Skip to main content

CODITECT Research Continuum: Vision Document

Executive Summary​

CODITECT is building the first Agentic Knowledge Infrastructure — a platform where research papers are not read but operated on by 150+ specialized AI agents that extract, synthesize, reason, and generate new insights at machine speed. We've achieved 100% Grade A extraction quality (0.915 avg) across 218 academic papers using our Universal Document Object Model (UDOM) with full provenance tracking. This isn't a better PDF reader. It's the operating system for computational research, where extraction is the I/O layer and agents are the kernel. The vision: compress a PhD literature review from 6 months to 6 hours, with citations that trace back to the sentence. Target customers: academic research groups, pharmaceutical R&D, law firms, financial analysts, and any organization drowning in unstructured knowledge. Revenue model: usage-based pricing on agent compute + knowledge graph API. The moat: multi-source extraction quality, compounding knowledge graphs with provenance guarantees, and a flywheel where every research query improves the system. We're not disrupting Google Scholar. We're building what comes after human-speed research.

The Insight: Extraction is the Front End for Agentic Research​

The world has spent 30 years building better search engines for documents. Semantic Scholar indexes 200M papers. Elicit uses LLMs to summarize abstracts. Consensus aggregates opinions. They're all solving the same problem: help humans find and read documents faster.

But here's the existential realization: no human will ever read 200 million papers. Not in one lifetime. Not in a hundred.

The bottleneck isn't discovery. It's cognitive bandwidth. A researcher can read 2-3 papers per day. A literature review for a PhD takes 6 months. A pharmaceutical company trying to understand all prior art on a protein target has 10,000 papers to review and 3 months before the patent filing deadline.

The insight: extraction without action is just expensive OCR.

What matters isn't that you can parse a PDF into clean markdown with 0.915 quality. What matters is what happens after extraction. Do you have agents that can:

  • Synthesize contradictory findings across 47 papers on the same topic?
  • Reason about why Method A worked in Paper X but failed in Paper Y?
  • Generate hypotheses by connecting insights from disparate fields?
  • Trace every claim back to the exact sentence, figure, or equation in the source?
  • Propose new experiments based on unexplored parameter spaces?
  • Write a coherent research brief that would take a human 40 hours, in 6 minutes?

This is the shift: from documents as endpoints to documents as inputs for computational research. Extraction is the I/O layer. The real product is the agentic reasoning loop that runs on top.

And here's the compounding insight: every research query improves the system. When an agent synthesizes 47 papers, that synthesis gets written back to the knowledge graph. The next researcher who asks a related question starts from that synthesis, not from zero. Knowledge compounds. This is the flywheel.

UDOM — the Universal Document Object Model — is the foundation. But the vision is the Research Continuum: the end-to-end system where documents flow in, agents operate, insights flow out, and knowledge accumulates.

Product Name & Category​

Product Name: CODITECT Research Continuum

Category: Agentic Knowledge Infrastructure

Alternate Framing: Computational Research Operating System

Why "Research Continuum"?​

The name signals continuity — not discrete document retrieval, but a continuous flow from raw documents to structured knowledge to agentic reasoning to synthesized insights to new hypotheses to recursive research. It's a continuum because the system never stops learning. Every query feeds the next.

Why "Agentic Knowledge Infrastructure"?​

Because this isn't software-as-a-service. It's infrastructure — the computational substrate on which research happens. "Agentic" signals that the system has agency: it doesn't wait for humans to ask questions; it proposes hypotheses, identifies gaps, and suggests next steps.

This is the category we're creating. There is no incumbent. Semantic Scholar is a search engine. Elicit is a summarization tool. We're building the operating system for machine-augmented research.

What It Is: The Full System​

The CODITECT Research Continuum is a closed-loop agentic research platform that ingests unstructured documents, extracts them into a universal semantic representation (UDOM), stores them in a provenance-aware knowledge graph, and enables 150+ specialized AI agents to reason, synthesize, and generate insights at machine speed.

