Skip to content

Delegation Capacity

Paper reference: Section 1, "Delegation capacity"; Theorem 1

Before optimizing review allocation or tuning a corrector, ask the most basic question: can this pipeline hit the target quality at all?

The operational ceiling

The delegation capacity \(C_\text{op}\) is the best possible output quality the pipeline can achieve under the stated model, topology, and budget. If the quality target \(p_\text{min}\) exceeds \(C_\text{op}\), local governance tuning alone cannot rescue the design; capability, review budget, task decomposition, or topology must change.

\[C_\text{op}(G, K, B) = \sup_{p(\text{task})} q^*(\text{output})\]

Here \(q^*(\text{output})\) is the shipped output quality at the sink after available correction. Raw support \(\sigma_\text{raw}\) remains the authorization signal because it is the less masked estimate of agent competence.

from minimal_oversight.capacity import check_feasibility

report = check_feasibility(pipeline, p_min=0.80)
print(report.explanation)
# INFEASIBLE: Quality target p_min=0.800 exceeds pipeline
# capacity C_op=0.725. Within the fixed model, topology, and
# budget, no local governance policy can rescue this design.

When the MSO has no solution

If \(p_\text{min} > C_\text{op}\), the theory prescribes no delegation — not "more authority." The task must be performed by a more capable agent, decomposed into subtasks, or the topology must change.

For a single node

A single node with observation rate \(\eta\), decay/reversion rate \(\delta\), and baseline support \(\sigma_0\) achieves maximum capacity when \(\sigma_\text{skill} = 1\):

\[C = \frac{\eta + \delta\sigma_0}{\eta + \delta}\]

With the conservative \(\sigma_0=0\) default, \(\eta = 10\), and \(\delta = 2\), \(C = 0.833\). This is the ceiling even for a perfect agent because stale evidence and environment shift pull measured support back toward the baseline.

For a chain

Each layer degrades the signal. The recursive formula accounts for the fact that each corrector stabilizes quality before passing it downstream:

\[\sigma_\text{corr}^*(i) = R\big(\sigma_\text{skill} \times \sigma_\text{corr}^*(i-1)\big), \quad \sigma_\text{corr}^*(0) = 1\]

where \(R(\cdot)\) is the Return Operator at fixed point (Equation 11).

Verify the math 
from minimal_oversight._formulae import recursive_chain_quality

# How does quality degrade with depth?
for D in range(1, 8):
    q = recursive_chain_quality(D, sigma_skill=0.55, catch_rate=0.65, eta=10, delta=2)
    print(f"D={D}: C_op = {q:.3f}")

The theory-observation gap is less than 0.002 across all 28 conditions tested in the paper (Experiment 6).

Correction model: theory vs simulation

The closed-form equations (Eq. 5-6) use raw \(c\) as the catch rate. The paper's simulator uses \(c \times K/N\) as the effective catch rate, where \(K/N\) is the fraction of outputs reviewed. This means:

  • Theoretical \(M^*\) = 1.83 (with \(c = 0.70\), Eq. 6)
  • Simulated \(M^*\) ≈ 1.4 (with \(c \times K/N = 0.70 \times 0.50 = 0.35\))

If your observed \(M^*\) is lower than the theoretical prediction, check your effective review coverage.

Critical depth

There is a maximum target-feasible depth beyond which adding layers drives the recursive corrected-chain quality below the required target:

\[D_\text{max} = \max\{D: C_\text{op}(D) \geq p_\text{min}\}\]

This is computed directly from the recursive formula (Eq. 11). For \(\sigma_\text{skill} = 0.55\), \(c = 0.65\), and the conservative \(\sigma_0=0\) default, a demanding target \(p_\text{min}=0.80\) permits one layer but not two. A lower target such as \(p_\text{min}=0.50\) has no finite cliff in the stabilized corrected-chain model.

Better correctors extend the useful depth significantly because each layer passes a more reliable corrected signal downstream.

The effective autonomy buffer

The buffer combines capacity, quality target, and workflow complexity into one number:

\[B_\text{eff} = C_\text{op} - p_\text{min} - \lambda H(W)\]
  • \(B_\text{eff} > 0\): delegated autonomy is feasible
  • \(B_\text{eff} = 0\): at the autonomy cliff
  • \(B_\text{eff} < 0\): the fixed design lacks sufficient margin to maintain quality

The complexity tax \(\lambda H(W)\) captures how routing entropy, tool-call variability, and timing uncertainty consume the quality margin. Each additional bit of process entropy costs approximately \(\lambda \approx 0.02\) in quality (Experiment 7).