# Ground truth for DRMHB researchers

## Four different kinds of information

**Measured input** is raw acceleration/velocity information and measurement
context: units, sample rate/frequency grid, point, mode, speed and timing.
**Construction truth** records the simulator's process, health and nuisance
states. **Waveform truth** explains how a captured source became a stored crop.
**Diagnostic truth (DiagGT)** carries diagnostic documents, observations,
findings and materialization rules. These are complementary, not interchangeable.

| Where | Use |
|---|---|
| `obs` | Snapshot identity and measurement context |
| `X`, `var` | Explicit signal-derived feature values and definitions |
| `obsm['raw_waveforms']`, `raw_spectra` and companions | Reusable raw signals, real lengths, positions, speeds and tachometer edges |
| `obsm['ground_truth']` | Snapshot-aligned construction labels |
| `obsm['waveform_ground_truth']` | Per-snapshot `records_json` with zero or more channel-bound waveform labels |
| `uns['vibframe_evaluation']` | Original evaluation files, byte-exact, with source/path/size/SHA-256 |
| `uns['vibframe_anndata']` | Source, conversion, alignment, channel and feature provenance |

## Import all the information

```python
from vibframe_anndata import import_raw_to_h5ad
base = import_raw_to_h5ad(
    "data/DRMHB-compact.vibframe.zip", "derived/DRMHB-compact.v0.3.0.raw.h5ad",
    config={
        "version": 1,
        "raw_import": {"on_missing_signal": "nan"},
        "ground_truth": {"enabled": True, "scope": "all",
                         "on_missing": "error", "max_sidecar_mib": 512},
        "output": {"dtype": "float32", "raw_compression": "none"},
    }, block_size_mib=8,
)
```

Full scope retains regular files under `evaluation/`, `ground-truth/` and
`ground_truth/`, root JSON/YAML documents and machine JSON context. This includes
unknown future annotation fields and files. Files outside those locations are
not indiscriminately copied. The evaluation archive is not a duplicate waveform
archive. Raw `trends.parquet` is never a source of production feature values.

## Identity and time

Use `(source, machine_id, snap_t)` for snapshot labels. Times are integer
microseconds since Unix epoch, UTC. Never infer the unit from the magnitude.
Waveform labels additionally bind point and available mode/definition/config
identifiers. Missing identity, duplicated bindings and ambiguous modes must not
be fixed by row-order or nearest-time guessing.

A source acquisition starts at `snap_t`; a stored crop starts at `t`:

```text
t = snap_t + crop offset
```

New imports retain `raw_waveforms_capture_t`, `raw_spectra_capture_t` and the
corresponding `*_known` matrices on the channel axis. Always check real lengths
and the known flags: an absent capture is not a date, and a legacy fallback is
not an explicitly measured time. Tachometer edges are relative to the stored crop.

## Read labels without loading signals

```python
from vibframe_anndata import (
    get_snapshot_ground_truth, get_waveform_ground_truth, list_evaluation_files,
    read_evaluation_table, read_evaluation_json,
)
path = "derived/DRMHB-compact.v0.3.0.features.h5ad"
snapshot = get_snapshot_ground_truth(path)
waveform = get_waveform_ground_truth(path, point_id="pump_DE_H", proc_mode_id="ACC_10K")
files = list_evaluation_files(path)
# Pick an exact source/path from files before calling a reader.
rows = files.loc[files["path"].str.endswith("observations.parquet")]
if not rows.empty:
    item = rows.iloc[0]
    diagnostics = read_evaluation_table(path, item["path"], source=item["source"])
```

Convenience views can stringify nested fields for HDF5 compatibility. The table
reader recovers the original Arrow schema; the byte reader verifies the original
hash by default. A selected table is materialized, but unrelated raw buffers are
not loaded. Avoid reading the entire `uns` group just to inspect metadata.

## Protocol decisions remain explicit

Normality training must not automatically consume fault identity, severity,
physical state, nuisance/OOD flags, simulation seeds or future labels. A known
operating state is an oracle input only under an explicitly declared experiment.
Do not infer health from missing DiagGT. Consolidated diagnoses and intervals
must be interpreted using their original document/materialization semantics.

The aligned views follow observation slicing, but the original archive remains
dataset-wide. Features added/recalculated/removed by the package preserve both.
Keep independently generated paired controls out of the longitudinal training
and evaluation splits; use a separate H5AD/source and pair identities.

## Existing H5ADs

Example 04 wraps `add_ground_truth_to_h5ad()`. It adds originals and aligned
annotations to an existing file without recomputing features or source raw
signals. Keep `obs.source`, `obs.machine`, `obs.snap_t` and the original source
set. The transactional copy requires additional disk space. It does not recreate
legacy capture-time companion arrays; new raw imports do. Original waveform
truth still provides offsets/durations for audit.

## Resource accounting

The H5AD grows by original sidecar bytes, aligned labels, the manifest and timing
companions. `list_evaluation_files()` reports exact original sizes. The 512 MiB
budget refers to encoded sidecars, not total RAM. No raw compression is changed.
The current longitudinal ZIP is 7,635,324,956 bytes. Measured complete H5ADs are
8,771,810,437 bytes (base) and 8,874,970,373 bytes (features), preserving 97
original files totaling 26,874,052 bytes. The [current audit](provenance/current_dataset.json)
verifies their source annotation bytes, identities and waveform lengths, plus
complete raw-buffer equality between base and feature files. The current original
snapshot-truth schema has 49 columns. Campaign windows are 0.25/0.4/0.75/1.6 s;
paired controls use 0.5 s. `fault_present` and `diagnostic_eligible` are distinct;
no RUL or failure-time label is defined.

Authoritative API and full coverage rules:
https://413hq.github.io/vibframe-anndata-docs/0.3.0/guides/ground-truth/