| ADR-001 | Modular component architecture | Accepted | Enables independent testing, replacement, and parallel development of merge, evolution, and evaluation subsystems | Increased interface complexity; requires strict contract definitions | None |
| ADR-002 | No MergeKit lock-in | Accepted | Prevents dependency on a single external library whose roadmap may diverge from EMEP needs | Must re-implement core merge strategies; full control over tensor operations | None |
| ADR-003 | Safetensors as primary artifact format | Accepted | Memory-mapped loading, fast I/O, no pickle execution risk, security scanning compatible | All model I/O must convert to/from safetensors; legacy pickle checkpoints require migration | None |
| ADR-004 | ed25519 artifact signing | Accepted | Modern, fast, widely supported elliptic-curve signatures for provenance and supply chain integrity | Key management infrastructure required; verification mandatory in enterprise mode | None |
| ADR-005 | Offline-first enterprise mode | Accepted | Satisfies air-gapped deployment requirements for regulated and secure environments | All external dependencies (Hub, package indexes) must be mirrorable; updates are manual | None |
| ADR-006 | Hidden Test Set isolation | Accepted | Prevents overfitting and benchmark contamination by strictly separating optimization and final evaluation data | EvaluationEngine must enforce split boundaries; operators cannot override | None |
| ADR-007 | Evolution non-optimizing against test data | Accepted | Extends ADR-006: evolutionary search uses only Optimization Set and Validation Set | FitnessEngine must reject Hidden Test Set metrics; any violation aborts experiment | None |
| ADR-008 | Semantic versioning for Enterprise Models | Accepted | Communicates breaking changes, compatibility guarantees, and release stability to downstream consumers | Version bump policy required; automated checks against API surface changes | None |
| ADR-009 | Evidence-tagged documentation | Accepted | Distinguishes established research from hypotheses, assumptions, and unvalidated results, reducing reader confusion | Every claim must carry a tag; untagged claims are flagged in review | None |
| ADR-010 | Mermaid as primary diagram format | Accepted | Plain-text, version-controllable, renderable in docs and GitHub, no proprietary tool dependency | Complex layouts may require manual adjustment; limited interactive features | None |
| ADR-011 | Python 3.11 as reference runtime | Accepted | Match target deployment environment; balance new features with stability | Components must run on 3.11; newer syntax is prohibited | None |
| ADR-012 | PyTorch as reference tensor engine | Accepted | De facto standard for open-weight models; extensive ecosystem for CUDA, quantization, and export | Ties EMEP to PyTorch release cadence; alternative backends require abstraction layer | None |