# PTC-WD-2026 data dictionary

## Shared identifiers and status fields

| Field | Meaning |
|---|---|
| `timeline_id` | SHA-256 identity of one frozen subject–property timeline |
| `version_id` | Stable timeline-local version identity |
| `source_row_index` | Position in the authoritative frozen SILVER JSONL |
| `subject_qid` | Wikidata item ID of the subject |
| `property_pid` | Wikidata property ID of the relation |
| `object_qid` / `target_object_qid` | Wikidata item ID of the dated answer |
| `statement_guid` | Exact Wikidata statement GUID |
| `human_approved` | Always false in this automatic release |
| `gold` | Always false |
| `model_eligible` | Always false; requires per-model pre-edit screening |
| `experimental_eligible` | Always false; no experimental split is published |

## `cases` (default viewer)

One row is one dated answer, presented in the conventional benchmark shape:
numeric `case_id`, readable subject, readable relation, readable answer,
validity interval, probe date, canonical question, and accepted aliases. QIDs,
statement IDs, and the SHA-256 timeline identity follow the readable fields.
Rows lacking a safe automatic prompt retain a null `canonical_question` and an
explicit `prompt_status`; they are not silently deleted.

## `timelines`

One row is a complete ordered history for one subject QID and property PID.

| Field | Meaning |
|---|---|
| `snapshot_date` | Frozen database date, `2026-07-20` |
| `structural_split` | Depth-2 control or depth-at-least-3 structural stratum |
| `chain_depth` | Number of retained dated versions |
| `distinct_object_count` | Number of distinct answer QIDs |
| `recurring_object_qids` | Answers that appear in multiple non-overlapping intervals |
| `track` | Primary chain, recurrence control, or scheduled-future diagnostic |
| `alias_status` | Whether automatic scoring aliases passed the alias gate |
| `prompt_status` | Whether prompts are available and why a row was blocked/excluded |
| `versions` | Nested compact ordered version records |

## `versions`

One row is one dated answer interval. `start` is inclusive. `end` is the
normalized upper boundary used by the construction pipeline; use the supplied
interior `probe_date` for unambiguous queries. A null `end` denotes an open
final interval at extraction time. `snapshot_current=true` means that exactly
one passing version contains the frozen snapshot date.

`accepted_aliases` are automatic English scoring surfaces. They are not human
approved. `reference_count` is the number of embedded Wikidata reference
blocks; zero does not mean the statement lacks all provenance.

## `prompts`

One row is one globally unique UTF-8 prompt for one target version.

| Field | Meaning |
|---|---|
| `prompt_id` | Stable prompt identity |
| `prompt_sha256` | SHA-256 of exact UTF-8 `prompt_text` bytes |
| `prompt_family` | Canonical, direct, paraphrase, full-interval, or bare-current |
| `temporal_order` | Whether the date occurs before/after the subject, or is absent |
| `probe_date` | Deterministic time inside the target interval |
| `accepted_aliases` | Normalized automatic answer strings without forced leading space |
| `completion_candidates` | Same answer strings with the release's completion boundary |
| `score_completion_status` | Tokenizer-specific preflight remains pending |

## `provenance`

One row retains the raw source payload for one version. Complex Wikidata
objects are stored as canonical JSON strings so the Parquet schema remains
stable and viewer-compatible:

- `mainsnak_json`;
- `qualifiers_json` and `qualifiers_order`;
- `temporal_qualifiers_json`;
- `references_json`;
- `source_statement_raw_json`.

The source entity/claim indices and source-record SHA-256 permit exact audits.

## `relations`

One row per property gives the total, depth-stratum, version, and recurrence
counts. `observed_*` fields are recomputed from the release tables and must
match the frozen relation-count file.

## `forecast_scheduled`

The five retained timelines containing a version that starts after the frozen
snapshot. They are published for transparency and future/scheduled-statement
analysis only, not current-fact evaluation on 2026-07-20.