\subsection{Repeated handoffs and error correction (E4)}
\label{sec:handoffs}
Twenty synthetic test documents were relayed through chains of 1, 2, 4 and 8 handoffs (Figure~\ref{fig:handoffs}, Table~\ref{tab:handoffs}).
Only the first agent sees the document; every later agent sees only the previous message (closed) or the previous message plus metered
re-reads of source sentences for the project codes it mentions (repairable; at most four lookups and 1\,500 bytes per hop). Both prose (A4)
and structured (A5) relays were run with a homogeneous Qwen chain and with a Qwen/Llama alternating chain; after each measured hop the last
agent answered all fifteen question types, and the evaluator-only ledger (62 atomic facts per document) measured retention and unsupported
additions in each message. 7\,200 answer trials were produced.

\begin{figure}[t]\centering\includegraphics[width=\textwidth]{fig5_handoffs.pdf}
\caption{Repeated handoffs on 20 synthetic test documents: retained ledger facts, unsupported numbers/identifiers per message, and question
accuracy after 1, 2, 4 and 8 handoffs. Solid = closed chains, dashed = repairable chains with metered source lookups; filled circles = Qwen
only, open squares = Qwen/Llama alternating.}\label{fig:handoffs}\end{figure}

\begin{table}[t]\centering\scriptsize
\caption{Repeated handoffs (means over 20 documents; 95\,\% cluster-bootstrap CIs). Retention = share of the 62 ledger facts whose value appears
in the message; unsupported = numbers ($\ge$ 4 digits) or identifiers in the message that do not occur in the document; fetched = cumulative
bytes re-read from the source by repairable agents.}\label{tab:handoffs}
\adjustbox{max width=\textwidth}{\input{tables/table_handoffs}}
\end{table}

\paragraph{The loss happens at the first hop; relays are nearly lossless afterwards (H5 rejected).} A 512-token prose summary retains
${\approx}32$\,\% of the ledger facts and supports ${\approx}40$\,\% of the answers after one handoff; after eight closed handoffs the figures are
${\approx}31$\,\% and ${\approx}39$\,\%---a drift of one to two points, within the intervals. Structured relays are even more stable: from hop~2 the
Qwen chain re-emits a byte-identical JSON (709 bytes) at every hop, a fixed point of the ``re-emit this evidence'' instruction. The
registered expectation of monotone decay below 70\,\% retention after four hops is rejected for both representations, and so is the
expectation that structured evidence would protect facts better than prose: at four handoffs the A5 chain answers 12 points \emph{fewer}
questions than the A4 chain (T5, paired difference $-0.12$, 95\,\% CI $[-0.21,-0.03]$, bootstrap $p=0.013$, 300 instances in 20 documents),
for the same reason as in the single-handoff study---the budget buys ${\approx}8$ quoted facts.

\paragraph{Unsupported additions are rare.} Across 7\,200 relayed messages the ledger metric found no unsupported identifier and at most one
unsupported number in a single repairable Qwen chain (0.05 per message); the 3B agents compress by \emph{omission}, not by invention, and the
receiver answers ``unknown'' for omitted facts. Alternating Qwen and Llama agents neither helped nor hurt (differences within one point).

\paragraph{Metered repair recovers a little at a growing price (H6 partially confirmed).} Allowing each relay agent to re-read the source
sentences for the project codes in the notes raised retention from 32 to 38\,\% and accuracy from 40 to 44\,\% for the prose chain, with no
further gain after the fourth hop; the cumulative fetched volume reached 5.9~kB by hop~8, i.e., about five messages' worth of traffic for a
four-point gain. For the structured chain the repair added two points of retention and nothing in accuracy: the agent could re-read, but the
re-emitted JSON did not grow. Evidence access therefore repairs what the previous message still \emph{names}; facts whose entity was already
dropped cannot be requested, which is the practical limit of reference-based repair without the original question.