Multi‑agent pipelines must both ingest a requester’s forms and author formal documents based on them. We present a deployed tender‑response system that runs an open‑weights model under sovereignty constraints and compare its output with the actual human‑written bids submitted by the same organization.
In a blind test where the system had no worked example, an LLM judge evaluated $55$ ground‑truth sections and found the system’s answers at least as good as the human ones in $40$ cases, better in $4$, with no missing answers and only one unsupported claim.
Classifying every gap identified by the judge shows that $68\\%$ stemmed from content absent in the system’s own sources—knowledge the human author possessed but the pipeline never received—so only $6$ of the $15$ adverse verdicts involved a deficiency the system could have avoided. Divergence from ground truth is more often an information‑availability issue than a writing‑quality one, and evaluations that do not separate the two underestimate such systems.
Against this backdrop we observe a conditioning asymmetry. It is well known that representing documents as structural markup rather than flat prose improves extraction, and we reproduce this on three reading tasks. However, the benefit does not transfer to conditioning: converting a bid’s \emph{instruction} material from prose to nested XML dropped answer quality from $74\\%$ to $48\\%$ in a paired comparison.
Further experiments reveal that explicitly forbidding a construction concentrates rather than eliminates it—$96\\%$ of the remaining defects fall into the two forms the prompt explicitly names—and that coupling a stochastic annotation with a deterministic windowing function reduces the extracted requirement count on a byte‑identical file from $68$ to $51$.
Thus, structure belongs where the model reads; prose and self‑applied tests belong where it writes.
Blogger's Review: This study highlights the divergent roles of structured markup and natural language in multi‑agent document generation, offering practical guidance for designing pipelines that separate reading‑time structure from writing‑time prose.