data/data_dictionary.md

# Data dictionary — dataset v1.0.0

Every exported artifact of this synthetic dataset, column by column: type, role, range,
allowed values and derivation rule (S4). Written in English, like every other artifact
in the project (S1).

The dataset contains no real patient data. `patients`, `telemetry_vitals` and
`centro_servizi_tickets` are generated from the fixed seed below; `operator_notes` is the
only generated text and it comes from a versioned cache, not from a live model call (S22).

## Dataset identity

| field | value |
| --- | --- |
| `dataset_version` | `1.0.0` |
| `seed` | `42` |
| `generation_timestamp` | `2026-07-01T00:00:00Z` (fixed, never the wall clock) |
| `operator_notes.model` | `qwen3:1.7b` |
| `operator_notes.model_digest` | `8f68893c685c3ddff2aa3fffce2aa60a30bb2da65ca488b61fff134a4d1730e7` |
| `operator_notes.prompt_version` | `notes_prompt_v1_en` |
| `operator_notes.cache_path` | `data/fixtures/notes_cache_qwen3-1.7b_v1.json` |

## How to read the tables

| column kind | meaning |
| --- | --- |
| role `id` | identifier, unique within its table; foreign keys are named in the derivation |
| role `time` | ISO 8601 UTC; every timestamp is derived, never the wall clock |
| role `feature` | usable as a model input |
| role `target`, `target_source`, `target_auxiliary` | labels; never use them as features (S5) |
| role `protected` | protected attribute, used only for fairness checks (S14) |
| role `split` | dataset split assignment |
| role `metadata` | descriptive, not a clinical signal |
| range / allowed values | the contract; `—` means unrestricted within the declared type |

## `patients.csv`

- entity: `patients`
- rows: `500`
- format: CSV, UTF-8 without BOM, LF, comma delimiter, header row

| column | type | role | range | allowed values | derivation rule |
| --- | --- | --- | --- | --- | --- |
| `patient_id` | string | `id` | P-0001..P-0500 | — | Seeded permutation of the id range. One row per patient and the id never appears in more than one split (S5). |
| `age_years` | integer | `demographic` | 18..90 | — | Uniform integer draw inside the declared range from the seeded per-patient generator; no clinical meaning is attached to age. |
| `biological_sex` | string | `protected` | — | F, M, O | Exact counts F 230 / M 230 / O 40 (S15) assigned by seeded permutation, so the marginals are exact rather than approximate. |
| `insurance_tier` | string | `protected` | — | public, basic, premium | Exact counts public 180 / basic 170 / premium 150 (S15) assigned by seeded permutation independent of biological_sex. |
| `split` | string | `split` | — | train, validation, test | Two-stage largest-remainder apportionment per sex x tier cell, then within each cell by seeded permutation; totals are exactly 350 / 75 / 75. |

## `telemetry_vitals.csv`

- entity: `telemetry_vitals`
- rows: `6000`
- format: CSV, UTF-8 without BOM, LF, comma delimiter, header row

