B-MAN MANIFESTO Human-Controlled AI-Assisted Development Version 1.0

B-MAN is a judgment framework for working with AI in software development.

It is not a tool, not a workflow, and not a promise of speed or automation. It exists to preserve human control, ownership, and accountability in a world where AI systems can produce convincing, fast, and incorrect output.

B-MAN does not ask: “How do we let AI do more?” It asks: “How do we prevent humans from losing control?”

AI systems will produce:

• partial understanding
• hidden assumptions
• confident mistakes
• locally correct but globally dangerous changes

Therefore, correctness cannot be trusted by default. Human judgment must remain the final authority.

Humans are the sole decision makers.

AI may propose, generate, explain, or optimize, but it never makes binding decisions.

Responsibility for outcomes cannot be delegated to AI. If a human cannot justify a decision independently, the decision must not be accepted.

Human control is a hard constraint, not an optimization goal.

A human may only own artifacts they fully understand.

Code, configuration, designs, and decisions must be:

• explainable without AI assistance
• readable by humans
• modifiable by humans

Artifacts that are opaque, magical, or accepted on trust violate ownership and must be rejected.

Learning the system is not a side effect. It is a safety requirement.

AI operates as an executor, not a conversational partner.

Work is performed asynchronously on a predefined task. There is no live chat steering during execution.

The workflow is strictly separated: define → execute → judge

This prevents scope drift, emotional momentum, and accidental delegation of judgment.

Resetting work must be psychologically and technically cheap.

Clarification

Asynchronous, task-bound execution applies to the execution of an explicit task.

Conversational interaction with AI is allowed during exploration, ideation, and problem framing, where no binding artifact or decision is produced.

Once a task becomes explicit and intent-bound, execution must be asynchronous and non-interactive to preserve human control and prevent decision drift.

Each task produces one atomic, intent-bound artifact.

An artifact represents a single, clear purpose and must stand on its own.

Artifacts must be independently:

• reviewable
• rejectable
• revertible

The artifact may be code, documentation, design, specification, migration, or decision record.

In code work, a common implementation is: one task → one commit but this is an implementation detail, not the principle.

AI does not infer intent. All intent must be stated explicitly.

Every task must define:

• its purpose
• its scope (what is in / what is out)
• expected outputs
• constraints and boundaries

Implicit assumptions are forbidden. Silent decisions are unacceptable.

B-MAN uses priority-ordered (lexicographic) optimization.

Primary objective (non-negotiable):

• human control
• explicit decision points
• artifact ownership
• explainability
• reset capability

Secondary objectives (allowed only if the above hold):

• speed
• cost
• throughput
• efficiency
• convenience

Optimizations that weaken primary objectives are forbidden, even if they increase productivity.

Optimizations that improve secondary objectives without harming control are allowed.

The following are expected consequences of the core principles, not independent rules:

• Reset is a first-class operation
• The system is designed for failure
• Changes are small and reviewable
• Transparency is favored over trust
• AI self-reports assumptions, decisions, and uncertainties
• Humans remain accountable for all outcomes

From Ownability ⇒ Comprehensibility follows a strict requirement:

Artifacts must be small and focused enough to allow thorough human review.

An artifact that cannot be realistically reviewed end-to-end by a human within reasonable time cannot be owned, and therefore must not be accepted.

This requirement is not about process preference. It is a safety constraint.

Large or multi-intent artifacts silently undermine ownership by turning review into sampling instead of understanding.

Code and artifacts must be optimized for human readability.

Readability is not a stylistic choice. It is a prerequisite for ownership.

AI systems may benefit indirectly from clear structure, explicit naming, and simple control flow, but such benefits are secondary effects.

Optimizing code primarily for AI consumption, compression, or cleverness at the expense of human understanding violates the Ownability principle.

Human-first readability is mandatory. AI-friendliness is a permissible side effect, not a goal.

Default assumption: All existing code and configuration are invariants until proven otherwise by conscious human decision.

In high-risk systems (security, money, isolation, SLA, compliance), critical invariants should be made explicit as first-class, human-owned artifacts.

AI may propose invariant changes. Humans must explicitly approve them.

This is an advanced safety mechanism, not a core requirement.

B-MAN is tool-agnostic.

External methodologies (e.g. requirement frameworks, checklists, templates, automation tools) may be integrated as secondary process optimizations.

Such integrations are allowed even if they do not improve human judgment directly, as long as they do not:

• bypass human decision points
• obscure responsibility
• weaken explainability
• reduce reset capability

B-MAN is the control kernel. Other methods are plugins.

B-MAN is not:

• autonomous AI development
• prompt engineering ideology
• a guarantee of correctness
• a replacement for human judgment
• a workflow optimized solely for speed

B-MAN assumes that both AI systems and humans will make mistakes.

AI systems make mistakes due to limited understanding and hidden assumptions. Humans make mistakes due to cognitive bias, fatigue, pressure, and misplaced trust.

B-MAN exists to make both kinds of mistakes visible, bounded, and recoverable.

B-MAN is designed not only to preserve human control, but to keep the human creator at the center of the development process.

By offloading bounded execution to AI working asynchronously, engineers regain cognitive space to think, design, and write.

The goal is not to turn humans into supervisors of machine output, but to keep them active creators who understand, choose, and build.

Human well-being, sense of ownership, and joy of creation are not side effects. They are essential for sustained responsibility and judgment.

B-MAN does not optimize AI autonomy. It optimizes human ownership.

AI is treated as a powerful, fast, and fallible tool. Humans remain responsible for every accepted artifact.

If a system cannot be confidently owned by a human, it must not be accepted—no matter how impressive the AI output.

This is the core of B-MAN.