Canonical operating model
Chapter 2: Model
Defines how Goals, bounded AI responsibilities, deterministic control, escalation, handoff, and human acceptance connect.
How to read the model
People first define the Goal and acceptance conditions. AI selects a Skill suited to the current state and performs one bounded responsibility. The Skill loads only the required Knowledge and records work, outputs, judgments, and unresolved items as Evidence. The workflow protocol deterministically verifies the execution record and invokes a separate quality review when required. A decision that exceeds the current authority returns to a person through Escalation. Moving work to another responsibility uses Handoff to transfer outputs and unresolved state. Closure checks progression completeness but does not decide business adoption. People make the final acceptance decision against the Goal. XIDs trace each definition and Evidence item so an improved version can be distributed under stable identifiers.
The Judgment Replication concept page defines the organizational problem this model addresses: distributing organization-specific judgment by stable reference rather than repeatedly retransmitting it through people or anonymous similarity retrieval.
Goal
Holds the desired business state and acceptance conditions. Finishing an individual Skill does not complete the Goal while its acceptance conditions remain unmet.
Semantic routing
Selects the next required Skill from the Goal, current state, and Skill responsibility definitions. AI performs this semantic candidate selection separately from deterministic verification.
Skill
A bounded responsibility with a purpose, judgment boundary, inputs, outputs, required Knowledge, and completion conditions. It narrows what AI handles at one time and exposes out-of-scope decisions.
Knowledge
Organization-specific facts, criteria, and judgment perspectives. The runtime selects required XIDs from a catalog and loads only the bodies needed by the current responsibility.
Workflow protocol
Records work items, outputs, concerns, roles, progression state, and handoffs for each responsibility execution. It makes omitted work detectable without replacing the Skill's domain judgments.
Evidence
Traceable outputs, rationale, unconfirmed items, risks, and used XIDs. Evidence supports resumption, review, and improvement decisions; it is not an instruction for automatic revision.
Verification and quality review
xrefkit skill verify deterministically checks work records, role separation, Evidence, and concern states. A separate quality review evaluates output validity when required.
Escalation
Returns authority, exception, risk, or irreversible decisions that the current responsibility cannot make to a person. It differs from Handoff, which transfers work to the next owner.
Handoff
Transfers outputs, Evidence, unresolved items, and next ownership to another Skill or person. It remains separate from Escalation, which returns decision authority.
Closure
Checks whether execution, verification, required quality review, and handoff records are complete. Closure does not automatically declare business adoption or Goal achievement.
Human acceptance
Evaluates outputs and Evidence against the Goal's acceptance conditions, then accepts, returns, stops, or routes the work onward. This is the accountable business decision point.
XID
A stable identifier for Skills, Knowledge, contracts, and Evidence. It keeps the managed object traceable when file locations or distribution paths change.