SCIENTIA Automatic Publication & Deep Research: Claims vs. Reality Audit & Power Architecture (2026)

This document audits the Vox Deep Research (vox-research-shim, vox-search, vox-cli-research, Axis ResearchView.tsx) and SCIENTIA Automatic Publication (vox-scientia-*, vox-publisher, scientia_phase_handlers.rs, Axis ScientiaSurface.tsx / PublicationsView.tsx) pipelines against prior architectural claims. It documents which historical gaps have already been closed in tree, isolates the remaining ten concrete illusions across backend and GUI surfaces, and specifies the three-phase unified milestone (Phases 52–54: RES-01..04, PUB-01..04, SCIGUI-01..04) to deliver realistic end-to-end power, live stage and cost observability, and cohesive visual flair.

1. Executive Summary & Correction of Stale 2026 Audits

Earlier audits (deep-research-implementation-divergence-audit-2026-08-01.md and scientia-self-publication-gap-map-2026.md) flagged several issues that have since been fixed in the codebase, while missing deeper structural illusions in the research-to-publication handoff, Architecture SSOT publisher, manuscript/LaTeX scaffolders, and Axis GUI telemetry.

1.1 What Actually Shipped Since the August 2026 Audits (Verified in Tree)

Subsystem	Prior Audit Claim	Current Codebase Reality (2026-10-14)
GUI Claim Verification	“ResearchView.tsx omits verifyClaims: true, bypassing claim verification.”	Fixed at crates/vox-gui/ui/src/components/surfaces/Research/ResearchView.tsx:407: startResearchAsync({ query: activeQuery, verifyClaims: true, lane: activeLane }).
GUI Report Rendering	“ResearchView.tsx renders Markdown inside a plain <pre> tag.”	Fixed at ResearchView.tsx:792 via ResearchReportMarkdown.tsx with clickable [N] citation pills and claimVerdicts.
CLI Eval Harness	“vox research eval calls execute_search_plan instead of the multi-hop pipeline.”	Fixed at crates/vox-cli-research/src/eval.rs:54-56: calls vox_research_shim::research::run_research_with_context.
Domain Research Engines (Shopping & CodeGen)	“Planned in roadmap.”	Shipped in backend at crates/vox-research-shim/src/research/domain/{shopping,codegen}.rs with Wave 2+ rustc/cargo check compilation self-correction in pipeline.rs:677-763 — but unreachable from the GUI (ResearchView.tsx has no domain_mode, scope, waves, or site_scope controls).
Discovery Scout & Archival Run	“Missing vox scientia scout and publication-archive-run.”	Shipped at crates/vox-cli/src/commands/scout.rs (vox scientia scout) and crates/vox-cli/src/commands/db/publication/archive_run.rs (Zenodo + OpenReview adapters).

---

2. Claims vs. Reality Matrix: The 10 Concrete Gaps Holding Back Realistic Power

Diagram source

graph {
    direction: right
    container DR: "Deep Research Pipeline (vox-research-shim)" {
        node Q: "ResearchQuery (General | Shopping | CodeGen)"
        node W: "Multi-Wave Gather & Verify"
        node DB: "VoxDb: research_sessions + research_Session_events" { shape: cylinder }
        node CAND: "discovery_bridge: scientia_finding_candidates"
        Q -> W
        W -> DB
        W -> CAND
    }
    container ILLUSIONS: "Current Broken Handoffs (Gaps 1–5)" {
        node DOC: "docs/src/architecture/*.md"
        node INBOX: "scientia_discovery_inbox"
        node LATEX: "IMRaD / LaTeX / arXiv / FindingPage"
        node CLAIMS: "scientia_claims"
        node COST: "Dashboard & Cost ($0.00)"
        INBOX -> LATEX: "Gap 3: Manual --scaffold X.json required on disk"
        LATEX -> CLAIMS: "Gap 4: MiniCheckBackend::Mock word-overlap default"
    }
    DR.DB -> ILLUSIONS.DOC: "Gap 1: render_doc_draft ignores DB & hardcodes empty vec![]"
    DR.CAND -> ILLUSIONS.INBOX: "Gap 2: finding-* ID written as publication_id (no manifest row!)"
    DR.W -> ILLUSIONS.COST: "Gap 5: UsageSnapshot not wired to scientia_cost_telemetry"
}

