Research Framework Overview

Prompt-first procedure: Describe the outcome you want in your agent conversation. The agent should select and load the appropriate AIWG assets, explain material changes, request any needed approval, and report verification evidence. Exact commands and flags appear only in the CLI reference.

The Research Framework helps teams turn source material into reviewable literature artifacts: search plans, acquired papers, grounded summaries, citations, quality assessments, gap notes, and provenance records. Use it when research needs to survive beyond a chat transcript and when citations or source boundaries matter.

Common Use Cases

• Build a literature review for an architecture decision or policy brief.
• Collect papers for a systematic scan and keep acquisition records tied to stable `REF-XXX` identifiers.
• Summarize PDFs with source-grounded claims and reviewer-visible uncertainty.
• Track citation networks, quality posture, and gaps before synthesis.

Why It Exists

Literature work often fails because sources, claims, and quality judgments drift apart. The framework addresses that by keeping each research output tied to its source identity, retrieval metadata, and review state.

Problem	What the Framework Does
Manual paper search across databases	Captures semantic searches, relevance notes, and gap-detection output
Tracking where papers came from	Assigns persistent `REF-XXX` identifiers during acquisition
Paywalled paper handling	Supports Unpaywall lookup and manual upload workflows
Reading full papers	Produces source-grounded summaries at multiple levels
Fabricated citations	Flags unsupported claims for review instead of treating them as accepted
Inconsistent quality assessment	Applies GRADE and FAIR-style checks with visible criteria
Fragmented notes	Produces literature notes that can be linked into a broader knowledge base

Workflow

1. Discovery      Find papers via semantic search, gap detection
2. Acquisition    Download PDFs, assign REF-XXX, validate integrity
3. Documentation  RAG summarization, structured extraction, lit notes
4. Citation       Format citations, build citation networks, bibliographies
5. Quality        GRADE assessments, FAIR validation, quality scoring
6. Synthesis      Create permanent notes, link related work
7. Gap Analysis   Identify research gaps and contradictions
8. Archival       OAIS-inspired archival workflow, integrity verification, provenance

Agent Roles

The framework includes role-specific agents for each workflow stage:

Agent	Purpose	Key Capability
`discovery-agent`	Semantic search, gap detection	Source discovery and citation network traversal
`research-acquisition-agent`	Download PDFs, assign IDs	FAIR validation, SHA-256 checksums, deduplication
`documentation-agent`	RAG summarization	Source-grounded summaries and literature notes
`citation-agent`	Format citations	Citation formatting and network analysis
`quality-agent`	Assess paper quality	GRADE methodology (High/Moderate/Low/Very Low)
`archival-agent`	Long-term preservation	OAIS compliance; SIP/AIP/DIP packages
`provenance-agent`	Lineage tracking	W3C PROV logging; reproducibility packages
`workflow-agent`	Orchestrate pipelines	DAG-based execution; parallel stages; failure recovery

Key Design Choices

REF-XXX Identifiers

Every paper gets a persistent `REF-XXX` identifier when acquired. This ID appears in summaries, citations, notes, and provenance records, providing a stable cross-reference that survives database reorganizations and file moves.

Source-Grounded Summaries

The documentation agent checks summary claims against the source PDF before treating them as accepted. Claims below the configured confidence threshold are flagged for user review rather than included silently. This is the reason the framework requires actual PDFs rather than just metadata.

FAIR-Oriented Review

Acquired papers are scored on Findability, Accessibility, Interoperability, and Reusability (0-100). Papers below threshold are flagged rather than silently accepted. The score records a workflow assessment; it is not a certification of compliance or source quality.

GRADE Quality Assessment

The quality workflow asks the agent to assess evidence using explicit criteria and record uncertainty. Review the suitability of GRADE for the question and evidence type; a generated assessment still requires domain review.

Storage Structure

All research artifacts go in `.aiwg/research/`:

.aiwg/research/
├── discovery/              # Search results, strategies, gap reports, acquisition queues
├── sources/                # Stage 2 output
│   ├── pdfs/               # REF-XXX-{slug}.pdf
│   ├── metadata/           # REF-XXX-metadata.json
│   └── checksums.txt       # SHA-256 integrity verification
├── knowledge/              # Stage 3 output
│   ├── summaries/          # REF-XXX-summary.md
│   ├── extractions/        # REF-XXX-extraction.json (structured data)
│   └── notes/              # REF-XXX-literature-note.md + permanent notes
└── config/                 # Per-agent configuration YAML files

Integration with SDLC

The research framework can be used alongside sdlc-complete. The artifact directories do not overlap (`.aiwg/research/` vs `.aiwg/requirements/` etc.). A common pattern is using the discovery and documentation agents during Inception and Elaboration phases to ground architectural decisions in the literature:

Use AIWG to complete this documented outcome: The research framework can be used alongside sdlc-complete. The artifact directories do not overlap (.aiwg/research/ vs .aiwg/requirements/ etc.). A common pattern is using the discovery and documentation agents during Inception and Elabora
Have it inspect the current state, explain the plan, ask before material
changes, and report the result with verification evidence.

References

• Quickstart — Deploy and first literature review
• `@$AIWG_ROOT/agentic/code/frameworks/research-complete/agents/` — Agent definitions
• `@$AIWG_ROOT/agentic/code/frameworks/research-complete/inception/vision-document.md` — Framework vision