research-document
Independent Replication Plan
Independent Replication Plan
Purpose
Define replication levels for the major Framework Engineering experiments.
Replication Levels
- same model, different prompt
- different model family
- human analyst
- domain expert
- blinded independent reviewer
FE-008 Identity-Capability Classification
- Same model, different prompt: Re-run classification with alternative prompt wording and reduced FE terminology.
- Different model family: Compare identity-capability outputs across model families.
- Human analyst: Use human reviewers with the same KACS and FEMS framing.
- Domain expert: Include experts familiar with selected artifacts but not with FE.
- Blinded independent reviewer: Review characterization records without seeing the originating method.
FE-011A Redesign Comparison
- Same model, different prompt: Re-run participant and evaluator prompts with tighter wording controls.
- Different model family: Separate participant and evaluator model families.
- Human analyst: Ask human practitioners to apply original and redesigned packets.
- Domain expert: Use domain specialists to judge whether traceability gains are meaningful.
- Blinded independent reviewer: Evaluate outputs without provenance and against non-FE structured baselines.
FE-012A Primitive Extraction
- Same model, different prompt: Repeat extraction with altered prompting and no seed-vocabulary priming where possible.
- Different model family: Compare extracted primitives and merges across model families.
- Human analyst: Use independent human extractors with merge-disagreement logging.
- Domain expert: Ask domain experts whether extracted sequences distort artifact meaning.
- Blinded independent reviewer: Review normalization decisions and reconstruction outputs without knowing FE expectations.
FE-012B Primitive Synthesis
- Same model, different prompt: Re-run synthesis on the same novelty set with varied prompt controls.
- Different model family: Compare synthesis coherence and missing-primitive pressure across models.
- Human analyst: Ask human analysts to synthesize from the frozen primitive set only.
- Domain expert: Ask domain experts whether synthesized procedures remain coherent under realistic domain pressure.
- Blinded independent reviewer: Evaluate synthesized frameworks for coherence, gaps, and genericity without knowing their origin.