DATA_DICTIONARY.md

# Data dictionary and statistical units

All numerical records are **synthetic simulations**, not microscopy, laboratory measurements, or training examples. Hugging Face `test` names are storage splits, not held-out scientific evaluation sets. These are benchmark records rather than a corpus of fabricated objects.

| Configuration | Records | Unit | Canonical source |
|---|---:|---|---|
| v3_manufacturing | 512 | One method × dimension × seed | results/gauge/runs.json |
| v2_manufacturing | 512 | One method × dimension × seed | results/runs.json |
| reserve_phase | 155 | One module count × reserve, aggregating 5,000 arrays | results/gauge/phase.json |
| claims | 20 | One scoped claim group; some group multiple labels | CLAIMS.json and THEOREM_INDEX.json |

JSONL exports flatten nested dictionaries using underscore-separated field names; variable-length traces remain in the canonical JSON. Missing values are JSON null. No number is rounded in these exports. `record_id` is an export identifier. `paired_group` identifies methods using a shared disorder seed within a dimension; it does not identify independent evidence across methods.

## Manufacturing fields

| Field | Meaning / units |
|---|---|
| dimension, nodes, modules | Spatial dimension, network vertices, scalar conductance edge modules |
| seed, method | Reproduction seed and controller/ablation label |
| complete | Controller completed the fabrication procedure; not by itself functional success |
| termination_reason | Recorded stopping cause, including infeasibility or lost reference |
| projective_function_pass | Complete object passed the declared projective/voltage specification |
| projective_error | Half-log generalized-eigenvalue condition number of grounded four-port responses; dimensionless |
| local_certificate_valid, false_local_certificate | Ground-truth audit of the local certificate, not an additional controller measurement |
| voltage_max_absolute_error | Maximum internal voltage-map discrepancy under the specified normalized drive |
| absolute_conductance_ratio | Actual-to-target absolute conductance; projective success does not require this to be 1 |
| chosen_relative_scale, physical_common_scale | Chosen representative in witness units and simulated physical scale |
| reserve_capacity, max_reserve_used | Conductance reserve in the simulator's normalized module units |
| additional_material_proxy, initial_material_proxy | Conductance-weighted material proxies, not measured mass or chemical yield |
| samples_per_pair, pair_edges, ratio_samples | Measurement allocation and total ratio sample count |
| diagnostic_radius | Simultaneous log-estimation radius used by the controller |
| maximum_numerical_log_error_bound | Checked inference numerical-error budget; ordinary floating-point implementation |
| growth_time, service_and_acquisition_time_subtotal | Abstract simulator time, not seconds |
| inference_latency_excluded_from_subtotal | Always true in this ensemble; inference and some hardware costs are not included |
| mean_cycle_residual_energy | Cycle residual diagnostic; gradient differential bias may remain invisible |
| guard_certificate_* | Flattened robust feasibility interval and stopping reason |
| config_* | Flattened v2 configuration, preserving its different objective and disorder model |

Remaining scalar names match their canonical JSON counterparts. See source comments and the manuscript for formulas. Automated field-type inventories are in schemas/data_schema.json.

## Dependence and negative results

The v3 study has 32 seeds per condition, two dimensions and eight methods. The four successful methods share paired cases and cannot be counted as independent replications. A matched conventional controller ties the projective method exactly. Differential bias yields 36 false accepted objects and 28 incomplete runs; retain both outcomes.

The phase study uses 5,000 independent arrays for each of five module counts: 25,000 sampled arrays total. Each array batch is reused at 31 reserve values. Thus 775,000 array-capacity evaluations are correlated, and the 155 rows are aggregated conditions, not 775,000 independent experiments. The curve is exact only in its ideal bounded uniform disorder model.

The preserved v2 data use a different absolute-response objective and disturbance process. Do not pool v2/v3 outcomes as a single efficacy estimate. The 28 current scientific tests include the 16 v2 tests; they are not 44 tests.