SCIENTIA Automated Research: Historical Roles and Extension Research (2026)

SCIENTIA Automated Research: Historical Roles and Extension Research (2026)

1. Executive summary

Vox SCIENTIA is more than a nanopublication emitter: it is a knowledge platform (vox-scientia) plus publication orchestration (vox-publisher), fed by deep research (vox-research-shim) and hybrid retrieval (vox-search). Nanopubs are one rigor tier on the output ladder — alongside Zenodo deposits, RO-Crate packages, manuscript scaffolds, and syndication — all behind a human-gated review SSOT.

This document:

1. Maps historical roles of computers and AI in scientific research (1960s–2026) to Vox capabilities and gaps.
2. Records a severity-ordered code review of the current SCIENTIA + research integration (June 2026).
3. Proposes a unified research loop that connects coding agents, spontaneous research, and optional gamification.
4. Sequences extension work aligned with existing plans (Tracks A–E plus Track F: Unified research loop).

Term disambiguation (common voice-transcription confusions):

Spoken / written	Vox canonical
TAN, TANTIVLY	Tantivy — Rust full-text index behind tantivy-lexical / VOX_SEARCH_TANTIVY_ROOT
Tavily, Tavoli	Tavily — live web retrieval vendor; Rust SDK in vox-search
Graphify	External Graphify pipeline + local graphify-out/ convention; P0 status in Vox

External research for §2 used Tavily Research API (2026-06-16) plus NLM/Cochrane/nanopub primary sources. Tavily CLI (tvly) was unavailable in the agent environment; MCP tavily_research was used instead.

---

2. Historical roles of computers and AI in research

Five eras structure how automated research evolved and what researchers still expect tools to save.

flowchart TB
  subgraph eras [Historical eras]
    E1["1960s-80s: IR and bibliographic DBs"]
    E2["1990s-2000s: Systematic review and citation graphs"]
    E3["2010s: Reproducibility and preregistration"]
    E4["2018-2023: Semantic search and ML screening"]
    E5["2024-2026: LLM agent deep research"]
  end
  subgraph vox [Vox SCIENTIA layers]
    L1[Signal producers and scout]
    L2[Retrieval: CRAG Tavily Tantivy Graphify]
    L3[Claim and novelty verdict ladder]
    L4[Human gate nanopub RO-Crate]
    L5[Scholarly syndication Zenodo]
  end
  E1 --> L2
  E2 --> L3
  E3 --> L4
  E4 --> L2
  E4 --> L3
  E5 --> L1
  E5 --> L2

2.1 Era timeline with citations

Year	Milestone	Capability	Researcher time impact	Primary source
1960s	MEDLARS mechanized indexing	Batch bibliographic processing at NLM	Weeks → days for curated index updates	NLM MEDLARS history
1971	MEDLINE online search	Interactive biomedical literature retrieval	Search weeks → minutes	NLM Circulating Now
1993	Cochrane Collaboration	Standardized systematic review methods	Reduced ad-hoc duplication; reviews still multi-month	Cochrane Handbook Ch. I
2010	Nanopublications (Groth et al.)	Atomic, signed, RDF-linked scientific assertions	Full paper → citable micro-assertion	Groth et al. 2010
2012+	OSF registrations / preregistration	Time-stamped, frozen study plans	Post-hoc bias → auditable protocol	OSF Registrations guide
2015	Semantic Scholar	AI-enhanced paper discovery and TL;DR summaries	Manual scan → ranked semantic results	Semantic Scholar About
2019	RO-Crate 1.x	FAIR packaging of data, software, metadata (JSON-LD)	Manual provenance assembly → structured crate	RO-Crate introduction
2023–24	Elicit systematic review	LLM-assisted screening and extraction (PRISMA-oriented)	Up to ~80% review time reduction (vendor claim)	Elicit systematic review
Feb 2025	OpenAI Deep Research	Multi-step autonomous web research agent	Tasks framed as days/weeks of human work → minutes	OpenAI introducing deep research
2025–26	Gemini Deep Research	Plan → search → read → iterate → cited report (async)	Literature review / due diligence acceleration	Gemini Deep Research API
2026	Tavily /research endpoint	Single-call multi-hop vendor research for agents	Replaces hand-rolled search loops for integrators	Tavily Jan 2026 ship blog

