Literature Review

Literature Review

This is the SOTA literature review behind DeepDelve's architecture decisions. It's a standalone working document that started on 2026-07-19. Every claim below is marked as either primary source verified (someone on this project actually read the paper or data directly) or not yet verified (only seen through a WebSearch/WebFetch AI generated summary). Don't upgrade a claim's confidence without doing the verification pass first.

Why this exists

DeepDelve's own bake off (10 local model candidates, 9 disqualified, gpt-oss:20b the only full pass, see the Model Bake off page) raised a question: is the project hitting a real, externally documented ceiling on small model agentic reliability, or is this a DeepDelve specific gap? This review checks that against the current (2026) academic and industry literature. It uses exact terminology found in primary sources to chain into further relevant work (citation chaining, sometimes called snowball search) rather than relying on generic keyword search alone.

Methodology note (2026-07-19, a self correction partway through the review)

The first pass of this review (papers found via WebSearch) was presented with more confidence than it had earned. WebSearch and WebFetch return AI generated summaries, not primary reading, and several PDF fetches failed to parse properly. The user caught this and asked for the actual scientific method: read primary sources directly, verify claims against the real text, data, and tables, and make corrections transparently wherever a summary turned out to be wrong or imprecise. Papers marked ✅ below are primary source verified. Papers marked ⚠️ are still only seen through a search summary and shouldn't be treated as verified fact.

Parts

1. Verified Papers: primary source verified papers, the bulk of the review.
2. Leads, Corrections & Open Questions: leads found through terminology chaining that aren't verified yet, citations we downgraded or corrected, reads we rejected, and questions still open.
3. Architecture Synthesis: the routing layer proposal, the comparative survey against other projects, the RAG reconsideration, the novelty assessment, and the citation fabrication solutions.
4. Hardware & Serving: ROCm versus CUDA, Ollama and llama.cpp tuning, and the Qwen3 think suppression investigation.
5. Bake off Findings: literature checked directly against live bake off incidents.
   • Part I: Ornith 1.0 9B, the Searcher over fetching root cause, and the extended comparative survey.
   • Part II: the Lisbon/Mexico convergence chase, the 2026-08-17 seven fix arc, and AgentFloor.