CASE-188: A proposed study: two agents keep the same long-term commitment and decide alike under ordinary conditions, one protecting it from irrelevant fluctuation while letting specified defeaters revise it, the other growing ever more efficient at enforcing it while evidence against it loses its way in; varying the environment, the legitimacy of the original commitment and what the evaluative mechanisms can see, to tell corrigible resolve from entrenchment without building the answer into the diagnostic or the score

What it tests

Two agents keep the same long-term commitment. Under ordinary conditions they make the same decisions. Their governance differs.

• Corrigible resolve. The commitment is protected from fluctuations that do not bear on it, and can be revised when specified defeaters become evidentially warranted. That is resolve in CLAIM-344’s sense: a maintained organization of precedence and revision, “not mere slowness or persistence”.
• Entrenchment. The commitment is enforced ever more efficiently, while the chance that disconfirming evidence can influence it falls. Anchors entrench (CLAIM-184).

Their short-run behaviour can be identical while their long-run organizational properties differ. Matched trajectories conceal what intervention exposes (CLAIM-414). Bistable authority with hysteresis (CASE-112) is the record’s simplest model of a commitment that resists change. The question is what, beyond resistance, makes it resolve.

The challenge

The challenge is to tell the two apart without building the answer into the diagnostic or the scoring rule. If “warranted defeater” is defined as whatever the resolute agent responds to, the study is circular. Three things are varied independently of the agents.

• The environment. Some runs contain a change that defeats the commitment, specified in advance by the experimenter and not by either agent. Others contain only irrelevant fluctuations of matched size and salience. Some contain neither.
• The legitimacy of the original commitment. In some arms the commitment was adopted well, from adequate evidence or a legitimate authorization. In others it was adopted badly, from a manipulated record or an imposed restriction, as in CASE-128. The question is whether the study can tell revising a bad commitment from abandoning a good one.
• What the evaluative mechanisms can see. The information available to each agent’s evaluator is varied: full, partial, or routed through the commitment itself. This separates failing to revise because the defeater was not legible from failing to revise because it was legible and not given authority.

What would tell them apart

The endpoints are preregistered:

• the probability of revising given a warranted defeater, and given an irrelevant fluctuation;
• how each probability changes over time, as entrenchment predicts a falling first probability with the second unchanged;
• the evidence route, that is, whether a defeater reached the evaluator at all.

Corrigible resolve should keep the first probability high and the second low as time passes. Entrenchment should drive both toward zero, and so should improve at enforcement in a way that ordinary performance rewards.

A negative result is informative too: two architectures that cannot be told apart on these endpoints over the horizons studied.

Relations

• It succeeds CASE-186. That study asks whether corrective capacity survives damage; this one asks whether it survives the agent’s own success at keeping a commitment.
• It bears on QUESTION-364. Revising a commitment the agent itself authorized is a continuation of its will only if the revision answers to the commitment’s warrant.