GrantGuardian: Multi-Agent Citation Integrity SaaS

Multi-agent SaaS that detects 2nd-order citation rot in grant proposals, quarantines retracted research, and enforces deterministic human-in-the-loop governance with cryptographic audit trails.

Comment

Inspiration

The $100M+ Silent Research Integrity Crisis

In high-stakes academic and industrial R&D, submitting federal grant proposals to agencies like the NIH, NSF, DARPA, and the European Research Council (ERC) is a multi-billion-dollar enterprise. A single federal grant award typically represents $1.5M to $10M in institutional funding.

Yet an invisible compliance liability lurks within scientific literature: Citation Rot & Retraction Contamination.

🚨 The Retraction Tsunami by the Numbers

  • >10,000 Scientific papers retracted in 2023 alone (Nature 624, 479–481)
  • >55,000 Discredited manuscripts cataloged in Retraction Watch
  • 94% of citations to retracted papers continue to cite them as valid science without acknowledging their retraction (Schneider et al., PNAS)
  • $100M+ in federal grant clawbacks, compliance investigations, and institutional penalties

When a Principal Investigator (PI) cites a seemingly clean, landmark paper in their proposal's Specific Aims or Methodology, that paper may quietly depend on fraudulent, fabricated, or contaminated foundation experiments retracted years prior. This is 2nd-Order Latent Citation Contamination.

Why Conventional AI Fails Here

Most current "AI research assistants" are thin prompt wrappers around single LLM calls. When asked to review literature, generic LLMs:

  1. Hallucinate Retractions: Speculating on retraction status based on internet discourse rather than publisher errata.
  2. Commit Uncontrolled Destructive Actions: Auto-deleting citations that only mentioned a retracted paper to critique or dispute it.
  3. Suffer Alert Fatigue: Generating unverified walls of text with zero cryptographic audit trail for institutional compliance officers.

The world doesn't need another chat demo. We built GrantGuardian as a production-grade, multi-agent SaaS platform that traverses multi-hop citation graphs across 4 global registries, calculates mathematical contamination vectors, and enforces deterministic human-in-the-loop governance.

What it does

GrantGuardian is an autonomous research-operations and citation integrity command center for Principal Investigators, research universities, and enterprise labs.

1. Autonomous 4-Way Registry Consensus

Rather than trusting a single search engine, GrantGuardian queries 4 authoritative global literature databases in parallel:

  • Crossref REST API: Official publisher records, Crossmark metadata, errata, and corrigenda notices.
  • Retraction Watch Database: Comprehensive registry of editorial sanctions, institutional inquiry decrees, and specific fraud/data fabrication reasons.
  • PubMed Central (NLM MeSH): NIH-curated biomedical database verifying public health and clinical trial integrity.
  • OpenAlex Graph: 250M+ scholarly works graph providing cross-field consensus, citation trajectories, and institutional affiliations.

2. 2nd-Order Latent Citation Traversal (The Cascade Detector)

Direct citations in a grant proposal often appear clean. GrantGuardian recursively extracts the 1-hop outgoing reference graph of every tracked citation via Semantic Scholar. It inspects all child references against international retraction registries:

  • Direct Contamination (1-Hop): The proposal directly cites a retracted paper.
  • Latent Contamination (2-Hop): Paper A (cited by the proposal) heavily relies on Paper B's experimental methodology, but Paper B was formally retracted. GrantGuardian isolates this hidden vulnerability before grant reviewers discover it.

3. Contamination Blast-Radius & CSI Calculus

When retracted foundation science is identified, GrantGuardian maps its structural impact across the grant proposal:

  • Specific Aims Blast Radius: Evaluates whether Core Aim 1 or 2 is invalidated by the retracted methodology.
  • Budgetary Risk Mapping: Assesses the dollar liability tied to experiments built on compromised premises.
  • Quantitative Scoring: Assigns a deterministic Contamination Severity Index ($CSI$) from 0.00 to 1.00 based on graph depth and section criticality.

4. Deterministic Safety Policy (classifyDecision)

To guarantee scientific reliability, the platform enforces hard-coded invariant gates:

  • Direct Retractions: Automatically quarantined from active grant proposal bibliographies with zero model speculation.
  • 2nd-Order Dependencies: AI is mathematically forbidden from deciding if a downstream paper is still valid. The case is packaged with root DOIs, retraction reasons, and impact analysis, and routed directly to the PI's Human Decision Inbox.
  • Clean Literature: Operates under a "Quiet-by-Default" philosophy. Clean literature logs silent heartbeat records without generating annoying notification spam.

