Dynamic Prompt Assembly — References

The aggregator is a fan-out with deadlines, so the useful literature is about latency under fan-out, partial failure, and reproducible artefacts.

Fan-out and deadlines

• The Tail at Scale (Dean & Barroso, Communications of the ACM, 2013) — why a request that waits on several independent sources inherits the tail of the slowest, and the standard mitigations: per-source deadlines, hedged requests, and returning a good-enough result from the sources that answered. The core latency argument for this component.
• Release It!: Design and Deploy Production-Ready Software (Michael T. Nygard) — the stability patterns and antipatterns, in particular unbounded waits on an integration point and the cascading failure that follows. Cited by name; the book is in its second edition.

Reproducing what was sent

• Reproducible Builds — the discipline of making an artefact a deterministic function of recorded inputs, including the usual culprits for non-determinism: timestamps, iteration order, and concurrency. Directly applicable to an artefact assembled from concurrent producers.
• OpenTelemetry semantic conventions — the attribute names for recording the model call and its surrounding spans, worth adopting so that per-producer timings and the assembled artefact share one trace identity.

Why composition order is not neutral

• Lost in the Middle: How Language Models Use Long Contexts (Liu et al., 2023), arXiv:2307.03172 — the finding that the position of material within a context changes how well it is used, which is what makes slot order a design decision rather than an implementation detail.

No verified reference is recorded for assembly manifests, declared degradation policies, or cross-producer deduplication of prompt material. These are engineering practice specific to this component and no published treatment of them is recorded here.