DeepSci

Agentic Deep Research and Auto-Scientist workbench.

DeepSci follows the same product pattern as BTCBoard and PRForge:

papers / local corpus / source adapters
  -> deep research agents
  -> Evidence Graph
  -> Citation Guard
  -> Hypothesis Engine
  -> Experiment Planner
  -> Paper Draft
  -> Reviewer Simulator
  -> Research Memory
  -> exportable research artifact

The product goal is a lightweight but useful research board: strong enough to support literature review, hypothesis generation, experiment planning, paper drafting, and reviewer simulation without a heavy distributed multi-agent stack.

Current Status

DeepSci has completed Phase 0 through Phase 6 plus the final persistence/export layer.

• Phase 0: project scaffold complete.
• Phase 1: Deep Research Core complete.
• Phase 2: Evidence Graph and Citation Guard complete.
• Phase 3: Hypothesis Engine complete.
• Phase 4: Experiment Planner complete.
• Phase 5: Paper Draft and Reviewer Simulator complete.
• Phase 6: Auto-Scientist Loop complete.
• Final integration layer: source adapter status, JSONL research memory, history API, Markdown/BibTeX/JSON export complete.

See docs/PHASED_ROADMAP.md for the staged plan.

Core Workflow

Read papers
  -> Find gaps
  -> Generate hypotheses
  -> Critic debate
  -> Score novelty / feasibility / impact
  -> Plan experiments
  -> Estimate reproducibility and compute risk
  -> Generate paper draft
  -> Simulate reviewers
  -> Build rebuttal / revision plan
  -> Save research memory
  -> Recommend next experiment

Agent Modules

• Intent / Field Agent: classifies the research task.
• Literature-Agent: retrieves seed and local papers.
• Paper-Reading Agent: extracts methods, claims, datasets, metrics, and limitations.
• Evidence-Graph Agent: turns paper claims into graphable evidence.
• Citation-Guard Agent: blocks unsupported claims.
• Contradiction Detector: finds evidence conflicts.
• Gap-Finder Agent: identifies research gaps.
• Hypothesis-Agent: generates ranked hypotheses.
• Critic-Agent: attacks and revises weak hypotheses.
• Novelty Scorer: scores novelty, feasibility, impact, and survival.
• Experiment-Planner Agent: turns ideas into minimum viable experiments.
• Baseline Finder, Dataset Finder, Metric Planner: fill the experimental matrix.
• Reproducibility-Agent: scores risk and blockers.
• Paper-Writer Agent: creates a cited paper draft.
• Related-Work Agent: clusters prior work.
• Reviewer-Simulator Agent: simulates reviewer feedback.
• Rebuttal-Agent: creates revision and rebuttal plans.
• Research Memory, Hypothesis Scoreboard, Lab Notebook, Reflect Agent: close the loop.

Outputs

DeepSci returns:

• research_answer
• key_papers
• evidence_items
• evidence_graph
• claim_evidence_map
• unsupported_claims
• contradictions
• citation_quality_score
• citation_guard_verdict
• hypotheses
• novelty_score
• feasibility_score
• impact_score
• critic_report
• experiment_plan
• baselines
• datasets
• metrics
• minimum_viable_experiment
• reproducibility_score
• paper_draft
• reviewer_feedback
• accept_probability
• revision_plan
• research_memory
• experiment_history
• hypothesis_scoreboard
• lab_notebook
• next_best_experiment
• external_sources_status
• research_history
• export_artifacts

Source Adapters

DeepSci is offline-first by default.

Ready now:

• bundled seed corpus,
• local_papers supplied in request payload,
• adapter status reporting.

Connector slots:

• arXiv,
• Semantic Scholar,
• PubMed,
• Zotero,
• GitHub paper/code repositories.

The connector slots are intentionally safe stubs until live credentials/network retrieval are configured.

Useful environment variables:

DEEPSCI_ARXIV_ENABLED=
DEEPSCI_SEMANTIC_SCHOLAR_KEY=
DEEPSCI_PUBMED_EMAIL=
DEEPSCI_ZOTERO_API_KEY=
DEEPSCI_GITHUB_TOKEN=

Backend

Run from:

cd D:\ir\DeepSci-main

Create a venv if dependencies are needed:

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install fastapi uvicorn pydantic

Start the API:

$env:PYTHONPATH="backend"
python -m uvicorn app.main:app --reload --host 127.0.0.1 --port 8000

Frontend

cd D:\ir\DeepSci-main\frontend
npm install
npm run dev

The frontend is a research workbench with:

• research input,
• citation guard metrics,
• agent cards,
• evidence board,
• hypothesis portfolio,
• experiment blueprint,
• paper draft,
• reviewer feedback,
• lab notebook,
• persistence/export panels.

API

• GET /health
• POST /api/v1/research
• GET /api/v1/history
• GET /api/v1/export/markdown
• GET /api/v1/export/markdown/{run_id}
• GET /api/v1/export/bibtex
• GET /api/v1/export/bibtex/{run_id}
• GET /api/v1/export/json
• GET /api/v1/export/json/{run_id}

Example Payload

{
  "research_question": "How can agentic deep research systems prevent unsupported scientific claims?",
  "field": "machine_learning",
  "mode": "deep_research",
  "depth": "deep",
  "source_filters": ["seed_corpus", "local"],
  "time_range": "all",
  "local_papers": [],
  "output_type": "research_report"
}

Persistence

Local research memory is stored as JSONL:

• backend/data/research_runs.jsonl

Each completed research run records:

• run id,
• question,
• field,
• citation quality,
• hypothesis count,
• accept probability,
• export URLs,
• full response payload.

Export

DeepSci can export the latest or selected run as:

• Markdown research report,
• BibTeX references,
• JSON archive.

Validation

Verified locally:

python -m compileall backend\app

Manual smoke checks passed for:

• deep research run,
• research memory persistence,
• history summary,
• BibTeX export.

pytest and frontend dependencies were not installed during the last validation pass. Install them before running full unit or frontend builds.