2.2 Historical capability → Vox mapping

Historical capability	Vox today (shipped)	Missing for “more than nano publisher”
Batch literature search	Tavily /search Tier 4 + OpenAlex/Crossref/S2 prior art	Tavily /research, /extract, /crawl; arXiv-native tier; citation-diversity gate
Citation chaining	publication-novelty-fetch federated prior art	Mandatory embedding path; SPECTER2/Qdrant KNN in novelty scoring
Systematic review screening	claim_extractor + MiniCheck verdict ladder	Batch screening UI; inter-rater export; PRISMA trace
Preregistration	Orchestrator preregistration/ + scientia_prereg table	Doc drift: gap map cites crates/vox-prereg/ (absent); see §8
Reproducibility packages	RO-Crate + replay runner scaffold	Measured replay score auto-written to worthiness rubric
Semantic literature AI	SCIENTIA claim pipeline on publication path	Auto-bridge from vox research run sessions
Agent deep research	run_research + CRAG multi-hop + MCP vox_research_run	Reliable planner; mesh emission; eval harness
Micropublications	Offline RSA nanopub + test-server gate	Aggregate approved claims → IMRaD manuscript (Gap C)
Codebase knowledge graphs	Graphify P0 (vox graphify status)	P1 lexical ingest + MCP query in retrieval bundle
Model routing observability	OpenRouter catalog + research Lane G synthesis	Provider Atlas → ModelRegistry feedback loop

Design implication: SCIENTIA should operate as a research OS — discover → retrieve → extract → assess novelty → human approve → emit at chosen rigor — with coding-agent sessions writing into the same ledger as spontaneous literature research.

---

3. Current architecture

sequenceDiagram
  participant Agent as CodingAgent
  participant RS as vox_research_shim
  participant VS as vox_search
  participant OR as OpenRouter
  participant TV as Tavily
  participant SC as vox_scientia
  participant Human as Researcher
  participant Out as NanoPub_Zenodo_Manuscript

  Agent->>RS: vox research run
  RS->>OR: planner + synthesize
  RS->>VS: CRAG multi-hop
  VS->>TV: Tier4 search optional
  RS->>SC: claim_extractor optional paths
  Note over RS,SC: FindingCandidate mesh emission not wired
  SC->>Human: discovery inbox review
  Human->>SC: ApprovalToken
  SC->>Out: nanopub / archive / syndicate

Crate responsibilities:

Crate	Role
vox-scientia	Claims, novelty inspect bridge, review flow, nanopub, RO-Crate, producers, manuscript scaffold
vox-publisher	Prior art fetch, worthiness, syndication, discovery ranking, research mesh intake
vox-research-shim	run_research pipeline, session persistence, web gather
vox-search	CRAG, web tier, Tavily client, optional Tantivy/Qdrant
vox-research-events	Contract types: FindingCandidateV1, ResearchEvent, mesh broadcast

Human-gated review SSOT: review_flow.rs. Four-layer no-auto-publish design: scientia-micropublication-ssot-and-surfacing-design-2026.md.

---

4. Code review findings (June 2026)

Ordered by severity. Status updated after June 2026 implementation waves.

Fixed June 2026: findings 1 (scientia-claims feature), 4 (empty/no-trace → InsufficientEvidence), 6 (planner_degraded + user-visible metadata), 14 (cache store via upsert_knowledge_node). Fixed implementation waves (June 2026 plan): 2 (embedder fail-fast + golden harness), 3 (bundle SSOT unification), 5 (discovery bridge + mesh), 7 (Tavily extract/research tiers), 8 (Graphify P1 lexical leg), 9 (assess_novelty callsite audit), 11 (research kudos when gamify enabled).

4.1 Critical — correctness / trust

1. Research shim claim path diverges from SCIENTIA claim extractor — FIXED June 2026 (scientia-claims feature routes through claim_extractor first)

   • claims.rs uses LLM cascade when runtime feature is on, else returns empty claims.
   • claim_extractor/pipeline.rs is the real VeriScore → atomic → MiniCheck pipeline.
   • Risk: vox research run can complete with zero claims while publication-extract-claims produces verdicts.

2. Novelty scoring without embedder degrades to lexical-only — MITIGATED June 2026 (require_embedder_for_online_novelty when VOX_SCIENTIA_REQUIRE_EMBEDDER=1; golden harness added)

   • scientia_prior_art.rs enriches semantic scores only when caller supplies Embedder.
   • Risk: false Novel/NotNovel at scale; no measured precision/recall harness yet (upgrade plan Track A).

3. Bundle type drift — FIXED June 2026 (NoveltyEvidenceBundleContract alias + parity tests)

   • Hand-written NoveltyEvidenceBundleV1 (publisher) vs typify types in vox-research-events — conversion seams at research_mesh.rs and MCP tools.

4. Empty query traces still score Novel — FIXED June 2026

   • novelty.rs gates InsufficientEvidence when all traces fail, but bundles with no traces and empty hits remain Novel.
   • Risk: offline runs appear novel when retrieval never executed.

