Skip to content

Research and evidence

Genesis treats a claim without a source as a defect. The research path has three checks.

The source registry records which sources are trusted for which kind of claim. Source anchors pin each claim to a retrievable URL. The registry recognizes 30 sources in 11 tiers. That is a catalog of sources Genesis knows how to rank, not a list queried on every run.

The sources depend on the surface you use.

Surface Research sources Network
Standalone CLI, genesis init GitHub repositories and issues, plus your configured LLM Yes
Standalone CLI, genesis research Merges, ranks and validates findings you or your agent collected beforehand; with no data it prints a collection guide No collection of its own
Agent skill (SKILL.md) GitHub, plus ecosystem and community search, package registries, security data and video metadata, through whatever web and MCP tools the host agent provides Through the host agent
Genesis MCP server None. Four read-only analysis tools over a local project No

Genesis researches across the software ecosystem using the sources available in the current execution environment. When it runs as an agent skill, its research workflow can combine GitHub with available web and MCP sources, community discussions, package registries, security data and relevant media. Which of these actually run depends on the tools the agent has; a thin result is reported as thin.

Using external MCP tools for research is a skill behavior of the host agent. It is separate from the MCP server Genesis exposes.

Citations are checked live. A URL that does not resolve fails validation, and the result is recorded rather than hidden.

ARCHITECTURE_EVIDENCE.md bundles the evidence behind a decision: research quality, repositories scanned, pitfalls found and the mitigation mapped to each. The pack states when it is unsigned, which is normal before you have confirmed an architecture choice.

When research is thin, Genesis says so in the pack (Quality: THIN) instead of presenting a confident scaffold. Treat a thin pack as a prompt to narrow the description or research further.

The counts for engines, gates and sources are published in data/public-product-facts.json and generated from the code.