Operating Principles (OS-10)
The Operating System for Reproducible AI Execution
These are not ideas. They are rules for operation. When in doubt, return to this OS. No exceptions without design.
OS-01 | Humans Own Responsibility
AI is not a decision-maker. Humans define goals, evaluation criteria, and risk tolerance.
OS-02 | Do Not Let AI “Think”
AI executes. It does not decide purpose or direction. All inputs must be structured; all outputs must be specified.
OS-03 | Prompts Are Contracts
Every generation requires a Message Contract: clear inputs, outputs, constraints, and acceptance criteria.
OS-04 | Outputs Are Deliverables
AI outputs are deliverables, not suggestions. Quality bars and acceptance rules must be defined.
OS-05 | No Validation, No Generation
If an output cannot be validated, it must not be generated.
OS-06 | Reproducibility First
One great result is irrelevant. Design for consistent quality across repeated runs.
OS-07 | Logs & Replay Are Mandatory
Inputs, contracts, outputs, models, and versions must be logged. Re-execution under identical conditions must be possible.
OS-08 | Exceptions Are Systemized
Exceptions must be designed into the system, not handled by intuition.
OS-09 | Design Is an Asset
Contracts, templates, and evaluation rules must persist as files/DB. They must survive team turnover.
OS-10 | Improve via Contracts
Failures feed back into contracts and validation. Do not fix problems with personal hacks.
Responsible Architect: Panolabo / Kenichiro Numa