4.2 High — pipeline fragmentation

5. Deep research and SCIENTIA are parallel, not unified — FIXED June 2026 (discovery_bridge.rs persists finding candidates; mesh emission + session telemetry)

   • run_research persists sessions but does not emit FindingCandidateV1 / DiscoverySignal into SCIENTIA mesh (deep-research roadmap §6).
   • Largest researcher time sink: manual re-entry into publication manifests.

6. Planner silent fallback — FIXED June 2026 (planner_degraded on plan + metadata JSON)

   • planner.rs passthroughs on cascade failure; CRAG multi-hop then runs on a single query.

7. Tavily /research and /extract unwired — FIXED June 2026 (tavily_extract.rs, tavily_research.rs behind env flags)

   • Only Tier-4 /search in web_dispatcher.rs. See tavily-integration-ssot.md — do not duplicate endpoint tables here.

8. Graphify not in retrieval bundle — PARTIAL June 2026 (P1 lexical leg in scientia_prior_art.rs; full MCP query tool still open)

   • P0 status only; no vox_graphify_search MCP tool (graphify-integration-research-2026-06-16.md).

9. Novelty assess path vs raw scorer — FIXED June 2026 (arch test guards; CLI/GUI/MCP route through assess_novelty)

   • ChronoFilter and EvidenceConflictDetector are wired in assess_novelty(), but not all CLI/GUI paths may call assess_novelty vs raw AtomicNoveltyScorer::score — verify at each callsite before claiming full temporal/conflict coverage.

4.3 Medium — UX and ops

10. GUI covers ~10% of scientia CLI — vox-gui-scientia-coverage-audit-2026.md.
11. Gamification = Console discovery exposure only — PARTIAL June 2026 (research_session_complete kudos when gamify enabled; publication never gated)
12. Prereg documentation drift — see §8.
13. graphify-out/ collision with CI artifacts blocks trustworthy P1 ingest.
14. Pipeline cache always miss — FIXED June 2026 (store + read both use knowledge_nodes via upsert_knowledge_node / list_memories_by_type) — pipeline_cache.rs.

4.4 Low

15. Production nanopub network publish intentionally unimplemented (correct posture).
16. OpenRouter spend HUD tile contract exists; GUI tile incomplete.
17. Tantivy behind heavy-retrieval — lean builds lack local lexical corpus.

4.5 Positive (preserve)

• Human-gated review + ApprovalToken in review_flow.rs.
• InsufficientEvidence verdict + empty bundle with None scores (June 2026 fixes).
• Contradicted claim verdict reachable via MiniCheck contradiction path.
• Web gather + CRAG multi-hop shipped; CLI + MCP research surfaces exist.
• Embedder seam + code-uniqueness producer scaffold for automated discovery.

---

5. Integration matrix

Integration	Shipped	Planned / stub	SSOT
OpenRouter	vox_actor_runtime::llm, routing policy, live catalog, research synthesis Lane G	Unified model-routing.v1.yaml; GUI spend tile	model-orchestration-ssot-audit-2026.md
Tavily	/search via TavilySearchClient; budget + fail-open	/extract, /research, /crawl	tavily-integration-ssot.md
Tantivy	lexical_tantivy.rs behind feature	Default-on for research builds; 0.22→0.26 bump	search-retrieval-ssot-2026.md
Vox Search	CRAG, web tier, retrieval bundle, run_multi_hop_web_research	Wikipedia/arXiv/S2 dedicated tiers	Same
Graphify	Registry + vox graphify status / MCP vox_graphify_status	P1–P3 search + MCP query tools	graphify-integration-research-2026-06-16.md

All LLM calls must stay on vox_actor_runtime::llm per root AGENTS.md — no direct vendor SDKs at callsites.

---

6. Researcher time-savings backlog (ranked)

Priority = estimated hours saved per active researcher per month divided by implementation cost (S/M/L).

Rank	Gap	Est. time saved	Cost	Track
1	Unified research → SCIENTIA mesh emission	4–8 h/mo (no manual re-entry)	M	F
2	Route vox research run claims through vox-scientia extractor	2–4 h/mo (consistent verdicts)	S	F
3	Mandatory embedder on novelty-fetch	1–3 h/mo (fewer false novelty decisions)	S	A
4	Tavily /extract for weak snippets	1–2 h/mo (less manual page reading)	M	Phase 3
5	Graphify P1 in retrieval bundle	1–2 h/mo (structural codebase answers)	L	Phase 3
6	Novelty eval harness (precision/recall)	Prevents costly false publishes	M	A
7	Collated results dashboard (session → archive)	1 h/mo (status visibility)	M	Phase 4
8	IMRaD manuscript scaffold from approved claims	2–5 h per publication	L	Phase 4
9	Optional gamify kudos for research milestones	Engagement; not blocking	S	Phase 5
10	Pipeline cache for identical queries	API cost; marginal wall-clock	M	Ops

