Research Standards

Research Methodology

This is the standard I want every research page to meet: a directed question, inspectable sources, separate specialist passes, explicit limitations, and a human publication decision. The existing 104 topic maps are being checked against individual sources. Generated search links are not counted as citations.

104

Active Domains

Published Datasets

Registered Sources

103

Pending Registries

4-Phase Research Process

The target workflow for new briefs, plus the standard older briefs are being brought up to.

Phase 01Continuous

Signal Detection

The intended intake combines directed manual searches, specialist-agent scans, academic papers, industry reports, vendor documentation, developer surveys, and regulatory sources.

Maintain a curated watchlist for the domains under active review
Track relevant product launches, research releases, and regulatory updates
Look for patterns across domains without treating correlation as proof
Record contradictions between vendor claims and independent evidence

Phase 022-4 hours per domain

Deep Analysis

For selected questions, specialist passes extract supported quantitative data, qualitative context, uncertainty, and possible implications.

Record a metric only when the cited source supports its scope and date
Separate observed facts, source claims, and FrankX interpretation
Identify methodology limitations in cited research
Seek two independent sources for consequential claims; label single-source evidence plainly

Phase 03Per claim

Validation & Rating

The target standard is to rate consequential quantitative claims by source quality, recency, and corroboration. Older briefs remain provisional until that review is complete.

Target high confidence: two or more independent, directly relevant sources
Medium confidence: one authoritative source or a documented expert consensus
Low confidence: vendor-sourced or weakly corroborated data — label it
Exclude or rewrite claims whose scope, date, or provenance cannot be verified

Phase 04Per domain update

Publication & Schema

Reviewed material is synthesized into readable briefs with registered sources, explicit limits, and structured metadata where the page content supports it.

Use TechArticle or FAQPage JSON-LD only when the visible page warrants it
Use question-based headings when they help a reader navigate the evidence
Link related domains without implying evidentiary support between them
Update lastUpdated timestamps only when a human review actually occurs

Quality Principles

The standards that govern what gets published and how claims are presented.

Specificity Over Generality

When a brief uses a statistic, pair it with its scope, source, and date. Unsupported precision is worse than a careful qualitative statement.

Cross-Reference Everything

Two independent sources are the target for high-confidence claims. Single-source evidence should be marked, not quietly upgraded.

Vendor Skepticism

Vendor-sourced metrics (funding, user counts) are treated as claims, not facts. We seek independent verification.

Recency Matters

AI moves fast, but newer is not automatically better. Use current sources for changing facts and foundational sources where they remain authoritative.

Source Hierarchy

Not all sources are weighted equally. Our evidence hierarchy prioritizes independent, primary research.

Source TypeExamplesWeight

Industry ReportsGartner, McKinsey, Forrester, IDCHigh

Academic ResearchArXiv, NeurIPS, ICML, ACL proceedingsHigh

Developer SurveysJetBrains, Stack Overflow, GitHub OctoverseHigh

Regulatory FilingsEU AI Act, FDA clearances, SEC filingsHigh

Vendor DocumentationOfficial docs, benchmarks, changelogsMedium

Press CoverageTechCrunch, The Information, ReutersMedium

Community SignalsGitHub stars, HN discussions, Reddit threadsLow — directional only

Domain Coverage

104 research domains at different stages of review. Counts below come from the approved source registry. The separately published Observatory dataset is identified on its own; held claim registries are labeled as pending.