Value Continuity Design
VCDesign
(Value Continuity Design)
Design principles not for "how to build", but for how to protect the created value
over time.
Why is it necessary?
AI isn't the only one crossing boundaries
I asked to "fix the UI".
The code changed. It worked.
I noticed a while later. The DB write validation had silently changed.
This isn't just about AI. The same thing happens when you ask a person.
There is no malice. They cross boundaries with good intentions.
If boundaries are not designed, anyone will cross them.
If "how far you can go" is not explicitly stated, optimization spreads endlessly.
Capability is not authority. An Agent without designed execution boundaries, responsible owners, and halt conditions becomes a source of accidents.
Choose your entry point
Start from your symptoms
From everyday situations
Give AI an "I don't know" lamp
Approval and responsibility are not the same. Understand the concept of boundaries through a story about expense approval.
→ idk-lamp.orgDevelopment design collapses
Boundary design to maximize AI
Hand the design document to AI and refine boundaries and responsibility placement through dialogue. The VC-AD YAML is the starting point.
→ VC-AD Getting StartedManagement decisions stagnate
Structure management troubles
"No money left", "No passion bubbling up". Design responsibilities and Decisions from symptoms.
→ VMSHow to Use
Bring your design documents and dialogue with AI
VCDesign is not a design document to read, but a design tool to pass to AI.
There's no need to memorize difficult concepts. Just start moving your hands.
Define value and boundaries
Articulate what you want to protect and what constitutes failure if broken. Use the Value Review Sheet to finalize Chapter Preconditions.
→ value-reviewHand design docs to AI and refine boundaries
Ask AI: "Tell me what's missing based on VCDesign." Create decisions.yaml and boundaries.yaml through dialogue. These become instructions for the AI agent.
→ VC-ADRun development with boundaries as a premise
Separate Loop A (Boundary Decision) from Loop B (Implementation by AI). Observe Δ (deviation) and select Action in Loop C.
→ TDDDCheck deviations from reality
Verify how Core and Policies connect to the actual AI operational loops.
→ Reality Fit / Adaptive Loop ModelAgent Era
Capability is not authority
VCDesign applies not only to the era of standalone AI models, but also to the era of AI agents that perform tool execution, chained operations, and long-running tasks.
As agents become more capable, decisions like "entrust it because it has capability" increase. However, an agent without designed execution boundaries, responsible owners, and halt conditions becomes a source of accidents.
5-Layer Structure (External Explanation Framework)
Model Layer
Inference, interpretation, proposal (output only, no execution)
Agent Layer
Execution boundaries, tool control, Multi-Agent coordination
Responsibility Layer
Who holds what, and who makes final decisions
Time Layer
Separation of time scales. A structure to prevent fast loops from overwriting slow loops.
Human Sovereignty Layer
Sovereignty where humans can ultimately intervene, halt, and explain
The capabilities of lower layers must not expand without the governance structure of upper layers.
See specs/core/agent-era-model.md for details.
Implementation Diagnosis
Maturity Profile
A framework to diagnose what stage an organization or system is at in each layer of AI governance. Evaluates each of the 5 layers in 4 stages.
| Level | Meaning |
|---|---|
| Unstructured | No structure. No corresponding design element exists. |
| Defined | Exists in documentation, but is not enforced in actual operation. |
| Enforced | Functions in actual operation, and controls trigger upon violation. |
| Governed | Includes auditing, improvement, and re-evaluation, and is maintained continuously. |
Diagnosis Example:
Model Layer: Enforced
Agent Layer: Defined
Responsibility Layer: Unstructured ← This is the limit of the whole
Time Layer: Defined
Human Sovereignty: Enforced
The layer at the lowest level determines the governance limit of the entire system.
See specs/validation/vcdesign-maturity-profile.md for details.
Design Philosophy
VCDesign Constitution and Manifesto
Constitution
Article 1: Continuity of Value
The purpose of the system is not the maintenance of function, but the continuity of value.
When value is impaired, the system bears the obligation to transform itself using Δ (deviation) as a starting point.
Article 2: Locality of Responsibility
For every behavior, there must be a "Chapter" that assumes responsibility for it.
Leaving ambiguous responsibility (owner_missing) unchecked means the demise of the system.
Article 3: Human Sovereignty
AI and automation aid judgment but do not hold responsibility.
Final determination of meaning and structural changes (Reframe / Retire) must always be led by a human and be explainable.
Manifesto
1. Embrace Finite Chapters
Chapters are not immutable truths. Due to divergence from reality and the accumulation of Δ, chapters will inevitably drift.
Change is not a failure, but a premise for transition (Reframe) for value continuity.
2. Celebrate Delta Signals
Δ (deviation) is the gateway to evolution. Observe discrepancies using responsibility metrics and declare drifting.
Prioritize observed anomalies over silent normality.
3. Consistency over Legacy
The Decision Log is not for reproducing the past, but for recording the "Purpose Linkage" to value.
All decisions hold accountability for the future.
4. Close or Defer
Ambiguous responsibility degrades the system. If you can solve it, Fix it. If not, Defer it.
Leaving things unchecked while holding responsibility is prohibited.
5. Retire the Meaningless
Chapters, code, and data that do not generate value must not be maintained.
Not continuing is also a design choice.
Reference Specs
Refer to Specs
Specs are not reading material, but tools for design and implementation support.
When passing to generative AI, the YAML in the GitHub repository is the shortest path.
Research & Analysis
Independent Reviews & Research Materials
Third-party research, analytical reviews, and independent evaluation materials regarding VCDesign.