---

7. Unified pipeline proposal

7.1 Spontaneous research during coding

When Socrates policy or explicit vox research run triggers:

1. Plan — OpenRouter cascade in planner.rs (must surface failure to user, not silent passthrough).
2. Retrieve — vox-search CRAG loop; Tavily when policy enables; future Tantivy for repo-local legs.
3. Ground — vox_scientia::claim_extractor on synthesis + evidence passages.
4. Assess — assess_novelty() with embedder + chrono/conflict on candidate claims.
5. Surface — emit FindingCandidateV1 to mesh → GUI DiscoveryInbox + WS scientia.discovery.surfaced.
6. Human gate — record_claim_review() → ApprovalToken.
7. Emit — user-selected rigor: nanopub build, Zenodo archive run, manuscript draft, syndication.

7.2 Collated results

One publication session (publication_session_id from review_flow.rs) should aggregate:

• Research session IDs (scientia_research_sessions)
• NoveltyEvidenceBundle artifacts
• Claim verdicts
• Review decisions
• Archive status (Zenodo / SWH / nanopub test server)

Extend ScientiaDashboard.tsx rather than adding a parallel surface.

7.3 Optional gamification

When [gamify] is enabled:

• Emit mesh kudos (not XP inflation) for research_session_complete, novelty_assessed, publication_approved.
• Reuse Console discovery ledger pattern (vox.scientia.review action id) — surface streaks in StatusBar only; never gate publication on gamify state.

---

8. Documentation drift: preregistration

Three sources disagree on prereg location:

Source	Claim
scientia-self-publication-gap-map-2026.md §0	crates/vox-prereg/ complete
where-things-live.md	Points to crates/vox-prereg/
vox-scientia/src/lib.rs	Comment: prereg still planned in this crate
Workspace reality (2026-06-16)	No vox-prereg crate in root Cargo.toml; implementation lives under vox-orchestrator/src/preregistration/ + scientia_prereg DDL in vox-db

Follow-up: Update gap map and where-things-live.md in a dedicated doc-hygiene PR; wire orchestrator prereg gate to SCIENTIA publication CLI where still missing.

---

9. Phased extension roadmap

Align with 2026-06-12-scientia-research-pipeline-upgrade.md Tracks A–E. Add Track F: Unified research loop.

Phase	Focus	Key deliverables
0	Doc hygiene (this file) + prereg drift ticket	Research index link; gap map correction PR
1	Track A remainder	Mandatory embedder; bundle SSOT unification; novelty eval harness; callsite audit for assess_novelty
2	Track F (highest ROI)	Mesh emission from run_research; scientia claim extractor in shim; discovery inbox auto-fill
3	Retrieval depth	Tavily extract/research eval; Tantivy corpus row; Graphify P1
4	Beyond nanopub	IMRaD scaffold; collated dashboard; Provider Atlas → model registry
5	Gamify optional	Kudos events; StatusBar streaks

Explicitly out of scope: production nanopub network publish; live arXiv API submission; new shell glue scripts (VoxScript-first policy).

---

10. Verification commands

Check	Command
Research pipeline compiles	cargo check -p vox-research-shim -p vox-scientia -p vox-search
Novelty unit tests	cargo test -p vox-scientia inspect_bridge::novelty
Prior art tests	cargo test -p vox-publisher scientia_prior_art
Novelty golden harness	cargo test -p vox-publisher novelty_golden_harness
Research CLI preview	vox research preview "test query"
Graphify status	vox graphify status --json
Doc frontmatter	cargo run -p vox-doc-pipeline -- --lint-only --paths docs/src/architecture/scientia-automated-research-historical-extension-research-2026.md

---

11. Related SSOT links

• Deep Research Prior Art and Vox Roadmap (2026)
• SCIENTIA micropublication SSOT and surfacing design
• SCIENTIA self-publication gap map
• SCIENTIA × mesh integration research
• Graphify integration research (2026-06-16)
• Search & retrieval SSOT (2026)
• Tavily integration SSOT
• Scientia research pipeline upgrade plan (2026-06-12)
• Where things live — SCIENTIA rows

---

Appendix A — Tavily research request metadata

• Request ID: 2339f716-ebaf-4b7a-977b-5783ee7347a0
• Date: 2026-06-16
• Model: mini
• Response time: ~60s
• Method: MCP tavily_research (CLI tvly not on PATH)