docs/data_specification.md

# Data specification

This release preserves all archived simulation labels and numeric tensor values.
Identity fields use the companion code's portable schema namespace.

## Coordinates and sets

For a rectangular scene of width W and height H, local metric coordinates are centered
on the scene. Normalized coordinates are u = x/W + 0.5 and v = y/H + 0.5.
Normalization is per scene; 0.1 in a 512 m scene is 51.2 m, not 12.8 m.
Z remains in meters. TX coordinates have shape [N, 3]; normalized XY has shape [N, 2].
N is variable and must not be inferred from padding. Set order is not a physical label.
Preserve alignment among candidate indices, positions, and per-TX quantities.

Arrays use [row_y, column_x]. Raster row zero is at minimum local y; plotting
the raster uses a lower origin. The cached scene tensor has shape [11, H, W].

## Scene channels

The channel order is:

1. `building_mask`
2. `road_mask`
3. `green_mask`
4. `water_mask`
5. `obstacle_mask`
6. `built_up_mask`
7. `explicit_open_mask`
8. `unknown_feature_mask`
9. `background_mask`
10. `tx_feasible_mask`
11. `evaluation_mask`

The source cached features are uint8 masks. Cached features and candidate coordinates
are bundled with each core dataset, separately from deployment rows.
They must not be duplicated in every deployment row.

**Keep three meanings separate:** `tx_feasible_mask` constrains placement; the cached
`evaluation_mask` is a scene-context feature (some source metadata calls this
all-non-water cells); `road_evaluation_mask` in a Radio Map identifies the actual
coverage denominator. Do not replace the eleventh trained-model input channel with a
road mask during export. In the inspected generator, labels use the source road mask,
resized with nearest-neighbor sampling to the radio grid.

## Coverage labels

All coverage fields are fractions in [0, 1]. Percent presentation multiplies by 100.
For example, joint coverage equals `joint_covered_count / evaluation_cell_count`.
Thresholds are inclusive. The existing `joint_target_met` flag uses a configured 0.9
joint target; this is distinct from the four per-cell metric thresholds and is not
a claim that all deployments attain 90%.

Path Loss is lower-is-better; SS-RSRP, SINR, throughput and coverage are higher-is-better.
SS-RSRP and effective throughput are configured simulation proxies. Throughput uses
the reference efficiency, resource share and spectral-efficiency cap rather than an
end-to-end protocol-stack measurement. Numerical parameters are in the RF profile.

## Optional Radio Maps

| Array | Shape | Unit / interpretation |
| --- | --- | --- |
| `path_loss_db` | [H, W] | dB |
| `ss_rsrp_dbm` | [H, W] | dBm |
| `sinr_db` | [H, W] | dB |
| `effective_throughput_mbps` | [H, W] | Mbps |
| `serving_tx_index` | [H, W] | Serving-TX index; validate masked/invalid cells |
| `road_evaluation_mask` | [H, W] | Coverage evaluation cells |
| Four metric `*_covered` masks | [H, W] | Per-metric threshold satisfaction |
| `joint_covered` | [H, W] | Joint threshold satisfaction |

Non-finite propagation values outside meaningful support are not zero-valued
measurements. Apply the evaluation mask and documented metric handling before
computing summaries. Per-TX path gains, when present in a source file, are optional;
inspect each NPZ array list rather than assuming their presence.

## Families and split safety

The deployment table contains unique physical deployments. The family-membership
table is a many-to-many relation: membership ID, family ID, sample ID, scene/split,
master sample, subset indices, role, strategy and simulation seed. A deployment's
primary family field does not enumerate all memberships. Joining memberships and
then averaging rows can double-count a deployment.

Train/validation/test assignments are fixed at scene/geographic-group level. An
adaptation extension has its own protocol and provenance; training on a test scene
changes the evaluation setting, even when only a fraction of its labels is used.

## Portable identity and integrity

Contract, sample, family, membership and parent/master references are migrated
together to the companion code's schema namespace. Physical positions, label
values, splits, channel order and numeric tensor payloads are unchanged.
Source-session timestamps, local file paths and unused contributor/contact tags
are not part of the distributed metadata. Geographic snapshots retain physical
classification/height tags and geometry. No propagation was rerun for export.

The 18 tutorial rows use the same migrated sample IDs as the complete deployment
table. Their `source_split` describes the primary benchmark assignment, while
the tutorial configuration itself is called `preview/examples`.

`metadata/checksums.parquet` records every payload's bytes and SHA-256 except
itself and the mutable transfer-status file. `radio_maps/index/*.parquet` maps
each available sample to a TAR shard, NPZ member and original payload checksum.
The NPZ payloads inside those shards are byte-identical to their source arrays.

## Adaptation extension

`extensions/reward_adaptation_targets` contains 3,590 unique deployment rows and
3,600 family memberships for paris_11 (128 m, 2 TX), new_york_12 (256 m, 4 TX),
and tokyo_01 (512 m, 8 TX). It has no archived Radio Maps. Its historical split
names label the three target scenes; they are not the within-target few-shot
train/validation/test selections. Select each target and construct the adaptation
protocol explicitly. Never merge these rows into the primary benchmark training
split: tokyo_01 is held out in the primary benchmark.