2.1 Automatic Publication Pipeline (vox-scientia-* / vox-publisher / vox-cli)

1. Gap 1 — “Publish Architecture SSOT” (render_doc_draft) is a Hardcoded Stub (crates/vox-orchestrator-mcp/src/gui_ops/research.rs:199-239):
   • Claim: Clicking "Publish Architecture SSOT" in ResearchView.tsx (DocPublishModal.tsx) synthesizes a repository-compliant docs/src/architecture/<slug>-research-2026.md SSOT from the completed research session’s verified claims, codebase references, citations, and wave stability.
   • Reality: render_doc_draft never queries VoxDb for session_id! It constructs a dummy DocSynthesizerInput with stability_score: 0.90, executive_summary: "Empirical investigation findings.", and empty vec![] for verified_claims, codebase_refs, competitive_matrix, sandbox_probes, gaps_and_recommendations, and roadmap_phases. Furthermore, the template in crates/vox-research-shim/src/research/doc_synthesizer.rs:41-52 omits required topic, domain, authorship, authorship_model, and <!-- kb:summary --> metadata, so any file written by write_doc fails vox-doc-pipeline --lint-only.
2. Gap 2 — Research → Publication Handoff Drops Candidate Class & Breaks Inbox Review (crates/vox-research-shim/src/research/discovery_bridge.rs:82-98, 226-231):
   • Claim: High-signal Deep Research runs automatically populate the SCIENTIA Discovery Inbox and can be reviewed and archived from the GUI.
   • Reality:
     • finding_candidate_from_research_result hardcodes candidate_class: EventCandidateClass::Other, title_hint: None, and novelty_evidence_bundle_id: None.
     • surface_research_discovery_inbox passes candidate.candidate_id ("finding-<uuid>") into the publication_id column of scientia_discovery_inbox, even though no publication_manifests row exists for that ID. Clicking "Open review" or "Archive" in DiscoveryInbox.tsx queries publication_manifests for finding-<uuid> and fails with "not_found". There is no 1-click/1-command “Promote Candidate / Research Session to Publication Manifest” bridge.
3. Gap 3 — Manuscript, LaTeX, arXiv Bundle, Replay & Finding Page Require Manual Offline JSON Files (crates/vox-cli/src/commands/scientia_phase_handlers.rs:32-66, 727-890):
   • Claim: SCIENTIA automatically drafts IMRaD manuscripts, renders venue-compliant LaTeX (neurips_2025, icml2025, acmart, jmlr2e), exports arXiv .tar.gz bundles, executes artifact replay, and renders public HTML finding pages.
   • Reality:
     • publication-manuscript-draft --scaffold X.json, publication-render-latex --scaffold X.json, and publication-arxiv-bundle --scaffold X.json require a hand-authored ScaffoldInput JSON file on disk. There is no assemble_scaffold_from_db(db, publication_id) thathydrates ScaffoldInput from publication_manifests + scientia_claims + novelty_evidence_bundle + ro_crate + prereg automatically.
     • publication-replay-execute (scientia_phase_handlers.rs:727-796) only appends an event "artifact_replayability_measured" and never updates metadata_json.scientia_evidence or WorthinessInputs::artifact_replayability.
     • publication-finding-page-render --page X.json requires a manual FindingPageInput JSON file instead of hydrating from publication_id.
4. Gap 4 — Claim Extraction Defaults to Trivial Word-Overlap Mock (crates/vox-cli/src/commands/scientia_phase_handlers.rs:304, 332-336):
   • Claim: publication-extract-claims performs NLI / MiniCheck grounding verification over extracted claims.
   • Reality: When VOX_MINICHECK_ENDPOINT is unset (the default local setup), ExtractionPipeline falls back to MiniCheckBackend::Mock (crates/vox-scientia-claims/src/minicheck.rs), which computes naive token overlap instead of routing through the workspace’s model-agnostic LLM boundary (vox_actor_runtime::llm::infer_with_retry).
5. Gap 5 — Dashboard Confidence & Cost Rollup Blind Spots (scientia_phase_handlers.rs:230, 391, 577-605):
   • Claim: vox scientia dashboard and vox scientia cost provide real-time candidate confidence, reply-window tracking, and per-stage cost attribution (extraction, novelty, critic, scholarly).
   • Reality:
     • scientia_dashboard() only queries publication_manifests with hardcoded confidence: 0.0, ignoring scientia_finding_candidates rows and hardcoding manifests_in_reply_window: Vec::new().
     • scientia_cost() records 0.0_f64 for extraction and critic and 0.0 for novelty and scholarly, never reading vox_research_shim::research::metering::snapshot() or LLM token telemetry.