| column | type | role | range | allowed values | derivation rule |
| --- | --- | --- | --- | --- | --- |
| `reading_id` | string (UUIDv5) | `id` | UUIDv5, documented namespace | — | UUIDv5 in the fixed namespace declared in creator_dataset.constants over 'patient_id|timestamp_utc' (S26); recomputable from the row. |
| `patient_id` | string | `id` | P-0001..P-0500 | — | Foreign key to patients.patient_id; 12 readings per patient (S19). |
| `timestamp_utc` | datetime (ISO 8601 UTC) | `time` | 2026-01-01T00:00:00Z .. 2026-06-24T11:00:00Z | — | t0(patient) + 1 h per reading index, where t0 is a seeded date inside S25's window at 00:00 UTC, so the internal 24 h trajectory always ends before 2026-06-30. |
| `heart_rate_bpm` | float, 1 decimal | `feature` | 30.0..220.0 | — | The patient's allocated point value selects an S8 band, and the value is sampled inside that band with a rounding margin (D1). |
| `systolic_bp_mmhg` | float, 1 decimal | `feature` | 50.0..260.0 | — | Sampled inside the S8 band chosen by its own point value and inside the interval allowed by diastolic_bp_mmhg (D3); systolic > diastolic + 10 holds by construction, in integer tenths. |
| `diastolic_bp_mmhg` | float, 1 decimal | `feature` | 20.0..150.0 | — | Sampled first among the two pressures, because its point value decides the systolic interval (D3). |
| `spo2_pct` | float, 1 decimal | `feature` | 70.0..100.0 | — | Sampled inside the S8 band chosen by its allocated point value. |
| `resp_rate_rpm` | float, 1 decimal | `feature` | 5.0..50.0 | — | Sampled inside the S8 band chosen by its allocated point value. |
| `body_temp_c` | float, 1 decimal | `feature` | 30.0..43.0 | — | Sampled inside the S8 band chosen by its allocated point value. |
| `device_id` | string | `metadata` | DEV-001..DEV-100 | — | One primary device per patient, assigned deterministically by creator_dataset.telemetry.device_map(); 25 test patients get the v1.3 pool (S29). |
| `device_firmware` | string | `metadata` | — | v1.2, v1.3 | Fixed per device by device_firmware_map(); v1.3 appears only in the test split because the device pool decides it, not a per-row check (S5, S29). |
| `battery_level_pct` | float, 1 decimal | `metadata` | 0.0..100.0 | — | Seeded bounded random walk per patient reading; never scored. |
| `signal_noise_index` | float, 3 decimals | `metadata` | 0.0..1.0 | — | Device metadata only (S29): v1.3 draws from a clearly higher mean than v1.2, and the six vitals are never perturbed by it in v1.0.0. |
| `quality_flag` | string | `data_quality` | — | OK | Always OK: this table contains only rows valid for normal use, which is why invalid shapes live in telemetry_edge_cases.csv (S5). |
| `news2_score_current` | integer | `target_source` | 0..18 | — | Sum of the six point values from the S8 table, and independently recomputable from the six exported vitals, which is what S28's recompute check asserts. |
| `clinical_severity_current` | string | `target_source` | — | Low, Med, Critical | S6 thresholds on news2_score_current: Low 0-4, Med 5-6, Critical >= 7. |
| `label_window_start` | datetime (ISO 8601 UTC) | `metadata` | — | — | Equal to timestamp_utc by definition (S19). |
| `label_window_end` | datetime (ISO 8601 UTC) | `metadata` | — | — | timestamp_utc + 6 hours, exactly (S19). |
| `future_max_score_6h` | integer | `target_source` | 0..18 | — | Maximum score over the 24 internal trajectory points that fall inside (timestamp_utc, timestamp_utc + 6h], so it is never below the current score. |
| `future_clinical_severity_6h` | string | `target_auxiliary` | — | Low, Med, Critical | S6 thresholds applied to future_max_score_6h. |
| `deterioration_risk_6h` | integer | `target` | — | 0, 1 | 1 if future_max_score_6h >= 7 or future_max_score_6h - news2_score_current >= 3; 0 otherwise. The two rules are evaluated independently (S11). |
| `critical_risk_6h` | integer | `target` | — | 0, 1 | 1 if future_clinical_severity_6h == Critical; 0 otherwise. |

## `telemetry_edge_cases.csv`

- entity: `telemetry_edge_cases`
- rows: `50`
- format: CSV, UTF-8 without BOM, LF, comma delimiter, header row

| column | type | role | range | allowed values | derivation rule |
| --- | --- | --- | --- | --- | --- |
| `edge_case_id` | string | `id` | EC-001..EC-050 | — | Sequential id in the fixed case-class order of S20. |
| `description` | string | `metadata` | — | — | Short English description of the defect the row carries. |
| `field_under_test` | string | `metadata` | — | heart_rate_bpm, systolic_bp_mmhg, diastolic_bp_mmhg, spo2_pct, resp_rate_rpm, body_temp_c | The one vital the row breaks, except in the cross-field class where both pressures are individually legal. |
| `heart_rate_bpm` | string (raw) | `feature` | raw cell, may be invalid on purpose | — | Raw text, not a float, so the fixture can carry NaN, an empty cell, non-numeric text and out-of-range numbers (S20). |
| `systolic_bp_mmhg` | string (raw) | `feature` | raw cell, may be invalid on purpose | — | Raw text; also carries the cross-field violation against diastolic. |
| `diastolic_bp_mmhg` | string (raw) | `feature` | raw cell, may be invalid on purpose | — | Raw text; legal in the cross-field class, where only sys > dia + 10 fails. |
| `spo2_pct` | string (raw) | `feature` | raw cell, may be invalid on purpose | — | Raw text, including the empty-cell case. |
| `resp_rate_rpm` | string (raw) | `feature` | raw cell, may be invalid on purpose | — | Raw text, including the non-numeric case. |
| `body_temp_c` | string (raw) | `feature` | raw cell, may be invalid on purpose | — | Raw text, including the boundary case just outside the valid range. |
| `expected_action` | string | `metadata` | — | reject, flag_invalid, clamp_with_flag | reject for every class except cross_field_invalid_sbp_dbp, which is flag_invalid because both cells are individually legal. clamp_with_flag is declared by the contract and unused in 1.0.0. |
| `notes` | string | `metadata` | — | — | Short English note naming the class the row belongs to. |