The Stack (Top to Bottom):​

  1. Extraction Layer (UDOM Pipeline v1.7)

    • Multi-source extraction: Docling (PDF), ar5iv (HTML), arXiv (LaTeX)
    • 100% Grade A quality, avg 0.915 across 218 papers
    • 25 component types: sections, equations, figures, citations, code blocks, tables, etc.
    • Full provenance: every component traces back to source page, bounding box, and extraction method
    • 15,478 LOC across 26 Python files

  2. Knowledge Graph (Provenance-Aware Storage)

    • Hybrid semantic + FTS5 search
    • Component-level indexing (not document-level)
    • Citation graphs linking papers at the claim level
    • Temporal evolution tracking (how claims change across paper versions)
    • Multi-tenant architecture with role-based access control

  3. Agent Layer (150+ Specialized Experts)

    • MoE (Mixture of Experts) routing: match task to best agent
    • Vertical specialists: market-researcher, competitive-analyst, trend-analyst, framework-specialist, synthesis-writer
    • Horizontal generalists: senior-architect, devops-engineer, security-specialist
    • Research-to-Artifacts (R2A) v2.0: orchestrates multi-agent workflows (lit review to synthesis to strategy brief)
    • Claude Code CLI bridge (ADR-167): zero-cost LLM access via local Claude instance

  4. Synthesis Layer (Strategy Brief Generator)

    • Pyramidal executive summaries (situation to challenge to recommendation)
    • One-page strategic briefs for time-constrained executives
    • Prioritized recommendations with timelines, owners, success metrics
    • McKinsey/BCG quality gates (word count, actionability, scannability)

  5. Interface Layer (UDOM Navigator)

    • Web-based viewer with 8 tabs: markdown, JSON, images, LaTeX, citations, entities, search, provenance
    • Real-time streaming of agent outputs
    • Diff view for iterative agent improvements
    • Export to markdown, PDF, DOCX, JSON

  6. Orchestration Layer (Research Loop)

    • Recursive research: synthesis to hypothesis to new query to new extraction to new synthesis
    • Gap analysis: agents identify missing data and propose new documents to ingest
    • Convergence detection: stop when new documents don't change the synthesis
    • Audit trail: every decision, every agent invocation, every data source logged

What Makes This a Unified Product (Not a Collection of Tools):​

The magic is in the closed loop. Extraction feeds the knowledge graph. The knowledge graph feeds the agents. The agents generate syntheses. The syntheses identify gaps. The gaps trigger new extractions. Every cycle compounds knowledge.

This isn't a pipeline you run once. It's a living research environment that gets smarter with every query.

The 8-Step Pipeline: UDOM Extraction as the Foundation Layer​

The Universal Document Object Model (UDOM) is the foundation. Here's how it works:

Step 1: Multi-Source Ingestion​

  • PDF: Docling v2.72.0 with Tesseract OCR (primary)
  • HTML: ar5iv.labs.arxiv.org (LaTeX to HTML5 conversion)
  • LaTeX: arXiv e-print source (equations, macros, .bib files)
  • Fallback: pymupdf4llm (if Docling fails) to basic fitz (if pymupdf4llm hangs)

Step 2: Parallel Extraction​

  • Three extractors run simultaneously (PDF, HTML, LaTeX)
  • Each produces a list of UDOM components with metadata
  • Timing: 5-7s for PDF, 2-3s for HTML, 1-2s for LaTeX (total: 10-35s per paper)

Step 3: Component Mapping​

  • Mapper aligns components across sources using fuzzy text matching
  • Coverage score: fraction of component text found in each source
  • Best source selected per component (highest coverage + reliability score)
  • Result: unified component list with multi-source provenance

Step 4: Assembly​

  • Assembler rebuilds document structure from components
  • Deduplication: abstract, title, authors (prefer LaTeX to HTML to PDF)
  • Section hierarchy reconstruction from flat component list
  • Bibliography resolution: .bib to .bbl to inline citations

Step 5: Image Extraction & Storage​

  • Docling: PictureItem.image.pil_image to PNG at 2x scale (11 images avg)
  • ar5iv: download figures from HTML <img> tags (6 images avg)
  • Storage: alongside markdown, referenced by filename only (portable paths)
  • Orphan cleanup: remove unreferenced images after assembly

Step 6: Quality Scoring​

  • Grade A: 0.85+ (publication-ready)
  • Grade B: 0.70-0.85 (minor issues)
  • Grade C: 0.50-0.70 (major issues)
  • Grade D: 0.30-0.50 (extraction failed)
  • Grade F: <0.30 (unusable)
  • Metrics: section depth, formula coverage, citation completeness, image extraction, LaTeX enrichment

