# Orislop AV Joint — T4 + open-data Colab kit

This kit trains the newer Orislop joint audio–visual detector. It predicts active
speaker, lip/audio synchronization mismatch, audio spoofing, visible speaking-face
forgery, joint forgery, forged time segments, and uncertainty.

## What you do

1. Open `Orislop_AV_Joint_Open_Data_Trainer.ipynb` in Google Colab.
2. Select a T4 GPU runtime.
3. Run the setup cells. When prompted, upload `orislop-av-joint-open-data-colab-kit.zip`.
4. The notebook creates `MyDrive/orislop-av-joint/` and downloads the official
   OpenCV YuNet face detector after verifying its SHA-256.
5. Run the acquisition cell to download up to 50 decimal GB of official AMI
   close-up video and individual headset audio, or put your own rights-cleared videos under
   `MyDrive/orislop-av-joint/data/`.
6. For AMI, the notebook now pairs every present close-up AVI with its correct
   meeting-specific headset WAV and generates genuine plus controlled-offset
   records automatically. Review `rights_ami_review_required.json` and explicitly
   approve it only after you verify the license, attribution, and consent evidence.
   For other datasets, replace the generic `manifest.jsonl` and `rights.json`.
7. Run the preflight, preparation, training, evaluation, calibration, and export cells.

The trained checkpoint, reports, and exported TorchScript artifact are written to
Google Drive, not ephemeral Colab storage.

The acquisition cell is intentionally restricted to the official AMI corpus.
It does not scrape YouTube, TikTok, Instagram, LinkedIn, or arbitrary sites.
Each completed file gets a SHA-256 and provenance record, partial downloads are
resumable, unsafe redirects are rejected, and the quota is enforced before the
next file starts.

Read `OPEN_SOURCE_PATH.md` and `data_sources.json` for the model/dataset license
matrix. Acquiring authentic AMI data does not create truthful deepfake labels.
Controlled timing offsets teach synchronization mismatch only and are labeled
`joint_forgery=0`.

## Windows local acquisition

The local runner defaults to `D:\OrislopAVData`, because a 50 GB corpus plus
working space does not fit on a nearly full system drive.

```powershell
python -m pip install -r training\orislop_av_joint\requirements-acquisition.txt
powershell -ExecutionPolicy Bypass -File tools\run_local_data_pipeline.ps1 -Mode Discover
```

Discovery queries only the official directory indexes; it downloads no media and
accepts no license. After you have read the AMI terms, the complete resumable run is:

```powershell
powershell -ExecutionPolicy Bypass -File tools\run_local_data_pipeline.ps1 `
  -Mode Full -TargetGB 50 -AcceptAmiLicense
```

The full mode downloads to `D:\OrislopAVData\raw`, then writes the paired
clip manifest, mapping snapshot, and review-required rights ledger under
`D:\OrislopAVData\manifests`.

## The only inputs this kit cannot manufacture

- A project-specific legal approval of the recorded license/consent evidence.
- Rights-cleared manipulated examples from generator families you intend to detect.
- A truthful final rights ledger and reviewed labels.
- Per-face labels for active speaker, synchronization mismatch, audio spoof,
  visual forgery, joint forgery, offset, and optional forged segments.

Read `DATASET_LABELING_GUIDE.md` before labeling. Do not mark ordinary dubbing,
translation, Bluetooth delay, editing delay, or video-call lag as forgery merely
because the lips and audio are offset.

## T4 profile

The notebook defaults to batch size 1 and two data-loader workers. Dataset
acquisition and preprocessing are I/O-heavy;
the T4 only materially helps model training. Keep the raw 50 GB in Drive, but
prepare smaller reviewed clips rather than repeatedly scanning the entire raw
corpus every epoch.

## Release versus smoke mode

Release training is intentionally blocked unless the manifest contains at least:

- 100 approved speakers;
- 50 approved hours;
- no speaker or source-video leakage across splits;
- at least one held-out synthetic generator family in the test split;
- approved commercial-use, model-training, derivative, and consent evidence.

The notebook includes an explicit `ALLOW_CORPUS_SMOKE` switch for testing the
pipeline with a small corpus. Smoke artifacts always remain unpromoted and must
never be used for automatic filtering.

## Outputs

Each run writes to:

```text
MyDrive/orislop-av-joint/runs/<timestamp>/
├── av_joint_v1.pt
├── rights_report.json
├── test_metrics.json
├── temperatures.json
└── artifact/
    ├── orislop_av_joint_v1.ts
    └── orislop_av_joint_v1.json
```

Training the Phase 1 AV artifact does not by itself enable OriSlop automatic
skipping. Production promotion still requires Temporal MoE Phase 2 fusion and
calibration, independent spatial agreement, offline metric gates, and at least
10,000 reviewed shadow decisions.