Academic Compliance Review

Academic Compliance Audit & Deep Research Literature Review

Audits DeepDelve against its primary reference paper, DelveAgent (arXiv:2606.18648), and reviews 2026 literature for improvement ideas.

Compliance audit against DelveAgent

Tier structure. DelveAgent uses two tiers (Planner then Workers), DeepDelve uses three (Planner, Searcher, Analyzer). Compliant in spirit, but functionally suboptimal: the extra hop doubles token overhead and lets a Searcher's mistake cascade into the Analyzer, working against DelveAgent's whole cost efficiency point.

Adaptive planning loop, dual granularity memory, hierarchical reflection: fully or structurally compliant on all three at the time. The Planner replans on gaps or contradictions as designed, and the grounding check acts as the verification block, just checking web citations instead of physics constraints.

Note

The memory claim is stale: knowledge_cache.py, the module this cited for the coarse/fine grained memory split, was deleted. Its real failure turned out to be a benchmark isolation bug (a later model's bake off trial hit an earlier model's cached answer for the same query), not the RAG failure shape this compliance angle assumed, see the Literature Review's RAG reconsideration section. A successor, rag_cache.py, exists now, a flat semantic similarity cache over verified findings, a materially simpler design than the dual granularity split described here, not a like for like replacement.

Two relevant papers

"Marco DeepResearch" argues the biggest threat to small research agents is error propagation, since verification only at the end means early errors are hard to correct. DeepDelve only grounded final_report.md at the time; a step by step verifier checking each Specialist's summary before it reaches the Planner would catch fabrication at the source. Shipped since: upstream grounding checks now run on every specialist's summary before it reaches the Planner, confirmed directly in orchestrator.py, part of the same reliability batch documented in Completed.

"DEEPPLANNER" addresses planning fatigue with a constrained replanning tool instead of free text reasoning. DeepDelve's Planner free texted its replan decisions at the time, exactly what this paper argues against; a structured replan_action tool would fix that. Tried and reverted: a replan_action tool making the Planner's replanning decision a structured, checkable call was actually built and shipped, then deleted in a later cleanup pass as unused in practice, per Completed. The Planner still free texts its replan reasoning today; this specific fix was tried for real and didn't earn its keep, not merely unimplemented.

Recommendations

Collapse Searcher and Analyzer into one Research Worker tier with the full toolset, cutting one LLM hop per task. Still not done, the three tier split is still the live architecture. The other half of this recommendation, upstream verification of each Specialist's output before it reaches the Planner, is done, see above.