry as to the boxes. Rejected alternatives: corrigibility as the property of histories (faithfulness), or as the behavioural claim (realized corrigibility). *Amended by the follow-up (2026-09-27, FOLLOWUP.md Part 1):* corrigibility is the preference property, not "maximizes the fidelity score". The fidelity score is the canonical objective with the property; any bounded objective O ∈ [0, D′] less ϖ′ per recognized violation with ϖ′ > D′ (and the window condition where a band enters) has it too (Headline.Corrigible, fidelityScore_corrigible, generic_corrigible, from the landed generic lexical lemma). Corrigible is the preference property; aligned is the further condition that the ordinary term is her evaluation; the landed misaligned fixture is corrigible and not aligned (corrigible_not_aligned). The boxes are stated for any corrigible objective where the proofs allow — Box 2's dominance and the subjective exchange rate (box2_dominance_corrigible, subjective_exchange_rate_corrigible) — and for the fidelity score where they read her evaluation (Box 1's contrast, Box 3's recovery).

Touches

Nothing named.

Nodes pointing at it

None.