XRefKit | Operating model and reference implementation

Carry AI work through
to business acceptance.

AI can finish a local task, but it does not automatically preserve business completion, decision ownership, or the state required by the next responsibility.

XRefKit is an operating base that connects AI work from a Goal to human acceptance with verifiable progression. Managed Skills, Knowledge, and execution contracts can be distributed to multiple users through MCP.

1. How XRefKit operates internally

The runtime responsibility boundary that connects one AI work unit from Goal definition to human acceptance.

What the model shows

The model is not primarily a task sequence. It shows where decisions belong when work is delegated to AI. Before execution, people define business completion and acceptance conditions as a Goal. During execution, a Skill bounds the AI responsibility. Decisions outside that boundary return to people through Escalation instead of being filled by assumption.

The workflow protocol records progression and evidence. Closure deterministically checks whether the execution record is complete, but it does not decide whether the result is correct or should be adopted. Accountability returns to people through human acceptance. Decisions, unresolved items, and sources remain as Evidence and can be traced through XIDs.

Evidence is input to human improvement decisions. XRefKit does not let AI rewrite organizational rules automatically from execution results.

Read the canonical definitions of the 12 model elements

2. How the operating base is shared

MCP delivers the same managed Skills, Knowledge, and execution contracts to AI clients used by multiple people.

Open the XRefKit MCP distribution model

What MCP shares

An organization maintains Skills, Knowledge, execution contracts, catalogs, and XIDs as managed XRefKit assets. MCP presents the catalogs, then selects and resolves only the assets needed for the current purpose. Users receive the same identifiers and managed versions without loading the entire repository into every AI context.

MCP is a read boundary for discovery, selection, resolution, and delivery. Skill execution, client-side tool side effects, Evidence recording, and closure remain on each user's client. People can therefore share the same Knowledge and procedures without merging their execution permissions or work records.

MCP shares managed definitions and Knowledge. The MCP server does not perform business work on behalf of its users.

Choose by purpose

Start from the canonical model, then read only the depth required for the current question.

1

Understand the need and the model

See which AI characteristics require an operating structure and learn the canonical elements.

2

Operate AI work

Understand how bounded responsibilities connect and where tool, AI, and human decisions belong.

3

Implement, improve, and distribute

Connect the repository structure to evidence-led improvement and distribution of the next version.

Canonical concepts

Permanent concept pages define the problems XRefKit is intended to solve and connect them to the operating model.

Permanent problem definition

Judgment Replication

Organizations do not scale by adding processing capacity alone. They scale by replicating judgment—and traditional organizations do this slowly, lossily, and repeatedly. XRefKit replaces judgment retransmission with stable, versioned reference distribution.

End-to-End Example: csharp_review

A bounded review responsibility for C# and system-boundary problems that Roslyn diagnostics alone cannot confirm.

1. Goal Define scope, desired state, and acceptance conditions. The target state has a result for every review category and makes every unconfirmed item explicit.
2. Semantic routing Select csharp_review from the purpose of reviewing C# beyond compiler diagnostics.
3. Skill and Knowledge Follow the Skill procedure and load only the needed parts of the C# review criteria (XID: 30E6A4F6F3AA) and common source analysis criteria (XID: 5F21C8A41001).
4. Review execution Establish Roslyn diagnostics as the baseline, avoid duplicate findings, then inspect language-specific behavior, lifecycle boundaries, exception paths, and uncertainty.
5. Evidence Record category results, finding severity, rationale, remediation direction, judgments, unconfirmed items, and the next owner.
6. Workflow protocol xrefkit skill verify deterministically checks work items, evidence, unresolved concerns, and role separation. It does not judge whether the review content is correct.
7. Quality and handoff An independent quality review returns blocked, needs-review, or proceed. Implementation, security, and design findings move to the corresponding Skills.
8. Human acceptance A person reviews the Evidence and verdict, then accepts the result, returns it for repair, or selects the next bounded responsibility.