Skip to content

AI-Centered Development with Accountable Humans

inspequte is an ongoing experiment to explore how far an AI-centered development workflow can go while keeping humans accountable.

This is not about letting AI decide everything.
It is about separating judgment from judgment criteria.

The core hypothesis is:

If judgment criteria are explicitly written, version-controlled, and traceable, AI can execute the workflow and humans can remain responsible.


Project Philosophy

The project is built on three principles.

1. Criteria Over Intuition

Traditional development relies heavily on human intuition during:

  • planning
  • specification writing
  • implementation
  • review
  • acceptance

In an AI-centered workflow, intuition must be replaced with explicit, testable criteria.

Every decision must be:

  • documented in text
  • machine-readable
  • reproducible

AI does not understand intent.
It executes written constraints.

Therefore, the responsibility of humans shifts from:

making decisions repeatedly

to:

designing and maintaining decision criteria.

2. Full Traceability

The workflow is structured so that every step produces artifacts:

  • planning documents
  • specifications
  • generated implementation
  • validation results

Each stage references the previous one, forming a traceable chain of reasoning.

If an implementation is accepted, the evidence remains.
If rejected, the cause is explicit.

This ensures that AI execution is auditable and reversible.

3. Mechanical End-to-End Flow

The goal is to make the full process executable without human intervention:

  1. Planning
  2. Specification
  3. Implementation
  4. Validation

If validation passes, the output is ready for final acceptance.
If validation fails, the implementation is routed back to the specification stage and revised.

Humans intervene only at the final acceptance step.

They review:

  • the generated feature set
  • documentation
  • validation results

Then decide to accept or reject.

Verified Achievements

The following have been successfully validated in controlled experiments.

End-to-End Automation

From feature idea to validated implementation, the entire process can be executed mechanically when clear criteria are provided.

This includes:

  • generating structured plans
  • deriving specifications
  • producing implementation code
  • verifying conformance against specifications
  • automatically rejecting non-conforming outputs

The system is capable of iterative correction loops without manual intervention.

Specification-Driven Validation

Verification is not stylistic review.
It is specification conformance checking.

When validation fails, the system identifies mismatches and routes the implementation back for revision.

This establishes a closed-loop quality control mechanism.

Human Role Redefined

In this workflow, humans:

  • define the planning principles
  • document constraints and quality targets
  • evaluate final deliverables

They do not manually supervise intermediate steps.

This model preserves accountability while reducing repetitive judgment labor.

Model Capability Observations

Experiments indicate that workflow reliability is strongly correlated with model capability.

Higher-capability models demonstrate significantly improved:

  • instruction fidelity
  • multi-step reasoning consistency
  • specification adherence

This suggests that the conceptual architecture is viable, and model advancement is a key enabler.

Current Scope

This experiment does not claim that real-world production systems can immediately adopt this model.

However, it demonstrates that:

  • AI-centered development is structurally feasible
  • accountability can be preserved through explicit criteria
  • end-to-end mechanical execution is achievable under controlled conditions

The limiting factor is no longer automation technology.
It is the explicitness and quality of human-defined standards.

inspequte continues as an exploration of a development model where humans design principles, AI executes processes, and responsibility remains transparent.