Testing Generative Engineering Representations: A Protocol and Worked Experiment for Minimal Stem Sets in Abstract Structural Synthesis

Protocol and complete executable research object establishing the experimental tractability, rather than general validity, of compact engineering representations (“stems”) in abstract structural synthesis. Two open-weight four-bit constructor strata received all eight subsets of three candidate stems across four development and four held-out-labelled goal families. The deposit includes a deterministic evaluator, eleven calibration fixtures, the exact 384-packet plan, model and runtime configurations, every captured response and extracted structure, 10,176 predicate observations, descriptive analysis, figures, notebooks, and prepared output-language and stem-form follow-ups. All scheduled principal runs were collected, but no run met the conjunctive structural-adequacy endpoint. The object also contains a separately frozen 96-run post hoc capability control using the same Qwen3-8B revision on a reduced single-authority problem. All 48 held-out control responses passed the output contract; compact-stem and direct-checklist conditions each reached 12/12 adequacy, compared with 6/12 under neutral and deliberately corrupted guidance. The principal treatment contrasts are exploratory because calibration exposed packets assigned to the held-out set, and the control is a task-relative calibration rather than a confirmatory reinterpretation of the principal null result. A further 192-run post hoc powerset divides the successful authority guidance into three fixed candidate components. D1+D3 is the unique sufficient held-out minimum, but the complete set is not sufficient because D2 introduces paired regressions. The deposit therefore reports a minimal basis and interference rather than strict distillation from a successful whole. The archive preserves the limitations and the complete evidential chain.