Delegation Cliff calculator

Move the twelve organisation levers and read how far the published coefficients say the quality margin moves, and which levers carry the movement.

The instrument

Each lever starts at zero, meaning no change from your current design. Move a lever by up to one increment either way.

Modelled change in the quality margin

0.00

Every lever is at zero.

Each lever's contribution: its change times its coefficient. The pale bar behind each row is what one full increment of that lever is worth. Bars to the right buy margin and bars to the left spend it; the scale is fixed, so a bar's length compares across settings.

What it computes

The change in the delegation margin, the gap between the quality an organisation of agents delivers and the floor it must hold, when you move one or more design levers away from wherever you start. It shows the net change, each lever's share of it, and how much reviewer capacity would cancel a net loss.

The formula in plain words

Multiply each lever's change, in increments, by its published coefficient, then add the twelve products. A negative total spends margin and a positive total buys it. This is the additive linear reading of the published coefficients: the article reports descriptive linear slopes and no intercept, so the instrument can give a change from your starting point and never an absolute margin.

What this does not say

It does not say whether your organisation sits above or below its quality floor, since the article publishes no intercept. It gives no interval, because the article prints none. The slopes describe configurations inside the transition band of the Sobol sample, so they are descriptive and say nothing causal about the space outside the band, and the interactions the article mentions between observability and capacity are absent from an additive reading. Model capability and agent self-check differ by 0.034 and the article reads them as tied.

Source

Every number comes from Pricing Agent Autonomy (Paper, Medium, 2026-09-03), with some from Accountability Makes Oversight Worse: A Pre-Registered Test of Liability-Exposed AI Supervision Across Irreversibility Tiers (Paper, 2026-09-10). The campaign record is Delegation Cliff.

Every number this instrument uses

Number	As published	Where
Linear-probe coefficients on the delegation margin, Sobol sample, band-only	labels: Delegation depth / Reviewer capacity / Task coupling / Model capability / Self-check calibration / Rework cost / Verification depth / Verification coverage / Detection lag / Workload volatility / Trust response / Escalation latency; values: −1.499, 1.456, −0.958, 0.949, 0.915, −0.803, 0.413, 0.308, −0.259, −0.147, 0.102, −0.007; reads: the price of autonomy / the lever that works / second-largest cost / tied for second / tied for second, paid once / friction compounds / complements capacity / complements capacity / slower is worse / minor / no spiral / no detectable effect	Paper, The numbers table, and the chart What moves the delegation margin
Configurations in the band-only fit	4948	Paper, Chart caption and footer
Gap between model capability and self-check calibration, read as tied	0.034	Paper, How to read these, and how not to
Unit of one increment, as the lab names it	band-standardised increment	Paper, Section 1, citing Pricing Agent Autonomy