2.2 Deep Research Pipeline & Axis GUI Observability / Visual Flair (vox-research-shim / vox-gui)

6. Gap 6 — ResearchView.tsx Hides Shopping, CodeGen, Scope, Waves, and Site Power Controls:
   • ResearchQuery supports domain_mode: ResearchDomainMode::{General, Shopping, CodeGen}, scope: ResearchScope::{Web, Local, Both}, waves: u8 (1..=5), and site_scope: Option<String>, plus verify_Compilation in Wave 2+ for CodeGen and weighted decision matrices for Shopping. None of these controls exist in ResearchView.tsx.
7. Gap 7 — Static Pipeline Timeline & Orphaned MultiWaveProgressTimeline.tsx:
   • During an active run, ResearchView.tsx:596 renders <PipelineTimeline stages={RESEARCH_STAGES} statuses={deriveStages('active')} />, which freezes stage 0 (Query Decomposition) as 'active' for the entire run instead of reflecting live session.status (planning → retrieving → verifying_claims → auditing_citations → synthesizing → persisting_artifact → completed) and research_session_events.
   • crates/vox-gui/ui/src/components/surfaces/Research/MultiWaveProgressTimeline.tsx (with wave stability bars, contradiction counts, and sandbox check badges) is implemented and unit-tested in tree, but is never imported or mounted anywhere in ResearchView.tsx.
8. Gap 8 — Synthetic Linear Epistemic DAG (ResearchView.tsx:762-778):
   • dagNodes and dagEdges in ResearchView.tsx are fabricated client-side by assigning wave: Math.floor(i / 2) and connecting dagNodes[i] -> dagNodes[i + 1] in a single chain (chain-${i}), rather than rendering the true hierarchical epistemic graph (Query -> Subquery (Wave N) -> Source Domain -> Extracted Claim -> Verification Verdict / Contradiction Edge).
9. Gap 9 — Verifier Resampling Ignores Fast Lexical/NLI Pre-Filter & abstain_threshold (crates/vox-research-shim/src/research/verifier.rs:15-26):
   • VerifierConfig.abstain_threshold is marked Phase 0a — fields are placeholders and ignored; verify_claims_With_resampling runs 3 LLM calls on every claim without a fast lexical/BM25/token-containment pre-filter or calibrated abstention band.
10. Gap 10 — Fragmented Scientia / Discovery / Publications GUI Surfaces:
    • ScientiaSurface.tsx (scientia view) only shows Dashboard | Claims | Harness Issues.
    • DiscoverySurface.tsx (activity view) buries Inbox | Review | Archive inside the generic Activity timeline.
    • PublicationsView.tsx (publications view) is a read-only Kanban board that invokes slow CLI subprocesses (execute_command running vox scientia ...) instead of typed /api/v2/scientia/* daemon routes, with zero buttons to promote candidates, extract claims, run novelty/critic gates, preview IMRaD/LaTeX manuscripts, or trigger Zenodo/OpenReview archival.

---

3. Target Architecture: Realistic Power, End-to-End Observability & Visual Flair

Diagram source

graph {
    container P52: "Phase 52: Deep Research Realistic Power & Live Observability (RES-01..04)" {
        node R1: "RES-01: DB-Grounded DocSynthesizer (kb:summary + taxonomy frontmatter + real claims/citations)"
        node R2: "RES-02: Full ResearchQuery API (domain_mode, scope, waves, site_scope) + Epistemic Graph Builder"
        node R3: "RES-03: Live Stage/Wave Telemetry Stream + UsageSnapshot -> scientia_cost_telemetry"
        node R4: "RES-04: Calibrated Verifier (Lexical/Containment Pre-Filter + abstain_threshold Enforcement)"
    }
    container P53: "Phase 53: SCIENTIA Automatic Publication Closure (PUB-01..04)" {
        node P1: "PUB-01: Candidate Classifier + 1-Click Candidate/Session -> publication_manifests Promotion"
        node P2: "PUB-02: DB-Backed ScaffoldInput & FindingPage Assembler (Zero Manual JSON Files)"
        node P3: "PUB-03: LLM-Backed Claim Grounding Verifier (vox_actor_runtime::llm) + Measured Replay Writeback"
        node P4: "PUB-04: Unified Dashboard (Candidates + Manifests + Reply Window) & Real Cost Rollup"
    }
    container P54: "Phase 54: Unified Research & Scientia Studio GUI (SCIGUI-01..04)" {
        node G1: "SCIGUI-01: Research Power Bar (General|Shopping|CodeGen, Scope, Waves, Site) + Mode Cards"
        node G2: "SCIGUI-02: Live Stage Telemetry Ribbon + Mounted MultiWaveProgressTimeline + Cost Meter"
        node G3: "SCIGUI-03: True Multi-Layer Epistemic DAG Canvas (Query->Subquery->Source->Claim->Verdict)"
        node G4: "SCIGUI-04: Unified Scientia & Publications Workbench (1-Click Promote, Worthiness Radar, Manuscript/LaTeX Preview)"
    }
    P52.R4 -> P53.P1
    P53.P4 -> P54.G1
}

