# Dataset dictionary

Release 3.0 contains eight independently loadable JSONL configurations, 297
records in total, all under the `train` split. This split is a storage convention
and carries no machine-learning training/evaluation meaning.

All records include `id` (deterministic string), `source_path` (repository-relative
provenance) and `evidence_status` (interpretation label). No record is an original
measurement of an integrated Ørïelith skin.

| Configuration | Other fields | Types and units |
| --- | --- | --- |
| `manuscript` | `section_index`, `title`, `text`, `record_type` | Index is an integer, 0 for the cover and 1–42 for numbered sections; other fields are strings. Text retains Markdown. |
| `design_parameters` | `parameter`, `value_json` | Strings. Decode `value_json` with `json.loads` to recover the original number, boolean, string, array or object. Units are retained in source keys and values. |
| `layers` | `layer`, `z0_um`, `z1_um`, `thickness_um`, `role` | Coordinates and thickness are integer µm; other fields are strings. Nominal stack spans 0–1600 µm. |
| `formulations` | `formulation`, `ingredient`, `amount_mmol`, `MW_g_mol`, `mass_g`, `group_amount_mmol`, `qualification` | All strings, preserving mixed numeric and unresolved entries. Units are indicated in field names. `group_amount_mmol` also names the functional group. These are candidate formulation inputs, not validated syntheses. |
| `technology_compatibility` | `profile`, `present_method`, `material_interface`, `status`, `not_claimed`, `qualification` | Strings describing a proposed integration route and required qualification. |
| `frontier_comparisons` | `system`, `evidence_type`, `specific_metric`, `form_or_scope`, `recovery_or_memory_scope`, `compatibility_limit`, `primary_source` | Strings. Metric units and source scope are explicit in each row. The Ørïelith row points to its local PDF; other rows point to primary sources. No common benchmark protocol is implied. |
| `reference_model` | `model_state_z`, `model_host_kb`, `model_total_force`, `active_force`, `reference_force`, `true_translation`, `reconstructed_tether_extension`, `true_tether_extension` | Floats, all dimensionless reduced-model quantities. These names must not be interpreted as measured SI force or displacement. |
| `capture_control` | `model_time_s`, `opened_fraction_m`, `captured_fraction_r`, `g_N_m`, `receptor_force_fraction`, `Ktotal_N_m` | Floats. Time is illustrative model seconds; fractions are dimensionless; stiffness units are N/m. This is the reference-free control model, not the final paired active/reference module and not experimental kinetics. |

## Derivation and preservation

1. Split the manuscript at its explicit page markers, preserving each block
   except leading/trailing whitespace. Extract the existing heading as `title`.
2. Preserve each top-level design JSON value in a `value_json` string. Nested
   structures are not flattened or rounded.
3. Convert the six source CSVs row for row. Cast only layer dimensions to integers
   and the two model tables to floats. Retain all formulation and comparison
   fields as strings, including empty strings and unresolved values.
4. Add deterministic IDs, provenance and evidence labels. Write UTF-8 JSONL
   without NaN or infinity.

`dataset_schema.json` gives exact row counts and Python-level primitive types.
`tools/build_dataset.py` reproduces these transformations from `research/`.
Original CSV and JSON files remain available there. Primary-source evidence
boundaries are in the full manuscript; the short per-record labels are not a
substitute for those qualifications.

## Missing values and suitable use

An empty string means the source table did not specify that field. Text such as
`verify`, `derive from verified MW`, and `not specified` is meaningful and must
not be silently replaced by a numerical default. The dataset is suited to
inspection, blueprint indexing, calculation reproduction and planning a
qualification experiment. It does not support empirical skin-performance
prediction or claims about subjective android sensation.