The Engine
How Sourcery researches for you
01 · Plan
Research Plan
LLM drafts key sub-questions, source types, search angles, and seed queries before touching the web.
02 · Search
Focused Queries
Plan-driven queries are issued; pages are fetched, HTML and PDFs extracted to clean Markdown locally.
03 · Extract
Structured Claims
Each source yields structured claims with entities, confidence scores, negation, and uncertainty flags.
04 · Expand
Recursive Crawl
Related URLs and queries are recursively followed to configured depth — only grounded in actual source content.
05 · Verify
Contradiction Check
Heuristic contradiction detection surfaces numeric and logical disagreements across all claims.
06 · Synthesize
Final Report
Chunk-parallel synthesis merges everything into a citation-linked, fully referenced Markdown report.
The Output
Published Research Reports
How Vulnerable Is Asia to a Gulf Energy Shock?
India, China, Japan, and South Korea assessed against the 2026 Iran-Israel war and the largest supply disruption in IEA history. Reserve levels, Hormuz exposure, IEA membership, ISPRL financial health, and three-scenario outlook — all citation-grounded across 17 sources.
The Coming Power Wars Between Humans and Datacenters
AI-driven data centers assessed against rising U.S. electricity prices, grid bottlenecks, capacity-market shocks, fossil-fuel lock-in, water stress, subsidies, and the unresolved question of who gets power priority during shortages. Citation-grounded across 54 sources.
What Happens If Compute Becomes a Sovereign Reserve Asset?
A deep analysis of whether nations could treat AI compute like strategic reserves: sovereign wealth funds buying data centers, national compute reserves, compute-backed collateral, and the possibility of future compute sanctions — grounded across 38 sources.
The Financialization of Compute Futures
How GPU-hours are becoming tradable, securitized, and potentially hedgeable assets. Covers compute spot markets, GPU-backed debt, futures exchanges, fixed-price inference contracts, energy hedging, tokenized compute, and the structural risks behind turning AI infrastructure into a financial market.
Next report in progress…
Sourcery is crawling, extracting, and synthesizing. Check back soon for the next deep-research dispatch.
Another topic under research…
More reports will appear here as Sourcery completes each deep-research run.
Why Sourcery
Built differently from the start
Zero hallucination tolerance
Every claim in a Sourcery report is anchored to a numbered source. Inline citations are not decorative — they map to real URLs that were fetched and parsed in that exact run.
Recursive, plan-driven research
Sourcery doesn't just search once. It plans, expands related URLs and queries to configurable depth, and only follows links it actually found in source content — not LLM speculation.
Persistent research memory
Sourcery stores normalized source markdown, quality scores, entities, and summaries in SQLite across runs — so follow-up research reuses fresh cached sources instead of refetching everything.
Contradiction detection built in
After collecting all source claims, Sourcery runs heuristic contradiction detection — surfacing numeric mismatches and logical negation conflicts so the final report explicitly flags disputes.
Resumable, fault-tolerant runs
Every source node is checkpointed to SQLite as it's processed. If a run is interrupted, it can be resumed exactly where it left off — without re-fetching sources already completed.
Chunk-parallel synthesis
Final reports aren't generated in one massive prompt. Sourcery splits sources into chunks, synthesizes intermediate reports in parallel, then merges them — preserving original citation numbers throughout.