5. NIST SP 800-92 Cryptographic Audit Trail

For every sweep, GrantGuardian generates an immutable HMAC-SHA256 provenance receipt. This cryptographic digest anchors the full 7-step inspection chain (registries queried, raw timestamps, model reasoning steps, and human decisions), giving university compliance officers tamper-evident proof of due diligence.

6. Multi-Tenant Lab Command Center

Provides dedicated workspace isolation across lab environments, featuring:

  • Executive PI Dashboard: High-level health score, active grant deadlines, and pending decision items.
  • Interactive Citation Graph Visualizer: Color-coded node-link graph distinguishing verified, quarantined, and review-pending citations.
  • Instant Demo Evaluation Workspaces: Pre-seeded with real-world research personas (Materials Science, Computational Oncology, Neurobiology) plus an unseeded blank workspace for live testing.

How we built it

GrantGuardian is engineered on a decoupled, production-tier SaaS architecture separating user experience, API orchestration, multi-agent reasoning, and deterministic safety.

System Architecture & Decision Flow

Grant Guardian: System Architecture & Decision Flow

The 4-Tier Architecture

  1. Tier 1: Client Application (React 18 + TypeScript + Vite + TailwindCSS): Executive PI dashboard, interactive D3 citation graph, Human Decision Inbox, and real-time sweep telemetry.
  2. Tier 2: API Gateway & Orchestrator (Node.js + Express + Drizzle ORM): Multi-tenant user context resolver, autonomous morning sweep cron scheduler, and circuit-breaker database fallbacks.
  3. Tier 3: Multi-Agent Reasoning Fleet (Python 3.11 + FastAPI + Claude 3.5 Sonnet): Sovereign Coordinator orchestrating specialized Citation Integrity and Governance subagents via the Agents-as-Tools delegation pattern.
  4. Tier 4: Deterministic Safety Boundary (classifyDecision): Mathematical guardrails in TypeScript and Python enforcing immutable scientific safety rules.

The 10 Specialized Tools (Across 4 Scientific Registries)

GrantGuardian decomposes complex workflows across two dedicated subagents under the Sovereign Coordinator:

Agent / Fleet Tool Name Registry / Protocol Downstream Forensic Action
Citation Integrity crossref_lookup Crossref REST API Resolves official publisher notices, errata, corrigenda, and Crossmark flags
(Literature Forensics) retraction_watch_lookup Retraction Watch DB Verifies institutional inquiry decrees, editorial sanctions, and retraction reasons
openalex_global_registry OpenAlex (250M+ Works) Establishes cross-database consensus, citation velocities, and field normalization
pubmed_retraction_verifier NIH NLM MeSH Archive Corroborates biomedical and clinical validity against NIH public health standards
semantic_scholar_graph Semantic Scholar API Recursively extracts 1-hop outgoing bibliographies for 2nd-order analysis
check_reference_retractions 8x Worker ThreadPool Concurrently audits all extracted child DOIs in parallel to eliminate latency
contamination_vector_calculator Graph Analysis Engine Computes Contamination Severity Index ($CSI$) and project blast radius
Governance Fleet draft_compliance_report Institutional Compiler Assembles preliminary NSF/NIH narrative sections from tracked evidence
(Oversight & Sign-Off) escalate_to_human Structured Dispatcher Routes ambiguous 2nd-order findings to the PI's Decision Inbox
provenance_proof_generator Cryptographic Digest Computes NIST SP 800-92 compliant HMAC-SHA256 tamper-evident proofs

Mathematical & Algorithmic Formulation

1. Contamination Severity Index ($CSI$)

To avoid arbitrary qualitative guesses, GrantGuardian models citation risk with a deterministic scoring function:

$$CSI = \min\left(1.0, \; \sum_{k=1}^{n} w_k \cdot \left( \frac{1}{\text{hop}_k} \right) \cdot \lambda_k \right)$$

Where:

  • $w_k \in [0, 1]$ represents proposal section weight (Specific Aims = $1.0$, Core Methodology = $0.85$, Background Literature = $0.35$).
  • $\text{hop}_k$ represents graph distance ($\text{hop}=1$ for direct retractions, $\text{hop}=2$ for foundation cascades).
  • $\lambda_k$ is the severity coefficient from Retraction Watch taxonomy (Data Fabrication = $1.0$, Unreproducible Protocol = $0.75$, Publisher Error = $0.20$).

