Deep Research Enhancement Program — Synthesis & Priorities (2026-08-01)

Deep Research Enhancement Program — Synthesis & Priorities

Date: 2026-08-01 Inputs: deep-research-verification-2026-08-01.md, deep-research-fundamentals-2026-08-01.md, deep-research-competitive-landscape-2026-08-01.md, deep-research-trust-novelty-scoring-landscape-2026-08-01.md, deep-research-model-agnostic-multi-provider-and-skills-publication-2026-08-01.md. Supersedes the priority ordering (not the technical content) of deep-research-capabilities-audit-2026-06-17.md.

---

Strategic framing

The competitive research is unambiguous: every major deep-research product (Google, Anthropic, OpenAI, Perplexity, You.com/Grok) already claims breadth, iteration, and speed — that axis is saturated. The one gap none of them has convincingly closed is verifiable trust infrastructure: citation-to-claim auditability, reproducible run traces, honest confidence signaling, and novelty/quality scoring users can actually check. Vox’s own pipeline independently corroborates this from the inside — gate.rs is a real multi-signal fusion but has no trust weighting; NoveltyEvidenceBundle/WorthinessSignalsV2 are fully-specified schemas with nothing populating them; citation grounding exists structurally but has no post-hoc verification pass. This convergence — what the market lacks and what Vox’s own schema already anticipates but hasn’t wired — is the strategic bet: differentiate on trust/novelty/verification rigor, not on out-executing breadth or speed.

Prioritized gap list (impact × effort)

P0 — Near-zero effort, do immediately

1. Fix stale PHASE_0a_STUB doc comments in crates/vox-research-shim/src/research/mod.rs and verifier.rs — they falsely claim stub/empty behavior over code that is fully real, actively misdirecting anyone (human or agent) who greps for stub markers. (Verification doc, New Finding #1)
2. Unify the G1/G3 pipeline split — LLM-driven CRAG expansion (web_gather.rs) and novelty scoring (vox-search::research.rs) each landed in only one of two independent research loops (the primary vox-research-shim pipeline vs. the autonomous-agent run_multi_hop_web_research loop). Both fixes already exist; share them so neither loop is missing a capability the other already has. (Verification doc, New Finding #3)

P1 — High impact, medium effort — the trust/novelty core (directly answers the user’s original ask)

3. Populate ResearchHit.trust_score via Crossref retraction lookup (hard-zero) + OpenAlex venue-type/author-h-index (soft weight); feed it into gate.rs::score_with_config so citation_score becomes trust-weighted, not a flat count. (Trust/novelty doc §2)
4. Add an NLI first pass to verifier.rs (ONNX MiniCheck/DeBERTa-mnli-fever-anli, in-process via fastembed/ort) ahead of the existing LLM cascade, escalating only low-confidence claims — cuts LLM calls ~40% and adds a signal independent of the current single-shot JSON path. Reconcile Verdict to the SciFact taxonomy the code already says is owed. (Trust/novelty doc §4)
5. Populate NoveltyEvidenceBundle via OpenAlex/Crossref/Semantic Scholar prior-art queries + a two-stage MinHash-LSH → embedding-cosine (fastembed/hnswlib-rs) pipeline, replacing the URL-only and finding_id-only dedup at both existing dedup sites. This is Vox’s own most conspicuous schema-without-implementation gap. (Trust/novelty doc §3)
6. Populate WorthinessSignalsV2 hard/soft/diagnostic gates (retraction hard-gate, peer-review-status soft-gate, statcheck-style numeric-claim diagnostic) — the schema is already the right shape (reason-coded, auditable), just unfed. (Trust/novelty doc §5)
7. Add a post-hoc citation audit pass: after synthesis, re-fetch every cited source and confirm the quoted claim actually appears in it, surfacing a per-claim verification badge. This is the single most-requested-by-implication fix from the competitive research — no reviewed competitor does this rigorously at open-web scale. (Competitive-landscape doc, “What would beat all of these” #1; Fundamentals doc §7)

P1 — High impact, medium effort — multi-provider routing (unlocks the GUI keys already collectable)

8. Route research stage selection through decide() (or port key_guard::available_inference_providers() into the scorer’s filter) so a resolved model always has a configured key. (Multi-provider doc, Part A2)
9. Make cascade_for_research_stage provider-aware beyond its current two lanes (local + OpenRouter) — dispatch to whichever of the 9 already-collectable provider keys actually owns the resolved model, using the ModelRouteBackend/ChatRouteBackend plumbing the chat path already has. (Multi-provider doc, Part A2) This is the prerequisite for #10.
10. Per-stage free-tier-aware selection: once #8/#9 land, route cheap/fast stages (planning, claim extraction) to Groq/Cerebras free tiers and stronger stages (synthesis, judge) to Gemini Flash/Mistral Experiment rather than a rate-starved single lane. Also fix the scorer’s blanket Performance ⇒ exclude free models rule so QualityLevel (currently dead plumbing) can actually express a preference. (Multi-provider doc, Part A3, Part B)

P2 — Well-scoped, moderate effort

11. Cross-encoder reranking — no implementation exists; published benchmarks show ~13pp recall@10 gain from adding this over RRF-only fusion. (Fundamentals doc §6, Verification doc G2)
12. Bump/parameterize synthesis_max_tokens — currently defaults to 1200, actually lower than the 1800 the original audit flagged as too small. Cheap fix, direct quality impact on report length. (Verification doc G6)
13. Semantic result caching — replace the exact-match FNV1a cache key with embedding-similarity lookup. (Verification doc G7)

P3 — Longer-horizon / infrastructure

14. Durable async research jobs — current tokio::spawn fire-and-forget loses in-flight runs on process restart; needs vox-orchestrator-queue oplog wiring. (Verification doc G8)
15. Per-hop span observability — structured telemetry, not just log lines. (Verification doc G9)
16. Reflexion-style iterative self-revision — current self-check (run_self_verification) is single-pass CoVE-style, not iterate-until-satisfied. (Fundamentals doc §3)
17. Internal evaluation harness — a small BrowseComp/DeepResearch-Bench-style rubric set for the composed pipeline; currently nothing validates end-to-end report quality, only unit-level correctness. (Fundamentals doc §9)

P4 — Deferred until P1 lands (explicitly gated, not just deprioritized)

18. Deep research as an ad-hoc SKILL.md — packaging a still-stub-heavy pipeline as a more-discoverable skill just exposes the stubs to more callers; do after the P1 trust/novelty work closes. (Multi-provider doc, Part C)
19. SCIENTIA/VoxGiantia auto-publication handoff — should trigger only on findings that clear the (currently unpopulated) WorthinessSignalsV2 gates from #6; auto-publishing before that lands would be strictly worse than the current human-gated status quo. (Multi-provider doc, Part D)
20. GUI settings-search indexing for API keys — small, independent fix (federatedSearchIndex.ts doesn’t index the Keys & Secrets page); not blocking, can land anytime. (Multi-provider doc, Part A1)

Recommended next step

This synthesis, together with the four research docs and the verification pass, is the complete input for an implementation plan. Per the brainstorming skill’s own terminal-state rule, the next step is not further research — it’s translating items #1–#10 (P0/P1) into a concrete implementation plan via the writing-plans skill, since those ten items are where impact and evidence both concentrate: they’re simultaneously what the user asked to prioritize (trust/novelty/bad-research surfacing, multi-provider intelligence) and what the competitive research independently confirms is the open market gap.