# Delegation Cliff

Source: https://dinand.com/research/campaigns/delegation-cliff/
Author: Dinand Tinholt

> Each added level of delegation depth costs 1.50 on the quality margin; reviewer capacity buys back the most at +1.46.

## The question

Every organisation deploying AI agents is betting that pushing work further from human review is cheaper than keeping it close. This campaign priced that bet: where does delegation stop paying, and which design choices buy the margin back?

## What was registered

Predictions registered before the sweep.

- The design space has a cliff, a sharp boundary between robust and fragile configurations.
- Better observability outranks raw reviewer capacity.
- An audit death spiral, where eroding trust raises audit load until capacity saturates and quality falls.

## What was measured

A simulated agent organisation with twelve design dimensions, including delegation depth, reviewer capacity, task coupling, model capability and self-check calibration. 90,880 configurations were each scored on the margin between the quality the organisation delivers and the floor it must hold. The campaign finished with zero error units. A revalidation certificate reran 200 random configurations from scratch: median difference 0.002, 95th percentile 0.0117, against a tolerance of 0.05.

## What it found

Delegation depth carries the largest weight in the space, and it is negative: each increment costs 1.50 on the margin. Reviewer capacity buys back the most at +1.46. Model capability (+0.95) and self-check calibration (+0.92) follow, close enough to read as tied. Task coupling is the second-largest cost at −0.96.

The quality floor decides the failure geometry. At a 0.95 floor, 60% of the space is a broad transition band with no cliff. At 0.99, 79% of the space is fragile and the cliff appears. The observability prediction failed: observability complements reviewer capacity and does not replace it. The spiral prediction failed: mean margin rises across trust-response quintiles, including in the high-stress stratum.

Headline numbers, as published
| Change in quality margin per increment of delegation depth (Sobol band-only fit, n = 4,948) | −1.50 |
| --- | --- |
| Change in quality margin per increment of reviewer capacity, the largest positive lever | +1.46 |
| Model capability against agent self-check calibration (read as tied) | +0.95 against +0.92 |
| Share of the design space in a broad transition band at a 0.95 quality floor (no cliff) | 60% |
| Share of the design space that is fragile at a 0.99 quality floor (one 256-unit sweep) | 79% |
| Revalidation certificate: 95th percentile difference on 200 rerun configurations, tolerance 0.05 | 0.0117 |

*What moves the delegation margin: linear-probe coefficients on the Sobol sample, band-only, n = 4,948. Positive buys margin, negative spends it. Rebuilt from the numbers in [Pricing Agent Autonomy](https://medium.com/@tinholt/pricing-agent-autonomy-0a2bb5b97caf).*

Show the numbers as a table

| Design dimension | Coefficient |
| Delegation depth | −1.499 |
| Reviewer capacity | 1.456 |
| Task coupling | −0.958 |
| Model capability | 0.949 |
| Self-check calibration | 0.915 |
| Rework cost | −0.803 |
| Verification depth | 0.413 |
| Verification coverage | 0.308 |
| Detection lag | −0.259 |
| Workload volatility | −0.147 |
| Trust response | 0.102 |
| Escalation latency | −0.007 |

## What it does not claim

- The coefficients are descriptive slopes on the Sobol sample inside the transition band, n = 4,948. They describe how the margin moves where a boundary exists. They are not causal estimates over the whole space.
- The 0.99 result rests on one 256-unit calibration sweep. The 0.95 result is replicated on the 8,192-unit Sobol set. The two are not equally evidenced.
- Reviewer capacity is a parameter in this design. Reviewers do not tire or queue. The follow-up adds both.

## Follow-up: reviewers who tire

A second study, pre-registered on 2026-09-04 and complete on 2026-09-05, ran 90,568 units with reviewers who tire and queues that back up. Reviewer capacity still pays, at +0.93 against the +1.46 published here, as long as sustained utilisation stays below 0.8. Above 0.8 the margin goes negative. Escalating more with headcount fixed scored −0.59. The same escalation with staffing scaled to match scored +1.15.

One pre-registered hypothesis was refuted as written: the clause that the margin never goes negative, and the flat-then-knee shape that came with it.

Follow-up numbers
| Reviewer capacity lever below 0.8 sustained utilisation, reviewers who tire (published lever +1.46) | +0.93 |
| Escalating more with headcount fixed | −0.59 |
| The same escalation with staffing scaled | +1.15 |

## Read the pieces

- [Pricing Agent Autonomy](https://medium.com/@tinholt/pricing-agent-autonomy-0a2bb5b97caf)Paper, Medium, 2026-09-03
- [The Oversight Yield Curve](https://dinand.com/research/oversight-yield-curve/)Follow-up paper, 2026-09-05
- [Staff Before You Escalate](https://dinand.com/research/staff-before-you-escalate/)Executive brief on the follow-up, September 2026

---
Site guide for AI agents: https://dinand.com/llms.txt