Step 7: JSON Serialization​

  • Canonical UDOM format: .udom.json
  • Schema: config/schemas/udom-v1.schema.json (25 component types)
  • Every component includes: type, content, metadata, provenance (source, page, bbox)
  • Human-readable + machine-parseable

Step 8: Output Generation​

  • Markdown: .md (primary human-readable format)
  • JSON: .udom.json (canonical machine format)
  • Content JSONL: .content.jsonl (block-level with provenance)
  • Audit trail: .audit.jsonl (full extraction provenance)
  • Pipeline report: .pipeline-report.json (timing, quality, source breakdown)
  • Images: {basename}-figure_*.png, {basename}-x*.png

Real Performance (v1.7):​

  • 218/218 papers: Grade A (100%)
  • Average quality: 0.915
  • Total time: 148.9 minutes (41 seconds per paper)
  • Zero failures

The Research Loop: Extraction to Agents to Synthesis to Hypothesis to New Research​

Here's where extraction becomes agentic research infrastructure:

Phase 1: Ingestion​

  • User submits a research question: "What are the most promising approaches to protein folding using geometric deep learning?"
  • System identifies seed papers (e.g., via arXiv search or user upload)
  • UDOM pipeline extracts 20 papers in 14 minutes

Phase 2: Initial Synthesis​

  • market-researcher agent scans all 20 papers for key methods, datasets, benchmarks
  • competitive-analyst agent builds a comparison matrix: which methods work on which datasets?
  • trend-analyst agent identifies temporal patterns: what changed between 2020 and 2024?
  • framework-specialist agent maps findings onto theoretical frameworks
  • synthesis-writer agent combines findings into a 1-page strategic brief with 3 prioritized recommendations

Time: 6 minutes (parallel agent execution)

Phase 3: Gap Analysis​

  • Synthesis identifies gaps: "Most papers use AlphaFold2 as a baseline, but only 3 papers test on multi-domain proteins."
  • System proposes new searches: "multi-domain protein folding geometric deep learning 2023-2024"
  • User approves and system ingests 8 more papers

Phase 4: Incremental Synthesis​

  • Agents re-run on the expanded corpus (now 28 papers)
  • Synthesis updates with new findings
  • Convergence check: does the new synthesis differ from the old? If not, stop.

Time: 4 minutes (incremental update)

Phase 5: Hypothesis Generation​

  • Agents propose: "No existing method combines SE(3)-equivariance with diffusion models for multi-domain proteins. This is an unexplored parameter space."
  • System generates a research proposal with background (synthesized from 28 papers), hypothesis (mechanistic reasoning), proposed experiments (based on gaps), expected outcomes (predicted from related work), and citations (traceable to specific sentences in source papers)

Time: 3 minutes

Phase 6: Recursive Research​

  • User says: "Now find all papers on diffusion models for 3D structure prediction."
  • System ingests 15 new papers, re-runs synthesis, updates hypothesis, detects new gaps, proposes next search

Total time for 6 phases (43 papers, 3 iterations): 27 minutes

Human equivalent: 40 hours (reading + note-taking + synthesis)

Compression ratio: 90x

The Flywheel:​

Every research query leaves behind:

  1. Extracted papers in the knowledge graph (reusable for future queries)
  2. Agent syntheses (reusable for overlapping questions)
  3. Citation graphs (which papers cite which, at the claim level)
  4. Hypotheses (unexplored parameter spaces for future research)

The 10th researcher to ask about protein folding starts with the accumulated knowledge of the previous 9. This is the compounding advantage.