2. Cryptographic Provenance Hash

Every forensic decision produces an immutable, non-repudiable audit certificate:

$$H_{\text{provenance}} = \text{HMAC-SHA256}\Big(K_{\text{lab}}, \; \text{DOI} \parallel \text{Timestamp} \parallel \text{ConsensusBits} \parallel \text{TraceTree} \Big)$$

The Verifiable 7-Step Provenance Trail

Every execution follows an explicit, verifiable inspection chain:

  1. Identify Dependency: Receive target DOI and extract metadata schema.
  2. Crossref Registry Query: Verify publisher Crossmark errata and status flags.
  3. Retraction Watch Cross-Check: Check sanctions register for institutional decrees.
  4. Relationship Analysis: Query OpenAlex and PubMed Central for corroborating signals.
  5. 1-Hop Reference Graph Traversal: Concurrently scan all child references via Semantic Scholar.
  6. Deterministic Safety Check: Apply classifyDecision() rule invariants.
  7. Human Escalation or Quarantine: Route ambiguous propagation to the PI or auto-quarantine direct retractions.

Challenges we ran into

  1. Multi-Hop Graph Rate Limiting: Recursively scanning 35+ outgoing references across 4 scientific registries per citation initially took >45 seconds due to external rate limits. We re-engineered the engine with an 8x asynchronous worker pool (ThreadPoolExecutor), localized SQLite caching of confirmed sanctions, and staggered parallel queries, slashing total sweep time to <8.5 seconds.

  2. Preventing LLM Hallucination in Scientific Validity: LLMs tend to over-confidently speculate on whether a downstream study is still "usable" despite a retracted foundation. We solved this by enforcing a hard deterministic code gate (classifyDecision) that halts model execution and forces human-in-the-loop review whenever ambiguity is detected.

  3. Multi-Tenant State Isolation: Academic labs require strict data boundaries. We built an isolated tenant resolver where every database row (citations, activities, drafts, preferences) enforces a foreign key constraint to users.id, ensuring that multiple PIs evaluating proposals simultaneously never leak confidential grant data.

Accomplishments that we're proud of

Benchmark Metric Verified Result Verification Mechanism
Full Sweep Latency <8.5 seconds 8x parallel ThreadPoolExecutor worker pool
Automated Test Suite 90 Passing Tests 66 TypeScript + 24 Python tests
Consensus Rigor 4-Way Multi-Registry Crossref + Retraction Watch + OpenAlex + PubMed
Safety Invariant 100% Deterministic Gate Zero unverified autonomous external actions
Audit Compliance NIST SP 800-92 Cryptographic HMAC-SHA256 non-repudiable digests
Production Status 3 Microservices Live Full deployment on Render with health checks
  • Built a Real SaaS, Not a Prototype: Complete with multi-tenant session isolation, pre-seeded academic personas (Materials Science, Computational Oncology, Neurobiology, Blank Slate), and interactive visual dashboards.
  • 90 Passing Automated Tests: Exhaustively tested across circuit breaker trips, graph traversal depth, HMAC-SHA256 signature verification, and edge-case prompt injection attacks.
  • Production Deployment: Frontend, Node.js API, and Python Agent services are live in production with health checks and zero cold-start crashes.

What we learned

  1. Real-world AI products require deterministic restraint: The most impressive agent is not the one that takes unchecked autonomous actions, but the one with the architectural wisdom to recognize when human expertise is legally and scientifically required.
  2. Domain separation prevents hallucination cascades: Subdividing complex workflows into specialized, single-purpose subagents with strict negative constraints is orders of magnitude more reliable than a monolithic prompt.
  3. The internet desperately needs better scientific AI: Automating administrative busywork is trivial; safeguarding the foundation of human scientific knowledge requires rigorous, verifiable engineering.

What's next for GrantGuardian: Multi-Agent Citation Integrity SaaS

  • [ ] Vector-Grounded Clean Alternative Recommender: When a foundational paper is quarantined, automatically retrieve and suggest verified, peer-reviewed replacement literature using embedding similarity over clean OpenAlex papers.
  • [ ] Direct Grants.gov & NIH eRA Commons Integration: A pre-submission browser extension that audits proposals directly inside university grant submission portals before final sign-off.
  • [ ] University Research Office Enterprise Tier: Centralized compliance monitoring across entire university departments, alerting Research Deans to emerging retraction risks across all active federal awards.
  • [ ] Automated Institutional Review Board (IRB) Audit Reports: One-click generation of PDF compliance packages certified with cryptographic provenance digests.

🌐 Live Production Verification

Built With

Share this project:

Updates

Submission history