ContextPipe mechanism ablations

Determine, while holding the rest of ContextPipe fixed, the effects of (A1) predictive versus reactive pressure, (A2) SessionLatches versus per-turn re-evaluation, (A3) TTL, deduplication, and capacity limits on EmergentContext versus an append-only design, and (A4) ProviderCachePolicy versus fixed cache-marker placement.

Background

The paper evaluates the complete Structured ContextPipe pipeline against a Flat append-only context-construction policy, but identifies this comparison as only the first mechanism ablation. The unresolved experiments are intended to isolate the contributions of ContextPipe’s predictive pressure estimator, session-level latches, safeguards for emergent context, and provider-specific cache-marker policy while keeping other components unchanged.

The proposed ablations address design choices that may affect context-window utilization, cache stability, deduplication, staleness control, and compaction behavior. The paper notes that predictive and reactive pressure differ only slightly in the preliminary evaluation, making a controlled threshold-sensitive experiment necessary to distinguish them.

References

Still open, each holding the rest of ContextPipe fixed: (A1) predictive vs. reactive pressure (the two signals differ by at most 1.7\% of the window here, so separating them requires a knob where tier decisions sit near a threshold); (A2) SessionLatches vs. per-turn re-evaluation; (A3) TTL+dedup+cap on EmergentContext vs. append-only; and (A4) ProviderCachePolicy vs. fixed marker placement.

— ContextPipe: Database-Inspired Context Assembly for Long-Horizon Agents  (2609.00749 - Xu et al., 1 Sep 2026) in Section 7, subsection “Ablation status” (also reiterated in Section 7, subsection “Limitations and threats to validity”)