10x Vision: Near-Term (12-18 Months)​

What We Build:​

  1. Vertical Research Agents (10-15 specialists) — Pharmaceutical R&D, Legal Research, Financial Analysis, Academic Research
  2. Real-Time Knowledge Graph Updates — Auto-ingest new publications, update affected syntheses
  3. Collaborative Research Workspaces — Multi-user environments with role-based access and version control for syntheses
  4. Integration with Lab Notebooks — Export hypotheses to ELN, capture results, feed back to agents
  5. Provenance Visualization — Interactive citation graphs, trust scores, contradiction highlighting
  6. Multi-Language Support — Chinese, German, French, Spanish academic papers

What This Unlocks:​

  • Academic Research Groups: Compress PhD lit reviews from 6 months to 6 hours
  • Pharmaceutical R&D: Analyze 10,000 papers on a protein target in 3 days (vs. 3 months)
  • Law Firms: Synthesize case law for a brief in 2 hours (vs. 20 hours of paralegal time)
  • Investment Research: Monitor 500 companies' earnings calls, extract signals in real-time

Near-Term ARR Target: $2M-$5M (50-100 customers at $20K-$50K ACV)​

100x Vision: What This Becomes at Scale (3-5 Years)​

The End State:​

CODITECT becomes the operating system for human knowledge. Every research paper, patent, legal case, clinical trial, earnings report, and government document flows through UDOM extraction. The knowledge graph contains 1 billion+ documents with full provenance. Agents operate at petaFLOP scale, reasoning over the entire corpus in seconds.

What We Build:​

  1. Global Knowledge Graph — 1B+ documents, real-time updates, cross-domain reasoning
  2. Meta-Research Agents — Identify research gaps across entire fields, propose new hypotheses
  3. Autonomous Literature Reviews — 60-100 page lit reviews in 2 hours
  4. Computational Peer Review — Cross-check every claim against the knowledge graph
  5. Research-as-a-Service API — $0.01 per query for developers
  6. AI-Generated Research Papers — Agents write, humans validate and experiment
  7. Patent Prior Art Engine — 100M+ patents + papers, invalidating references in 10 minutes ($5K per report vs. $50K manual)
  8. Clinical Decision Support — Point-of-care synthesis for physicians

Long-Term ARR Target: $500M-$1B​

Target Customers: Beachhead + Expansion Markets​

Beachhead (Year 1-2):​

SegmentSizePainACVSales Motion
Computational Biology Labs200-300 US100+ preprints/week$500-$2K/moBottoms-up (grad students)
Boutique IP Law Firms500-1,000 US$10K-$50K prior art searches$5K/reportFreemium
Biotech Pre-Clinical R&D200-300 US3-6 month IND lit reviews$100K+/yrEnterprise POC

Expansion (Year 3-5):​

SegmentACVSales Motion
Investment Research (hedge funds, PE, VC)$50K-$200K/yrEnterprise
Large Pharma R&D$500K-$1M/yrTop-down enterprise
Academic Institutions (site licenses)$100K-$500K/yrLibrary procurement
Clinical Decision Support (hospitals)$10-$50/queryEHR vendor partnerships
Government Agencies (NIH, FDA, DoD)$1M-$5M/yrFedRAMP + contracting

Total Addressable Market (TAM): $1.8B/year​

  • Academic Research: $500M | Pharma/Biotech: $250M | Legal (IP): $250M
  • Finance: $100M | Clinical: $500M | Government: $200M

The Moat: Why This is Defensible​

1. Multi-Source Extraction Quality (0.915 Average)​

No one else does 3-way alignment (PDF + HTML + LaTeX) with full provenance. GROBID gets ~70% on equations. We get 100% Grade A. 3-4 year technical lead.

2. 150+ Specialized Agents​

Competitors build general-purpose Q&A bots. We build vertical specialists with curated prompts, tools, and knowledge. 2-3 years to replicate.

3. Provenance Guarantee​

Every claim traces to the exact sentence in the source. Not document-level — component-level (Section 3.2, Equation 7, page 14). Critical for FDA, patent law, scientific integrity. 1-2 years to replicate.

4. Compounding Knowledge Graph​

Every research query leaves behind extracted papers, syntheses, citation graphs, hypotheses. The 100th customer benefits from the previous 99. Data flywheel — impossible to replicate without time.

5. Integration Ecosystem​

Lab notebooks, reference managers, grant tools, EHR systems, legal tech. Each integration increases switching costs. 2-3 years to replicate.

6. Network Effects​

More researchers = more syntheses = more valuable knowledge graph = more researchers. Classic network effect. First-mover advantage.

Combined moat: 5-7 year lead.