## `centro_servizi_tickets.jsonl`

- entity: `centro_servizi_tickets`
- rows: `500`
- format: JSONL, UTF-8 without BOM, LF, one JSON object per line

| column | type | role | range | allowed values | derivation rule |
| --- | --- | --- | --- | --- | --- |
| `ticket_id` | string | `id` | TCK-00001..TCK-00500 | — | Sequential id in export order; unique. |
| `created_at` | datetime (ISO 8601 UTC) | `time` | 2026-01-01T09:00:00Z .. 2026-06-29T09:00:00Z | — | 2026-01-01T09:00:00Z + index * 179 / 499 days, so the window is a property of arithmetic and no ticket sits on the closing second of S25's range. |
| `category` | string | `metadata` | — | device_alert, connectivity, battery, calibration, software, user_error, alarm, other | Hamilton apportionment over the declared weights, then a seeded shuffle, so every allowed value is present deterministically (D8.1). |
| `priority` | string | `metadata` | — | low, medium, high | Same allocation scheme; weights 35 / 40 / 25, all three values present. |
| `device_id` | string | `metadata` | DEV-001..DEV-100 (may be empty) | — | Seeded draw over the shared device pool; omitted for a small share of the user_error and other categories, where a report can arrive without a device (D8.2). |
| `firmware_version` | string | `metadata` | — | v1.2, v1.3 | Equal to device_firmware_map()[device_id] whenever a device is named, which is how S29's shared device -> firmware map cannot drift; a device-less ticket takes the fleet default v1.2. |
| `ticket_subset` | string | `split` | — | train, golden | Exact counts train 350 / golden 150 assigned by seeded shuffle (S21). |
| `operator_notes` | string | `feature_nlp` | 20..120 words, English | — | Read from the versioned cache by ticket_id + prompt_version; generated once by qwen3:1.7b under the frozen L17 configuration with temperature 0, seed 42 and thinking disabled. The generator fails if a key is missing and never regenerates silently (S22). |
| `golden_flag` | boolean | `metadata` | — | true, false | true if and only if ticket_subset == golden (S21). |
| `review_status` | string | `metadata` | — | not_required, pending_review | pending_review for golden rows, not_required for train rows (S21). |

## `audit_traceability_log.csv`

- entity: `audit_traceability_log`
- rows: `50`
- format: CSV, UTF-8 without BOM, LF, comma delimiter, header row

| column | type | role | range | allowed values | derivation rule |
| --- | --- | --- | --- | --- | --- |
| `audit_id` | string | `id` | AUD-0001..AUD-0050 | — | Sequential id in export order; unique. |
| `event_timestamp` | datetime (ISO 8601 UTC) | `time` | 2026-07-01T00:00:00Z .. 2026-07-01T00:49:00Z | — | 2026-07-01T00:00:00Z + 60 s per row (S25), never the wall clock. |
| `entity_type` | string | `metadata` | — | patients, telemetry_vitals, telemetry_edge_cases, centro_servizi_tickets, manifest, dataset | Fixed quota per value (8 / 12 / 8 / 8 / 8 / 6 = 50) interleaved round-robin, per D7. |
| `entity_id` | string | `metadata` | — | — | The audited file name; dataset_v1.0.0 on the six dataset-level rows, where the audited artifact is the export directory (D7). |
| `actor` | string | `metadata` | — | dataset_generator | Constant: the generator is the only actor in this pipeline. |
| `action` | string | `metadata` | — | create, validate, export, checksum | Fixed quota per value (14 / 12 / 10 / 14 = 50) laid out over the entity types by the 6x4 matrix in creator_dataset.audit.ACTION_MATRIX (D7). |
| `dataset_version` | string | `metadata` | — | 1.0.0 | Constant dataset version. |
| `artifact_path` | string | `metadata` | — | — | outputs/dataset_v1.0.0/<file> for file-level rows, and the export directory outputs/dataset_v1.0.0/ for dataset-level rows (D7). |
| `artifact_sha256` | string (hex) | `metadata` | 64 hex characters, or empty | — | sha256 of the audited file. Empty on the rows that reference manifest.json, because the log is written before the manifest (D6) and never hashes itself (S23). Dataset-level rows carry the digest of the five exported tables. |
| `requirement_id` | string | `placeholder` | REQ-PLACEHOLDER-001..050 | — | Unique sequential placeholder. Requirements belong to the consuming project, so no real requirement is invented (S23). |
| `test_name` | string | `placeholder` | TEST-PLACEHOLDER-001..050 | — | Unique sequential placeholder; no real test name is invented (S23). |
| `requirement_status` | string | `metadata` | — | pending_reconciliation | Constant until the consuming project reconciles the placeholders. |
| `result` | string | `metadata` | — | pass | Constant: the pipeline fails loudly instead of writing a failing audit row. |
| `evidence_path` | string | `metadata` | — | — | Equal to artifact_path, except on validate rows where it is outputs/dataset_v1.0.0/validation_report.json (D7). |
| `notes` | string | `metadata` | — (may be empty) | — | Optional column, empty in 1.0.0: nothing is invented to fill it. |

