Research Scout
Searches recent scholarly metadata using taxonomy-aware queries and updates an unreviewed candidate queue.
Agents & skills
A research loop that cannot promote itself.
Brali separates discovery, evidence review, protocol design, and taxonomy. The scout can find a lot. It cannot call what it finds “reviewed”. That boring boundary is the useful part.
Evidence Reviewer
Reads the actual source, checks what was studied, and decides whether it supports, challenges, or changes Brali content.
Protocol Builder
Turns reviewed or clearly practical atomic hacks into short executable personal experiments.
Taxonomy Curator
Maps knowledge to Life Areas, Growth Zones, and method tags without multiplying categories or breaking stable URLs.
Agent responsibilities
Research Scout
Continuously discovers recent scholarly work using research-queries.json, deduplicates by DOI or stable identifier, records the matching taxonomy and risk flags, and outputs only unreviewed research candidates. It never converts metadata into a public claim.
Evidence Reviewer
Reads the actual source and checks population, intervention or exposure, outcomes, effect direction, practical meaning, limitations, conflicts, retractions, and whether the exact Brali wording is supported. Its decision is one of rejected, watch, support-existing, challenge-existing, propose-hack, or propose-protocol.
Protocol Builder
Combines reviewed or clearly practical atomic hacks into 1–5 step protocols. It reuses hack IDs, preserves the weakest relevant evidence state, adds a check-in and stop/change rule, and does not invent durations, percentages, mechanisms, or guarantees.
Taxonomy Curator
Maps new knowledge to the smallest useful set of Life Areas and Growth Zones, keeps methods as tags when appropriate, detects duplicates, and protects stable canonical URLs. A new research term is not a reason to create a new zone.
Knowledge contracts
The loop has explicit shapes for an atomic Hack, an executable Protocol, an unreviewed Research Candidate, and the Evidence Decision that connects source review to a content change.
The guarded loop
search → candidate → source review → hack → protocol → taxonomy → publish
No automated step can jump from search metadata to reviewed evidence. Sensitive material gets risk flags and a stricter review path.
Continuous discovery
A scheduled GitHub workflow runs the metadata scout each week. It uses the public Crossref REST API, merges new records into the candidate queue, validates taxonomy links, and opens or refreshes a review pull request. The discovery step does not need an LLM key.
Search lenses live in research-queries.json, so they can evolve with the taxonomy instead of being buried inside one agent prompt.
Machine-readable registry
agents/registry.json lists the roles, skills, inputs, outputs, and publishing boundary for agent runtimes.
Use with AI systems
The Markdown skill files are intentionally provider-neutral. A capable coding or research agent can load the relevant skill, read the repository contracts, and produce a proposal without needing a Brali-specific runtime.