nucleation

LIT-1880: When Can Safe Controllers Adapt? Information before Commitment

1 min read

Properties5
StatusActive — read in full 2026-10-31 (NOTE-1466), with the main proofs; worth reading for the precise sense in which evidence must precede commitment: the observation a committing action produces cannot justify it, and bounded information before commitment leaves a fixed fraction of the oracle gap to every uniformly safe controller (THEORY-1254). It also proves the positive case, safe recovery from a repeatable informative experiment, and shows certainty equivalence failing at an active constraint (THEORY-1256).
Version2
Tagscontrol-theory, mathematics, learning-theory
Cited as “paper” byNOTE-1466 — Saligrama — information before commitment in safe adaptive control
Cited as “source” byTHEORY-1254 — A controller that must stay safe under every model it has not ruled out cannot use the evidence its committing action produces: if the information it can gather before its first action that forecloses a safe continuation under an alternative is bounded by β, every such controller forfeits at least about ½e^{−β} of the value reachable only by committing, so a linear oracle gap gives linear regret, and acting on an exclusion at risk α needs about log(1/(2α)) nats gathered beforehand, THEORY-1256 — When a conservative experiment is safe under every plausible model, can be repeated from a common state and keeps informing, a controller that acts on the lower end of an anytime confidence sequence stays safe with probability 1 − δ and approaches the safe oracle, at O(√(T log T)) regret in a braking example; a certainty-equivalent controller in the same example is unsafe with probability ½ in every episode however much data it has

Graph View

Saligrama, V. (2026). When can safe controllers adapt? Information before commitment. arXiv:2607.16895.

Read in full in NOTE-1466.

Why this is here

A20 §11 used it for the condition that inquiry needs information gathered before a correction is committed to: a discrepancy that cannot be distinguished in time cannot be corrected, whatever the update rule. It bears on CLAIM-312 and TERM-085’s deadline. The reading bears A20’s use out and makes it exact. The floor on the evidence needed before the deadline is about log(1/(2α)) nats, and evidence that the corrective act itself would generate does not count (THEORY-1254). Changing the estimator cannot help; changing the experiment, the sensor (a persistent imprint rather than a transient signal) or the degree of specialization can. The “guarantee or exclude” choice, staying safe under every live alternative at a price fixed by information, bears on CLAIM-511 and QUESTION-285.

Backlinks