Blame for docs/data-dictionary.md

# Data dictionary

## Images and ImageFolder rows

`data/test/metadata.jsonl` has one row per image. `file_name` is relative to
`data/test/`; the manifest uses paths relative to the repository root.

| Field                         | Meaning                                                   |
| ----------------------------- | --------------------------------------------------------- |
| `image_id`                    | Stable string identifier; join key across native files    |
| `category`                    | `synthetic` or `trap`; both contain generated images      |
| `width`, `height`             | Original raster dimensions in pixels                      |
| `plausibility`, `naturalness` | Final ordinal ratings, integers 1–5                       |
| `no_violation_detected`       | True only for a usable negative                           |
| `method_id`                   | Ground-truth resolution method, e.g. `GT-004`             |
| `objects.category_id`         | List of integer class IDs, 1–97                           |
| `objects.bbox`                | Parallel list of normalised `[x, y, width, height]` boxes |

The loader exposes `image` as a decoded PIL image. Empty object lists are valid
negative examples. There is one evaluation `test` split, no training split.
`release/image_manifest.jsonl` adds the original image SHA-256 and byte count.

IDs and categories are strings; dimensions and class IDs are integers. Boxes
contain four floating-point values. Ratings are integers, and
`no_violation_detected` is a boolean. The two `objects` lists have the same
length: elements at the same index describe one instance. `category` records the
source grouping, not the final positive/negative judgement.

## Native annotations

`annotations/ground_truth.json` is an array of records with `image_id`, `kind`,
`gold`, `base_annotator`, `source_annotators`, `adjudicated_by` and
`edit_summary`. `gold.violations` contains `use_case_id` (`U01`–`U97`) and
normalised XYWH `bbox`. It also contains the quality ratings and usable/negative
state. The complete resolution provenance is in `edit_summary.resolution`,
including `method_id`, name and version. `kind` alone does not identify how gold
was obtained.

`gold.reported` is false for every released final annotation. A usable negative
has an empty violation list and `no_violation_detected: true`; a positive has
one or more violations and `no_violation_detected: false`. Each violation
retains its class name. Bounding boxes have positive width and height, with all
edges inside the full-image range [0, 1].

`agreement_at_gold` is an optional historical agreement snapshot, not a new
comparison of final labels. `base_annotator` and `adjudicated_by` can be null;
actor references are opaque pseudonyms, not sortable identities. For rating
meanings, use the [annotation rubric](annotation-protocol.md#the-rating-rubric).

`annotations/first_pass_annotations.json` is an array joined by `image_id`.
`first_pass` holds original per-annotator payloads; `n_first_pass` is their
count. Actor IDs remain pseudonyms. Administrative text, personnel timestamps
and local replay paths are nulled; free-form notes are emptied. Class labels,
box coordinates, ratings, state, record order and provenance identifiers are
preserved. The original local files have different hashes because these
administrative fields were removed. Source hashes are retained in provenance.

First-pass records also retain their original agreement snapshot and state
flags. A reported or unfinished submission must not be inferred to be a clean
negative merely because it has no boxes. `n_first_pass` is the actual number of
submissions; do not assume exactly two. Null administrative fields mean that
information is not distributed, rather than that the event never occurred.

## COCO compatibility

COCO stands for Common Objects in Context, a widely used object-detection
annotation format. `ground_truth_coco.json` is a derived compatibility export:

- `images[].id` is a deterministic integer starting at 1, in sorted stable-ID
  order. `images[].image_id` retains the native string join key.
- `images[].file_name` is relative to the dataset repository root.
- `annotations[].image_id` references that integer image ID.
- `category_id` 1 maps to `U01`, through 97 mapping to `U97`.
- `bbox` is **pixel** XYWH. Multiply native x/width by image width and native
  y/height by image height. Coordinates are not rounded.
- `area` is pixel width × height, and `iscrowd` is 0. Segmentation is absent.

All images, including negatives with no annotation rows, occur in `images`. COCO
does not carry the full native rating and resolution provenance; use the native
file when those fields are required.

## Results and provenance

`evaluation/dataset/` holds aggregate composition, first-pass agreement and
final rating distributions. `evaluation/models/synthetic11534/results.json`
holds valid/common scopes, coverage and per-class metrics; CSV is a summary.
Prediction JSONL stores `image_id`, `model_name`, `latency_ms`, `parse_status`
and `detections`. Each detection uses an integer `category_id` (1–97), a
normalised XYWH `bbox` and, where available, a `score`. Retained scores are
evidence and do not control evaluation.

`release/checksums.sha256` covers every publication file except itself.
`release/provenance.json` identifies the release, licence and original source
filenames and hashes. Source hashes identify the original artifacts before
administrative fields were removed; the checksum inventory covers the
distributed files.

## Worked example

Suppose an image is 1,000 pixels wide and 800 pixels high. A native violation
with `use_case_id: "U03"` and `bbox: [0.10, 0.20, 0.30, 0.25]` is litter with
its top-left corner at pixel (100, 160), width 300 and height 200. The
equivalent COCO box is `[100, 160, 300, 200]`; its bottom-right corner is (400,
360). This example is illustrative rather than an additional dataset record.

To trace a real example, take `image_id` from a loaded row, find the same key in
both native annotation arrays, and use `release/image_manifest.jsonl` to locate
its image. Join COCO annotations through the COCO integer `images[].id`; that
integer is not the native string ID. Class integer 3 maps to `U03` in every
representation.

A class can occur in several boxes on one image. Keep those instances for
localisation; deduplicate only the class IDs for image-level classification.