jordan-gibbs/hyperresearch

What it solves

Hyperresearch is a deep research harness designed to transform Claude Code into a high-capacity research agent. It solves the problem of shallow, one-shot research runs that often suffer from hallucinations, lack of source provenance, and the loss of data between sessions. By implementing a rigorous 16-step pipeline, it ensures that reports are adversarially audited, verified for citation accuracy, and grounded in a persistent, searchable vault of sources.

How it works

The system operates through a tier-adaptive pipeline that routes queries based on complexity (light, full, or dissertation). It uses a series of specialized subagents (fetchers, analysts, critics, and patchers) to decompose queries, perform width and depth investigations, and synthesize a final report.

Key mechanisms include:

• The Vault: A persistent SQLite-indexed collection of markdown notes where every fetched source is stored, allowing future sessions to reuse data.
• Adversarial Review: Four parallel critics attack the draft, and a tool-locked patcher applies surgical edits rather than regenerating the text to prevent drift.
• Verification Gates: A "ship gate" verifies quote integrity, checks for retracted scholarly sources, and audits every citation-sentence binding to block hallucinations.
• Source Ranking: Sources are scored based on authority, utility, and PageRank centrality, with a specific audit to ensure syndicated copies of the same press release are not counted as consensus.

Who it’s for

Researchers, academics, and power users who need high-fidelity, long-form reports (from 30-minute surveys to 80,000-word dissertations) with full source provenance and a verifiable audit trail.

Highlights

• Massive Source Capacity: Capable of processing 250+ sources in a single run, with dissertation mode supporting 300-450 sources.
• Scholarly Integration: Single-query access to OpenAlex, Crossref, CORE, DOAB, ClinicalTrials.gov, SEC EDGAR, and FRED.
• Persistent Knowledge: A searchable markdown-plus-SQLite vault that compounds knowledge across sessions.
• Hallucination Blockers: Structural enforcement of quote integrity and citation verification; retracted sources block the shipment of the report.
• Resumable Runs: Manifest-based tracking allows crashed runs to resume exactly where they stopped.
• MCP Server Support: Includes an MCP server to allow other tools like Claude Desktop or Cursor to interact with the vault.