3.1 Phase 52 (RES-01..04 — Wave 2 Backend): Deep Research Realistic Power, DB-Grounded Architecture SSOT Synthesis & Live Stage Observability

• RES-01 (DB-Grounded Architecture SSOT Synthesizer):
  • Upgrade crates/vox-research-shim/src/research/doc_synthesizer.rs to emit repository-compliant YAML frontmatter (title, description, category: "Architecture SSOTs", status: "current", topic: "scientia-research", domain: "knowledge", authorship: "llm-assisted", authorship_model) and a 40–120 word <!-- kb:summary --> block directly under # <Title>.
  • Upgrade render_doc_draft in crates/vox-orchestrator-mcp/src/gui_ops/research.rs to load the real research_sessions row (query_text, report_markdown, artifact_json, config_json) and research_session_events from VoxDb, populating verified_claims, codebase_refs, competitive_matrix (when Shopping mode or comparison tables exist), sandbox_probes (when CodeGen compilation probes exist), gaps_and_recommendations, and wave_stability.
• RES-02 (Full ResearchQuery Control Surface & True Epistemic Graph Serializer):
  • Extend StartResearchBody in crates/vox-orchestrator-mcp/src/gui_ops/research.rs and crates/vox-gui/src/commands/scientia.rs to accept domain_mode ("general" | "shopping" | "codegen"), scope ("web" | "local" | "both"), waves (1..=5), site_scope, and verify_compilation.
  • Add a deterministic build_epistemic_dag(session, events) helper in gui_ops/research.rs that emits typed nodes (query, subquery, source_domain, claim, sandbox_probe) and typed edges (decomposes_to, retrieves_from, grounds_claim, contradicts, verifies_compile) on get_session.
• RES-03 (Live Stage/Wave Event Telemetry & Cost Metering Persistence):
  • Emit granular research_session_events (wave_started, subquery_executed, claim_verified, contradiction_detected, sandbox_check_completed, wave_completed with stability_score, entropy_delta, and UsageSnapshot) and wire vox_research_shim::research::metering::snapshot() into VoxDb::record_scientia_cost.
• RES-04 (Calibrated Claim Verifier with Fast Containment Pre-Filter & abstain_threshold):
  • Remove the PHASE_0a_STUB placeholders in crates/vox-research-shim/src/research/verifier.rs: add a fast lexical/n-gram containment pre-filter that immediately confirms verbatim/high-overlap citations without 3x LLM resampling, routes ambiguous claims through vox_actor_runtime::llm, and enforces VerifierConfig.abstain_threshold to classify low-margin results as ClaimLabel::Abstain rather than silently coercing to Contradicted.

