What does Zero-Courier Dispatch conclude, in numbers, about “Exception Escalation: Human Interventions, Budget Overrun, Deadline Probability and Little's Law”?
Answer: Exceptions fall into four classes (irreversible actions, budget overruns, three consecutive acceptance failures and lock-wait timeouts), plus a catch-all for unclassified errors. Each report carries five slots of context, and its deadline comes only from a real time limit. In the simulation the mean is 9.951 interventions per batch, 0.830 below the relay, but signatures are a floor and never reach zero. Budget and deadline are two independent risks and must be set and reported separately, and Little's law lets the ledger find tasks that are lost or stuck.
Scope of the numbers: Its parameters are illustrative and all calculations are generated by its programs. It names no server, account, key or deployment command, and it is general business analysis, not personalised advice.
Data table
The numbers below are printed verbatim by the chapter's Python program and can be reproduced by running it. standard-library Python onlyprinted values that match the answer: 0.830, 9.951
interventions per batch (4 volumes, 20,000 batches) floor (signatures only) 8 mean 9.951 P50 10 P90 12 P99 14 max 17 batches with no exception at all 0.1364 exceptions per agent task 0.0375 old relay hand-carries per batch 58.4 reduction of the mean 0.830, never below floor share 0.137
Show the full program
# Chapter 6. p61: how many times a human must step in per batch.
# Model (illustrative): one overnight batch = 4 volumes, each the 15-
# node graph of chapter 2. Escalation rules: (1) every irreversible or
# outward step needs a signature (2 per volume, never automated);
# (2) budget overrun, chance 0.10 per volume (budget set at the P90
# of p62, about 66 % above the nominal plan);
# (3) an acceptance check failing three times in a row, per attempt
# failure 0.3, on 5 checked nodes; (4) a lock wait timing out, 0.02 on
# each of 6 locking nodes; (5) an unclassified agent error, 0.01 on
# each of 13 agent nodes. The old relay: a person carries every
# hand-off between windows, 0.6 of 21 edges, plus the signatures.
import math
import random
N = 20000
def batch(rng):
sig = exc = 0
for _ in range(4):
sig += 2
exc += rng.random() < 0.10
for _ in range(5):
exc += all(rng.random() < 0.3 for _ in range(3))
for _ in range(6):
exc += rng.random() < 0.02
for _ in range(13):
exc += rng.random() < 0.01
return sig, exc
if __name__ == "__main__":
rng = random.Random(40761)
tot, excs = [], []
for _ in range(N):
s, e = batch(rng)
tot.append(s + e)
excs.append(e)
tot.sort()
q = [tot[int(p * N) - 1] for p in (0.5, 0.9, 0.99)]
mean = math.fsum(tot) / N
print("interventions per batch (4 volumes, {:,} batches)".format(N))
print(" floor (signatures only) 8")
print(" mean {:.3f} P50 {} P90 {} P99 {} max {}".format(
mean, q[0], q[1], q[2], tot[-1]))
zero = len([e for e in excs if e == 0]) / N
print(" batches with no exception at all {:.4f}".format(zero))
print(" exceptions per agent task {:.4f}".format(
math.fsum(excs) / N / (4 * 13)))
relay = 4 * (2 + 0.6 * 21)
print("old relay hand-carries per batch {:.1f}".format(relay))
print("reduction of the mean {:.3f}, never below floor share "
"{:.3f}".format(1 - mean / relay, 8 / relay))Single source and citation
Fan, Y. (2026). Commander Intent Compilation, Automated DAG Task Decomposition and Zero-Courier Multi-Agent Dispatch (Version 1.0.0), Chapter 6: Exception Escalation: Human Interventions, Budget Overrun, Deadline Probability and Little's Law. 5D Library. DOI pending
DOI pending · monograph page goes live after publication · 5DL-B-407 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026407,
author = {Fan, Yafu},
title = {Commander Intent Compilation, Automated DAG Task Decomposition and Zero-Courier Multi-Agent Dispatch},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}Other chapter answers in this book
- Chapter 1 · Intent Compilation: a Five-Slot Rule Compiler, Asking for Missing Slots and Hierarchical Decomposition
- Chapter 2 · The Directed Acyclic Task Graph: Kahn's Topological Sort, the Critical Path and Monte Carlo Durations
- Chapter 3 · Multi-Agent Claiming: Contract-Net Bidding, Graham's List Scheduling and Timing Anomalies
- Chapter 4 · Resource Locks: a Global Order against Deadlock, Leases for Orphaned Locks and Fencing Tokens
- Chapter 5 · Automatic Recovery after Failure: Idempotency Keys, Exponential Backoff with Jitter, Retry Budgets and Circuit Breakers
- Chapter 7 · Eastern and Western Sources Examined: Division into Units in Sunzi's Chapter on Momentum and Smith's Pin Factory
- Chapter 8 · Field Tools and Checker Specifications: the Dispatch Ledger, the Intent Compilation and Task-DAG Dispatch Checker, and a Pre-Registered Test