Registry row
Check identity and status
Start with the role registry for version, authority level, active or superseded status, human-gate state, default output form, and validators.
AI research-agent deep dive / Phase 5B
Roles And Skills
AEther Flow uses role contracts and governed skills to keep work legible, bounded, and reviewable. A role name or skill entry point is not permission by itself; task-local records and AgentJob allowlists define what the current transaction may actually do.
Role families
Different roles exist for different bounded work.
Role families help classify work, but they still inherit the current transaction boundary. Human-gated and claim-promoting authority remains protected.
Research routing
Director roles
Director of Research and Project-System Director route one bounded step. Their normal authority is routing or project control, not claim promotion.
Physics drafts
Science-draft roles
Ontology Formalizer, Candidate Constructor, Refuter, Smuggling Auditor, and Theoretical Continuation Selector work inside draft/control boundaries.
Scientific gate
Gate Chair
The Gate Chair is defined as human-gated. Its existence does not authorize execution or promotion without explicit tracked approval.
Project system
Maintenance roles
Project-Control Maintainer, Validator Engineer, Memory-System Maintainer, and Documentation Curator maintain project surfaces without promoting physics claims.
Governed skills
Skills define procedures, not blanket permission.
A skill may tell an operator how to continue research, improve the project system, refresh memory, or publish explainers. It still works inside source authority, role contracts, validators, and allowlists.
continue-research
Research continuation
Routes research-control continuation and physics AgentJobs. It is not project-system repair by default.
improve-project-system
Project-system repair
Handles classifiers, sidecars, signals, receipts, and one bounded project-system AgentJob. It is not physics derivation.
project-memory-system
Memory and derivatives
Bootstraps and validates memory, wiki, registry, and derivative surfaces while keeping generated artifacts non-authoritative.
publication skills
Governed public explainers
HTML and visual explainer skills support reader surfaces only through the governed source-spec and publication-brief path.
Overread prevention
Catalog visibility must not become authority drift.
The main risk is reading a visible catalog entry as active permission. The repository prevents that by making status, execution records, allowlists, and gates explicit.
Superseded roles
Historical does not mean active
Old role versions remain visible so old execution records are interpretable. The status field controls the current role state.
Skill mentions
Entry point does not equal permission
A skill can define a workflow, but the current AgentJob allowlist still controls actual writes and outputs.
Task overlays
One-job adjustments stay local
A task overlay adapts a registered role for one job. It does not become a reusable role version by convention.
Public catalog
Navigation, not registration
This website page cannot register roles, supersede roles, change skill contracts, or expand authority.
Reader path
Use the catalog with workflow and source authority.
Roles and skills are easiest to read after the workflow page: the workflow explains the transaction chain, while this page explains why role and skill labels remain subordinate to current job authority.
Back to AI research-agent trackReturn to the track landing page for the AI research-agent page family.Research-agent workflowRead how Director decisions, AgentJobs, completions, and handoffs form the workflow chain.Source authority pageReview how website pages, generated explainers, registries, and tracked sources relate.Upstream roles and skills catalogGenerated noncanonical source-backed catalog for role and skill inspection order.
Source authority
What this page can and cannot establish.
This page can orient readers to role families, skill entry points, and the authority inspection order. It cannot register roles, supersede roles, change skill contracts, expand task-local allowlists, change routing behavior, or promote physics claims.