Lightcap/pcmt-artifact

Proof-Carrying Multimodal Timelines Artifact

This Hugging Face Dataset repository hosts the runnable artifact for:

Proof-Carrying Multimodal Timelines: Finite-Trace Modal Certificates for Video-Audio Consistency

Authors: Faruk Alpay and Hamdi Alakkad.

The artifact is organized as a dataset-style file tree rather than a zip archive. It is intended to support an arXiv submission whose source package stays below arXiv's upload limit while keeping the full runnable code, traces, certificates, logs, and figures available in one public location.

Keywords: finite-trace modal logic, LTLf-style monitoring, multimodal trace semantics, video-audio consistency, audio-visual synchronization, active speaker consistency, subtitle-video alignment, certificate checking, counterexample extraction, perturbation sweeps, logical defect metrics, YouCook2, AVE, AVA ActiveSpeaker, TVQA, ActivityNet Captions.

What This Contains

The contribution is the logic and certificate layer:

• finite multimodal trace semantics for synchronized video, audio, and subtitle/OCR atoms;
• modal/temporal consistency formulas for speech-speaker, audio-visual event, subtitle-video, scene continuity, and edit-induced temporal shock checks;
• deterministic certificate checking and counterexample extraction;
• logical defect metrics under counterfactual multimedia edits;
• GPU-accelerated extraction and sweep execution where available, with CPU fallback paths.

This is not a GPU benchmark repository. The GPU is used as an execution engine for detector-scale evidence and perturbation sweeps.

Repository Layout

pcmt/                         Core Python package: atoms, logic, schema, pipeline, plots
tests/                        Smoke and unit tests for schema, logic, edits, audit, pipeline
scripts/                      Packaging, plotting, remote GPU, and artifact upload helpers
docs/dataset_adapters.md      Notes for YouCook2, AVE, AVA ActiveSpeaker, TVQA, ActivityNet adapters
data/fixtures/                Tiny fixture-backed videos and annotations for tests
data/real_videos/             Small public MP4 clips used for real-video artifact paths
runs/smoke/                   Minimal ingest -> atoms -> trace -> certificate outputs
runs/run-small/               Small defect-atlas run
runs/real-video/              Real-video case-study traces and certificates
runs/remote_gpu/              Larger GPU-generated detector and sweep outputs
runs/robustness-atlas/        Certificate-stability and robustness tables
figures/                      Manuscript-grade figures and figure CSV values

Large certificate directories are sharded as certificates/shard_XXX/ to preserve individual JSON certificates while avoiding repository directory-size limits.

Quick Start

Clone with Git LFS enabled, because MP4 videos and generated figure assets are stored as LFS objects:

git lfs install
git clone https://huggingface.co/datasets/Lightcap/pcmt-artifact
cd pcmt-artifact

Install the package in editable mode:

python3 -m pip install -e ".[dev]"

Run the main verification commands:

make smoke
make test
make run-small
make figures

Expected generated outputs include:

runs/<name>/traces.csv
runs/<name>/windows.jsonl
runs/<name>/certificates/*.json
runs/<name>/logs/*.jsonl
figures/*.pdf
figures/*.png

Required Trace Schema

The artifact emits formula-level rows with the fields used in the manuscript:

video_id
time window
visual atoms
audio atoms
subtitle/OCR atoms
edit type
formula verdict
certificate hash
counterexample window
defect score

Certificates are JSON records whose verdict, violating index set, defect score, and first counterexample window can be recomputed by the checker from the serialized trace and formula identifier.

Dataset Adapters

The code includes adapters for small subsets of:

• YouCook2
• AVE
• AVA ActiveSpeaker
• TVQA
• ActivityNet Captions

If the real datasets are not locally available, the test and smoke paths use documented synthetic fixtures while preserving the same ingest, trace, certificate, and plotting code paths.

Check adapter availability with:

python3 -m pcmt.cli adapter-status

Reproducibility Notes

The arXiv source package contains only the manuscript source, bibliography, and figures. This Dataset repository contains the larger artifact tree. README and license-style files are intentionally not part of the arXiv ancillary source package, but this README is included here so the Dataset repository is self-describing.

The artifact does not include access tokens, SSH keys, environment dumps, or private credentials.

Citation

@misc{alpay2026pcmt,
  title        = {Proof-Carrying Multimodal Timelines: Finite-Trace Modal Certificates for Video-Audio Consistency},
  author       = {Alpay, Faruk and Alakkad, Hamdi},
  year         = {2026},
  note         = {arXiv artifact},
  howpublished = {\url{https://huggingface.co/datasets/Lightcap/pcmt-artifact}}
}