Clerk

How It Works

One topic in. 100+ parallel investigations. A statistically-backed S-Rank report out.

Phase I

Decomposition

— Commander Agent

A topic like “Web3” is broken into 100 specific sub-niches by frontier reasoning models. Output: a Research Map with unique IDs, timeframes, and search parameters.

Phase II

Distributed Extraction

— Worker Agents ×100

100 simultaneous Temporal workflows fan out — each researching one sub-niche with neural search, clean scraping, and schema-enforced extraction of reliability scores, technical claims, and sentiment.

Phase III

Aggregation & Statistics

— Analyst Agent

All results are loaded into DuckDB. Correlation matrices, trend vectors, and predictive “odds” are calculated with deduplication across sub-niches.

Phase IV

Synthesis

— Synthesizer Agent

Claude 3.5 compiles the statistics into an S-Rank enterprise research paper — executive summary, citations, visualizations, and predictive odds analysis.

Research Output Structure

docs/
└── [Research-Topic-ID]/          # e.g., /web3-analysis-2026
    ├── summary.md                # The S-Rank Executive Report
    ├── statistics/               # Odds calculations & trend vectors
    ├── niches/                   # 100+ Sub-niche deep dives
    │   ├── layer2-scaling/       # raw-distilled.json, year-by-year.md
    │   └── zkp-origins/
    └── assets/                   # Generated graphs and PDFs

Browse Sample Research