The Behaviorist Lab

In most review interfaces, approving costs one click and objecting costs a written justification, a context switch, and a social cost. If that asymmetry alone drives approval rates, independent of the risk of the underlying decision, then a large share of human-in-the-loop oversight is decorative, and the fix is interface design rather than training or policy.

A controlled study varying only the cost of the escalation path while holding decision risk constant, measuring approval rate and review time. This is cheap to run and has not been run.

The scale treats reversibility, consequence magnitude, confidence, precedent, and regulatory exposure as five distinct inputs. If reversibility and consequence magnitude are in practice one factor, the instrument is measuring four things and claiming five, and the weights are wrong in a way that matters.

Factor analysis over a large corpus of scored decision points from more than one organisation. This requires data the lab does not yet have enough of.

The Checkpoint Scale's weights are theory-driven. They were set from the behavioural literature, not fitted to outcomes. Until they are calibrated against what actually went wrong, the instrument is a structured opinion, and it should be described that way.

Decision points scored before the fact, paired with their real outcomes, across enough cases to fit and then validate weights out of sample.

This is the load-bearing claim under most of the practice. The argument is that pressure takes deliberate reasoning offline first, and that an AI assistant arriving at that moment gets used to confirm an existing prior rather than widen the decision. The narrowing-under-stress half is well established. The AI half is not, and it is asserted far more often than it is tested.

An experiment varying time pressure while holding the task constant, giving every participant the same AI assistant, and measuring whether they seek confirming or disconfirming information. Preregistered, so a null result counts.

Nobody has published what production agentic systems actually decide without a human, at what risk profile, at what volume. Every current claim about where autonomy sits, including ours, rests on anecdote and a handful of vendor case studies.

A structured corpus of real decision points across many systems, reported in aggregate. This is what the scoring API accumulates as a by-product, and it is the lab's intended first report once there is enough of it to be worth anything.