3.2 Phase 53 (PUB-01..04 — Wave 2 Backend): SCIENTIA Automatic Publication Pipeline Closure

• PUB-01 (Classified Candidate Handoff & 1-Click Candidate/Session → Publication Manifest Promotion):
  • Upgrade finding_candidate_from_research_result in crates/vox-research-shim/src/research/discovery_bridge.rs to infer EventCandidateClass (BenchmarkDelta, ArchitectureDecision, ReproducibilityAudit, CorpusFinding, etc.), title_hint, and novelty_evidence_bundle_id from the research artifact.
  • Add promote_candidate_to_publication(db, candidate_id_or_session_id, venue) in crates/vox-orchestrator-mcp/src/gui_ops/scientia.rs (and CLI vox scientia candidate-promote --id <id>) that creates a canonical publication_manifests row, links scientia_discovery_inbox.publication_id to the real manifest, and seeds initial scientia_claims from the research session’s verified claims.
• PUB-02 (DB-Backed ScaffoldInput & FindingPageInput Assembler — Zero Manual JSON Files):
  • Add assemble_scaffold_from_db(db, publication_id, venue) and assemble_finding_page_from_db(db, publication_id) so vox scientia publication-manuscript-draft, publication-render-latex, publication-arxiv-bundle, and publication-finding-page-render work directly from --publication-id <id> (with --scaffold kept only as an optional override) and are exposed over /api/v2/scientia/publications/:id/{manuscript,latex,arxiv-bundle,finding-page}.
• PUB-03 (LLM-Backed Claim Grounding Verifier & Measured Replay Writeback):
  • Upgrade crates/vox-scientia-claims/src/minicheck.rs & scientia_phase_handlers.rs so when VOX_MINICHECK_ENDPOINT is unset, claim verification routes through vox_actor_runtime::llm (structured entailment prompt) with MiniCheckBackend::Mock restricted to deterministic unit tests.
  • Update publication_replay_execute (scientia_phase_handlers.rs:727-796) to write the measured replayability score (1.0 on exit 0 + digest match, 0.5 on exit 0 without digest, 0.0 on failure) back into publication_manifests.metadata_json.scientia_evidence and recompute WorthinessInputs::artifact_replayability.
• PUB-04 (Unified Scientia Dashboard & Real Cost Attribution):
  • Upgrade scientia_dashboard and /api/v2/scientia/dashboard to merge scientia_finding_candidates (with real confidence_json and candidate_class), publication_manifests, and active reply-window items (openreview / arxiv manifests with pending comment windows).
  • Wire real UsageSnapshot / token-cost telemetry into scientia_cost_telemetry across extraction, novelty, critic, and scholarly stages.

3.3 Phase 54 (SCIGUI-01..04 — Wave 3 GUI Slot 17): Unified Research & Scientia Publication Studio GUI

• SCIGUI-01 (Deep Research Power Controls & Domain Mode Cards in ResearchView.tsx):
  • Add tasteful, compact power controls to the ResearchView.tsx query bar: Domain Mode selector (General | Shopping Matrix | CodeGen Verified), Scope selector (Web | Local Repo | Hybrid Both), Wave budget stepper (1–3 Waves), and optional site: filter pill.
  • Render dedicated rich cards when domain_mode === 'shopping' (weighted criteria & scored options table from ShoppingDecisionMatrix) or domain_mode === 'codegen' (sandbox cargo check/rustc compilation badge + repair iteration log).
• SCIGUI-02 (Live Stage Telemetry Ribbon, Mounted MultiWaveProgressTimeline.tsx & Cost/Latency Pill):
  • Poll/stream getSession(sessionId) during active research runs so PipelineTimeline highlights the true live backend stage (planning → retrieving → verifying_claims → auditing_citations → synthesizing → persisting_artifact → completed).
  • Mount MultiWaveProgressTimeline.tsx inside the active/completed ResearchView.tsx workspace and display a live telemetry header ribbon (Wave N/M · Stability Δ · Verified Claims X/Y · Cost $Z).
• SCIGUI-03 (Interactive Multi-Layer Epistemic DAG Canvas in ResearchDagCanvas.tsx):
  • Upgrade ResearchDagCanvas.tsx to consume the backend’s true epistemic graph (Query -> Subquery -> Source Domain -> Claim -> Verdict / Contradiction Edge), with distinct visual encodings for Supported (emerald pulse), Contradicted (amber/crimson dashed conflict edge), Abstain (muted slate), and CodeGen Probe (cyan badge), plus click-to-inspect node detail drawer and "Promote to Scientia Publication" CTA along the header.
• SCIGUI-04 (Unified Scientia & Publications Studio Workbench):
  • Unify ScientiaSurface.tsx, DiscoverySurface.tsx, and PublicationsView.tsx into a cohesive studio backed by typed /api/v2/scientia/* client methods (crates/vox-gui/ui/src/api/scientia.ts):
    • Add Discovery Inbox, Worthiness & Review, and Manuscript & Archive Studio tabs directly inside ScientiaSurface.tsx.
    • Add a 1-click “Promote to Publication Manifest” button on every DiscoveryInbox.tsx candidate and completed ResearchView.tsx session.
    • Add interactive stage progression controls on PublicationsView.tsx and DiscoveryReview.tsx (Extract & Verify Claims, Compute Novelty Bundle, Run Critic Gate, Preview IMRaD / LaTeX, Run Artifact Replay, and Archive to Zenodo / OpenReview) with a visual 6-axis Worthiness Radar / Meter (Epistemic, Novelty, Replayability, Signal-to-Noise, Venue Fit, Claim Grounding).