The DATA methodology

LighthouseHive is built on a four-stage workflow we call DATA: Democratic Accountability Through Analysis. Ingest · Analyze · Validate · Deliver. Each stage has a purpose; each is bounded by what came before it. The methodology is what allows a Hive to return an answer in seconds that would otherwise take a researcher days — without giving up the rigor that makes the answer trustworthy.

 

Ingest

Bounded before built. Every Hive begins with a question of scope, not scale. Which institutions count as sources? Which document types? What time range? What geographies? We work through this with domain experts until the boundary is documented and signed off before any technical work begins. It is the single most important determinant of what a Hive can and cannot answer — and it is the discipline that keeps a Hive honest about its own limits.

Curated sources, not the open web. Hives are restricted to verified official and institutional sources within the delineated scope — UN databases, treaty archives, government portals, academic repositories, and equivalent institutional record sets for each domain. We do not train on general internet content. We do not scrape social media. The corpus is what the delineation document says it is, nothing more. Automated acquisition runs continuously and is audited against a completeness target before a Hive advances to the next stage.

 

Analyze

Structure before search. Raw documents are not meaningfully searchable. Before a Hive can answer a question, every document is tagged against a structured metadata schema — institution, document type, date, language, geographic scope, topic — and those tags are built directly into the knowledge structure that sits underneath retrieval. This step does more for accuracy than any model choice. It is the difference between a search that finds a document and a search that understands what the document is, who wrote it, when, and why it matters.

Testing before release. Every Hive is tested against a purpose-built query suite spanning factual retrieval, temporal comparison, entity-specific analysis, cross-source synthesis, and gap identification. No Hive is released below a rigorous retrieval accuracy bar. If initial accuracy falls short, we do not tune the message; we tune the underlying methodology until the result is earned.

 

Validate

Independent experts, not self-certification. Accuracy is not asserted. It is verified. Before a Hive goes live, a panel of independent domain experts reviews sample responses, scores them against a defined quality scale, and signs off on release. Experts also verify that every response carries a valid source citation. Expert validation and provenance verification are permanent records, retained for every Hive.

For the Comprehensive Disarmament Initiative — our first operational Hive — this role is held by Professor Stuart Maslen, LighthouseHive's Chief Academic Adviser, of the Geneva Academy of International Humanitarian Law and Human Rights and the University of Johannesburg, and one of the world's leading authorities on arms control and international humanitarian law.

Provenance on every answer. Every answer a Hive returns links back to the specific document it came from, with page-level citation where available, and carries a confidence score reflecting retrieval strength. Users see the evidence as well as the conclusion. Nothing we return is a black box. If it cannot be sourced, it is not surfaced.

 

Deliver

Live, not frozen. A Hive is not a one-time build. Once live, it runs under continuous monitoring: automated availability checks, routine accuracy tests against the benchmark suite, and rapid integration of new documents as they appear in source repositories. Structured user feedback — ratings, issue categories, low-confidence flags — is reviewed on a weekly rhythm, and domain experts conduct a full audit each quarter. The methodology is versioned. When we improve it, we say so.

​​

What makes this different

Curated by design. Each Hive draws only on verified official and institutional sources within a documented scope. The corpus is deliberate, not scraped — and that boundary is what makes the analysis defensible.

 

Evidence-bound answers. Every response is anchored in the source material a Hive can actually cite. Where the record is silent, the Hive says so. The platform's job is to surface what exists, not to fill in what doesn't.

 

A tool for expert judgment, not a substitute for it. LighthouseHive accelerates the research that journalists, lawyers, and scholars already do well. It compresses days of archival work into seconds, so that human judgment can be spent where it matters.

​

Transparent by design. Accuracy improvements, source additions, and methodology changes are versioned and documented. When a Hive gets better, users can see how — and when.