# Cardinality / Range Dynamic Inventory

## 1. Related files found

| Path | Type | Granularity | Main fields | Used? | Why |
| --- | --- | --- | --- | --- | --- |
| `README.md` | markdown | project-level | benchmark scope, evaluation layers, model roster, repo layout | yes | Defines project framing and final scientific takeaway. |
| `Paper\69b27219c555c38a69bb2156\sections\introduction.tex` | tex | paper-section | evaluation narrative, scoring interpretation, main-text phrasing | yes | Aligns the analysis language with the paper draft. |
| `Paper\69b27219c555c38a69bb2156\sections\evaluation_methodology.tex` | tex | paper-section | evaluation narrative, scoring interpretation, main-text phrasing | yes | Aligns the analysis language with the paper draft. |
| `Paper\69b27219c555c38a69bb2156\sections\evaluation.tex` | tex | paper-section | evaluation narrative, scoring interpretation, main-text phrasing | yes | Aligns the analysis language with the paper draft. |
| `Paper\69b27219c555c38a69bb2156\sections\appendix_scoring_standard.tex` | tex | paper-section | evaluation narrative, scoring interpretation, main-text phrasing | yes | Aligns the analysis language with the paper draft. |
| `doc/synthetic_data_scoring_protocol_v0_4.md` | markdown | metric-level | step-by-step validation protocol, applicability rules | yes | Cross-checks the implementation against the documented scoring standard. |
| `src/evaluation/synthetic_validation_v4.py` | python | metric-level | exact score logic, per-column detail fields, channel composition | yes | Primary source for the exact discrete and continuous formulas. |
| `Evaluation\validation\runs\20260426_145322\summaries\validation_details__all_datasets.jsonl` | jsonl | dataset-model asset with column-level detail | dataset_id, model_id, cardinality_range discrete/continuous details, per-column real/synthetic support and ranges | primary | Main analysis source because it preserves the per-column structures needed for dynamic buckets. |
| `Evaluation\validation\runs\20260426_145322\summaries\validation_summary__all_datasets.csv` | csv | dataset-model asset | dataset_id, model_id, cardinality_range_score and other validation scores | secondary | Used for schema cross-checking and run-level coverage confirmation. |
| `Evaluation\validation\runs\20260426_145322\manifest.json` | json | run-level | dataset_count, asset_count | secondary | Used to select the highest-coverage validation run. |
| `logs/runs/ (legacy/v1 benchmark and grounded-SQL artifacts)` | directory | grounded-SQL run artifacts | query_results.jsonl, run manifests, traces | context only | Confirms repository positioning as workload-grounded, but not used for cardinality/range metric values. |

## 2. Primary analysis source

- Selected run: `20260426_145322`
- Details file: `Evaluation\validation\runs\20260426_145322\summaries\validation_details__all_datasets.jsonl`
- Coverage before deduplication: `513` assets across `51` datasets
- Coverage after deduplication and alias normalization: `496` dataset-model assets across `51` datasets and `14` normalized models
- Column-level analysis units built: `33815`
- Duplicate normalized dataset-model pairs resolved: `17` dropped assets

## 3. Field availability

| Source | dataset id/name | model/generator | discrete/continuous channel | score | real distinct | synthetic distinct | real min/max | synthetic min/max | column name | overall validation score |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| validation_details__all_datasets.jsonl | yes | yes | yes | yes | yes (discrete per-column) | yes (discrete per-column) | yes (continuous per-column) | yes (continuous per-column) | yes | yes |
| validation_summary__all_datasets.csv | yes | yes | no | channel-level only | no | no | no | no | no | yes |

## 4. Field mappings used by the loader

- `asset.model_id` -> normalized `model` via alias map `{rtf: realtabformer}`.
- `report.validation_channels.cardinality_range.details.discrete_profile.per_column[].real_distinct` -> `real_distinct_count`.
- `report.validation_channels.cardinality_range.details.discrete_profile.per_column[].syn_distinct` -> `synthetic_distinct_count`.
- `report.validation_channels.cardinality_range.details.continuous_profile.per_column[].real_min/real_max` -> `real_min` / `real_max`.
- `report.validation_channels.cardinality_range.details.continuous_profile.per_column[].syn_min/syn_max` -> `synthetic_min` / `synthetic_max`.
- `report.validation_channels.cardinality_range.discrete_profile_score` / `continuous_profile_score` remain attached as official channel-level references.

## 5. Files used only for cross-checking or context

- Older validation runs under `Evaluation/validation/runs/20260426_100004`, `20260426_100103`, `20260426_100210`, and `20260426_105754` were excluded because they only cover 1--2 datasets.
- The run-level summary CSV is not rich enough for dynamic-bucket analysis because it lacks per-column support and range envelopes.
- Synthetic CSVs under `SynOutput/` and `SynOutput-5090/` were not read directly because the unified validation details already materialize the necessary real-vs-synthetic comparisons.

## 6. Schema inconsistencies and missing fields

- Model alias inconsistency: `rtf` appears alongside `realtabformer`; the analysis normalizes both to `realtabformer`.
- The full validation run contains duplicate `dataset × normalized_model` assets for 17 pairs, mostly across `SynOutput` vs `SynOutput-5090`; the analysis keeps the newest generation timestamp and records dropped duplicates.
- Coverage is incomplete for several generators even in the highest-coverage run, so the effective panel is not a complete `51 × 14` matrix.
- Windows-style paths are embedded in the unified evaluation outputs, but they are treated as metadata strings rather than local file dependencies.

## 7. Coverage gaps that matter for interpretation

- The normalized roster still reaches all 14 target generator names, but some generators have much smaller dataset coverage than the benchmark maximum.
- `codi`, `forestdiffusion`, and `tabdiff` are especially sparse, so their high-dynamic conclusions should be read as conditional on smaller sample sizes.

## 8. Final decision for the main analysis

- Use `validation_details__all_datasets.jsonl` as the main source.
- Use column-level units whenever possible.
- Keep discrete and continuous channels separate until the final cross-channel summaries.
- Treat the discrete unit score as a derived support-retention view built directly from official per-column support-loss counts; note this explicitly in the report.