# Mechanical Softwering Method — Draft 0.1

The cycle is iterative: **Intent → Drawing → Reference → Decomposition → Fabrication → Inspection → Assembly → Bench Testing → Real Testing → Acceptance → Release → Feedback → Drawing**.

| Phase | Objective | Inputs | Activities | Output | Human responsibility | AI role | Done when | Main risks |
|---|---|---|---|---|---|---|---|---|
| 1. Intent | Define purpose, user and expected result | Need, context, constraints | Frame value, exclusions and harm | Product Intent | Own purpose and priority | Organize evidence, expose ambiguity | Outcome and exclusions are accepted | Vague purpose, hidden stakeholders |
| 2. Drawing | Make architecture, interfaces, tolerances and acceptance explicit | Intent, standards, constraints | Model boundaries, contracts, qualities | System Drawing | Choose design and tolerances | Explore options, check consistency | Interfaces and criteria are reviewable | Premature design, missing qualities |
| 3. Reference | Recover behavioral knowledge | Prior systems, data, users, records | Observe, measure, characterize, classify | Behavioral Reference Map | Decide what is authoritative | Extract cases and contrasts | Sources and confidence are recorded | Copying defects, misreading behavior |
| 4. Decomposition | Create buildable, verifiable units | Drawing and reference map | Split by contract and integration seam | Component Contracts | Preserve system coherence | Suggest seams and dependencies | Each unit has bounded scope and test | Coupling hidden by neat boundaries |
| 5. Fabrication | Produce a scoped component | Contract, Fabrication Brief | Implement, test locally, record assumptions | Candidate component | Authorize scope and protected areas | Fabricate code and evidence | Brief is met; assumptions declared | Hallucination, scope creep, plausible code |
| 6. Inspection | Establish conformity before integration | Candidate, diff, logs | Review code, provenance, assumptions, security | Inspection record | Judge conformity and corrections | Explain diff, run static checks | Findings are closed or accepted | Automation bias, superficial review |
| 7. Assembly | Integrate without breaking contracts | Inspected components | Merge incrementally, resolve interfaces | Integrated increment | Control interface and architecture | Assist integration and conflict analysis | System builds and contracts hold | Semantic mismatch, hidden shared state |
| 8. Bench Testing | Produce controlled, repeatable evidence | Integrated increment, fixtures | Automated tests, simulation, differential checks | Bench Test Report | Define coverage and interpret results | Execute, compare, classify failures | Required checks pass with evidence | Overfitting, weak oracle, false confidence |
| 9. Real Testing | Validate in operational conditions | Bench-qualified build, hardware/users | Exercise representative workflows and edges | Real Test Report | Ensure safety and realistic protocol | Capture logs, assist analysis | Required real scenarios have results | Unsafe tests, unrepresentative environment |
| 10. Acceptance | Make an explicit go/no-go decision | Evidence, residual risks | Review criteria, deviations and readiness | Acceptance Record | Sole acceptance authority | Summarize evidence, never decide | Named human signs decision and limits | Schedule pressure, unclear authority |
| 11. Release | Promote a verified candidate | Accepted increment, checklist | Package, version, communicate, rollback-plan | Release candidate/product | Authorize promotion | Automate reproducible packaging | Checklist complete and identity fixed | Configuration drift, irreversibility |
| 12. Feedback | Turn findings into better design | Field evidence, defects, decisions | Update drawing, contracts and registers | Revised drawing and backlog | Decide what changes the design | Cluster evidence, propose revisions | Learning is traced to an artifact | Knowledge loss, patching without learning |

The cycle may loop at any phase. A failed test returns to the earliest artifact whose assumption it disproves. Completion is evidence-based, not a fixed amount of agent activity.