AI Research System / Roles and Schemas

Roles and Schemas

Roles classify responsibilities and schemas constrain records; neither creates proof or live permission without the task-local workflow boundary.

Read matrix Workflow context Human gates

Role labels, schemas, registry rows, and current-job allowlists are separate authority-inspection layers.

The conservative reading is simple: role identity is not proof, schema compliance is not claim adoption, and registry visibility is not live authority. A registered role becomes current execution authority only when a task-local execution-role record and one AgentJob allowlist bind it.

Static diagram

Keep role labels below schema and job authority.

The role stack shows that labels, schemas, execution-role records, AgentJob allowlists, and validators occupy distinct authority layers.

Diagram showing role labels below registry status, execution records, AgentJob allowlists, and validators. — The diagram illustrates the role-authority stack from role labels through registry status, execution records, AgentJob allowlists, and validators.

Claim-authority matrix

Authority starts narrow and stays record-bound.

This matrix groups roles by stable authority class. It does not replace active-version resolution from the upstream role registry.

Role family	Primary authority	Can do	Cannot do
Director of Research	Routing control	Select one bounded next step and create or reuse one AgentJob when no human gate is required.	Modify sources, promote claims, or prove a scientific result by identity.
Science-draft roles	Draft/control construction, audit, stress, or routing	Construct candidates, formalize assumptions, refute candidates, audit smuggling, or select one theoretical continuation inside the job boundary.	Adopt ontology, promote benchmarks, derive GR, or create proof authority by role label.
Documentation and project-control roles	Project documentation, validators, memory, or controlled repair	Maintain allowed project surfaces, public documentation, memory tooling, and deterministic validators inside an owning AgentJob.	Change physics claim status or treat generated surfaces as source authority.
Gate Chair	Human-gated scientific decision path	Render protected promotion, closure, or suspension decisions only after explicit tracked approval.	Execute autonomously or become active because a page, registry row, handoff, or validator mentions it.

Schema map

Schemas are control contracts, not decorative references.

Each schema matters because it defines required fields, validators, and authority consequences for a record class.

Schema	Controls	Reader explanation	Boundary
Role Schema	Registered role contracts	Defines role identity, version, authority fields, validators, outputs, and human-gate status.	A registered contract is a template until a task-local record binds it.
Director Decision Schema	Director Decision Records	Makes routing reasoning inspectable and superseded rather than silently rewritten.	A decision routes one job; it does not settle scientific truth.
AgentJob Schema	Bounded executable contracts	Defines objective, allowed reads, allowed writes, validators, memory preflight, expected outputs, and claim boundary.	The AgentJob is one transaction, not reusable permission.
Execution Role Schema	Task-local role bindings	Distinguishes registered role, task overlay, and one-job provisional role semantics for one AgentJob.	Task overlays and provisional roles expire with the owning job unless separately registered.
Documentation Impact Schema	Project-system documentation receipts	Records whether project-system changes require documentation updates and how generated derivatives were handled.	A receipt accounts for effects; it does not make derivatives authoritative.

No-proof status

The proof-authority column is always no.

Roles can contribute routing, construction, audit, refutation, documentation, validation, or human-gated decision records. Proof still depends on source mathematics and tracked governance.

Question	Safe reading	Proof boundary
Can route work	Some roles may classify, select, or recommend one bounded next step.	Routing is not proof and does not grant writes outside the AgentJob.
Can create draft/control artifacts	Roles may create candidates, audits, refutations, documentation, validators, or repair receipts inside an allowlist.	Output creation does not adopt source claims or prove the artifact's target.
Can modify project-control files	Project-control roles may modify allowed control surfaces inside a tracked job.	Project-control authority does not modify physics source status.
Can promote protected scientific claims	Only explicit human-gated authority can promote, close, or suspend protected claims.	The default answer for automated roles is no.
Creates proof authority	No role, schema, registry row, execution record, validator, or website page creates proof authority by itself.	Proof still depends on source mathematics, tracked approvals, and applicable governance records.

Source basis

Active authority belongs upstream.

This route explains how to read roles and schemas. It does not resolve a fresh registry snapshot, register a role, or activate a human gate.

Source area	Used here for	Authority boundary
Role contracts and role registry	Active-version resolution, role status, authority level, outputs, and human-gate state.	Registry visibility supports provenance; it is not proof or execution authority.
Schema contracts	Role, Director decision, AgentJob, execution-role, and documentation-impact artifact shapes.	Schema shape is control contract behavior, not claim adoption.
Workflow and claim-boundary records	Execution-role binding, one-job allowlist, completion evidence, and blocked-overread language.	The current job boundary governs any actual execution.

Related internal routes