Market Context: Competitive Landscape​

PlayerWhat They DoWhat They Don't DoOur Advantage
Semantic ScholarIndex 200M papers, semantic searchFull-text extraction, synthesis, agentsWe replace the reading step
Elicit (Ought)Abstract summarization, Q&AFull-text extraction, provenance, multi-sourceWe synthesize full papers with provenance
ConsensusAggregate expert opinionsFull-text extraction, agents, hypothesesWe reason over evidence
ScholarcyPDF to summary flashcardsMulti-source, agents, knowledge graphWe build research infrastructure
GROBIDPDF to XML extractionMulti-source alignment, agentsWe use GROBID-level extraction as one of three sources
Docling (IBM)PDF to markdown with OCRMulti-source, agents, knowledge graphWe use Docling as our PDF engine, then build on top

Why No One Has Built This Yet:​

  1. Extraction is hard — 90%+ quality requires multi-source alignment (3-year R&D)
  2. Agents are hard — 150+ vertical specialists requires MoE routing, prompt engineering, tool integration (systems engineering, not LLM problem)
  3. Knowledge graphs are hard — provenance-aware, component-level, temporal versioning (database problem, not vector store)
  4. The vision requires all three — extraction alone (GROBID), agents alone (Elicit), knowledge graphs alone (Semantic Scholar) are not products. You need all three together.

The VC One-Liner​

"We're building the operating system for computational research — where AI agents reason over 1 billion documents with full provenance, compressing PhD literature reviews from 6 months to 6 hours. $1.8B TAM, 5-7 year moat, network effects kick in at 10K researchers."

Alternative (Provocative):

"Human researchers read 2-3 papers per day. Our agents read 1,000. We're not building better search — we're building what comes after human-speed research. The knowledge graph compounds. The agents get smarter. The first 10K researchers will never leave."

What Would the Founders Say?​

Yann LeCun (Meta AI, Turing Award Winner):

"Your UDOM is a world model for documents. Your agents are planners. The compounding knowledge graph is the key. If you get to 1 billion documents with provenance, you'll have the best training data for the next generation of research models. This is infrastructure for AGI-level research assistants."

Demis Hassabis (Google DeepMind, CEO):

"AlphaFold solved protein structure. But the literature review that led to AlphaFold took 3 years. You're compressing that to 3 days. When agents can propose hypotheses, identify gaps, and trigger new research autonomously, you've automated the scientific method."

Dario Amodei (Anthropic, CEO):

"Your provenance guarantees solve the 'honest' part for research synthesis. Every claim traceable to the source? That's the integrity layer the field desperately needs. And the MoE routing — 150+ vertical agents instead of one generalist — that's how you get reliability at scale."

Revenue Model​

PhaseTimelineRevenueModel
Usage-BasedYear 1-2$500K ARR$0.10/paper + $0.02/agent call
Enterprise SaaSYear 2-4$10M ARR$100K ACV (pharma, legal, finance)
Platform/APIYear 4-7$100M ARR$0.01/query + data licensing
AI Research StudioYear 5-10$500M-$1B ARR$50-$500/mo per researcher

The Existential Question​

If CODITECT succeeds, reading papers becomes obsolete — not because humans won't want to read, but because agents will be faster, cheaper, and more comprehensive.

A pharmaceutical company has 10,000 papers to review. Do they hire 10 humans to read 30 papers per day? Or do they use CODITECT to synthesize all 10,000 in 3 days and spend the remaining time thinking instead of reading?

The shift: from reading as work to reading as leisure.

And the deeper question: What does research become when reading is free?

If literature reviews take 6 hours instead of 6 months, does the PhD take 3 years instead of 6? If prior art searches are instant, do law firms charge less? If clinical decision support is real-time, do doctors make fewer mistakes?

The best researchers in 2035 won't be the ones who read the most papers. They'll be the ones who orchestrate agents the best — who ask the right questions, spot synthesis gaps, and know when to trust the agents and when to dig deeper.

CODITECT isn't replacing researchers. It's upgrading them.

Document Metadata:

  • Authors: CODITECT Strategy Team (synthesis-writer agent + human review)
  • Version: 1.0.0 (Draft)
  • Status: Draft (awaiting /moe-judges evaluation)
  • Next Review Date: 2026-02-18