## Notes generation (S22)

`operator_notes` is generated once, validated, and then read from the versioned cache
`data/fixtures/notes_cache_qwen3-1.7b_v1.json`. The generator reads the cache for every row and fails loudly if a
key is missing; it never regenerates a note silently.

| parameter | frozen value |
| --- | --- |
| model | `qwen3:1.7b` |
| model digest | `8f68893c685c3ddff2aa3fffce2aa60a30bb2da65ca488b61fff134a4d1730e7` |
| prompt version | `notes_prompt_v1_en` |
| length rule | 20–120 words |
| request options | `temperature=0.0`, `top_k=1`, `top_p=1.0`, `repeat_penalty=1.0`, `seed=42`, `num_ctx=2048`, `num_predict=220` |
| language | English only |
| cache key | `ticket_id|prompt_version` |

The cache, not the model, is the reproducibility contract: two identical Ollama calls are
not bit-reproducible, so a rerun must reuse the cache instead of comparing generations.

## `manifest.json`

Not a table, so it is documented field by field. The manifest hashes the five data files,
the audit log and this dictionary, and never itself (S29).

| field | type | role | range | allowed values | derivation rule |
| --- | --- | --- | --- | --- | --- |
| `dataset_version` | string | `metadata` | — | 1.0.0 | Constant dataset version (S4). |
| `seed` | integer | `metadata` | 42 | — | The single seed every random decision derives from (S13, S24). |
| `generation_timestamp` | datetime (ISO 8601 UTC) | `time` | — | 2026-07-01T00:00:00Z | Fixed by S24; the wall clock is never used. |
| `notes.model` | string | `metadata` | — | qwen3:1.7b | The Ollama model that produced operator_notes (S22, S24). |
| `notes.model_digest` | string | `metadata` | 64 hex characters | — | Digest of the model as installed, recorded because a tag can be re-pushed. |
| `notes.prompt_version` | string | `metadata` | — | notes_prompt_v1_en | Names the prompt config file the notes were generated with (S24). |
| `notes.cache_path` | string | `metadata` | data/fixtures/notes_cache_qwen3-1.7b_v1.json | — | Path of the versioned notes cache, which is an input and not an output (D5). |
| `notes.cache_sha256` | string | `metadata` | 64 hex characters | — | sha256 of the notes cache file (S24). |
| `tables` | object | `metadata` | — | — | One entry per exported table with its file path, its exact row count and its column list, read from the produced file (S4). |
| `files` | object | `metadata` | — | — | sha256 and byte size of the five data files, the audit log and the data dictionary. The manifest never hashes itself (S29, D6). |

## Invariants that apply to every table

- No real patient data, no PHI, no personal names and no real institutions anywhere.
- Every timestamp is ISO 8601 UTC and derived from a fixed date; the wall clock is never used.
- Splits are assigned per `patient_id`, and no `patient_id` appears in more than one split (S5).
- `telemetry_vitals` and `telemetry_edge_cases` are separate on purpose: the first contains
  only rows valid for normal use, the second contains only invalid or boundary cases (S5).
- Label and future columns are targets and must never be used as model features (S5).
- CSV artifacts are UTF-8 without BOM with LF endings and fixed decimal places;
  `centro_servizi_tickets` is JSONL (S4, S27).