---
title: ReplayForge
emoji: ⏪
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 6.20.0
app_file: app.py
python_version: "3.12"
pinned: false
license: apache-2.0
short_description: Turn a screen recording into an evidence-linked bug report
models:
  - microsoft/Mage-VL
  - microsoft/VibeVoice-ASR-BitNet
tags:
  - video-understanding
  - software-testing
  - bug-reporting
  - multimodal
startup_duration_timeout: 1h
---

# ReplayForge — The Multimodal Bug Time Machine

**Upload the crash. Reconstruct the truth.**

ReplayForge turns a screen recording into an editable, evidence-grounded incident report. It extracts a bounded
set of timestamped frames, reads visible UI text locally, optionally transcribes narration locally, asks Mage-VL
to reconstruct only the events supported by the recording, and links each claim back to an evidence frame.

## What it exports

- Markdown bug report
- Structured timeline JSON
- Automatically redacted evidence frames
- Annotated evidence replay
- ZIP evidence bundle (never the original recording)

## Privacy and safety

- Uploaded content is processed inside this Space and is never sent to a third-party media service.
- The original recording is not placed in the downloadable bundle.
- Temporary uploads and generated files expire automatically.
- Application logs contain numeric timing/count metadata, not prompts, transcripts, OCR text, filenames, or media.
- Visible or spoken instructions in a recording are treated as untrusted evidence, never as commands.
- Credential-like OCR is removed before model analysis, and generated text is sanitized again before export.
- Masked password fields are never interpreted as revealing a password value.
- Automatic PII detection is best-effort. Do not upload credentials, private customer data, or confidential
  production footage to a public demo.

## Limits

- 60 seconds and 100 MB per recording
- H.264 MP4 is the most reliable input
- Hidden cursors, tiny text, rapid transitions, and cropped dialogs can make actions ambiguous
- Low-confidence claims must be reviewed before filing the report
- Narration that looks silent, repetitive, or degenerate is discarded instead of being treated as evidence
- The generated report is an AI-assisted draft, not a verified fact record

## Models and attribution

- [`microsoft/Mage-VL`](https://huggingface.co/microsoft/Mage-VL), pinned at
  `5c78cab61938e73859b63724d9bf5cb88c477eaa` (Apache-2.0)
- [`microsoft/VibeVoice-ASR-BitNet`](https://huggingface.co/microsoft/VibeVoice-ASR-BitNet), pinned at
  `66e78021ab8f5f06133d1ab421ba4d348bda97c9` (MIT)
- VibeASR.cpp pinned at `70b3ebb8ad75b5f37aee948df34f15cc84951d05`

Mage-VL loading and codec preprocessing are derived from Microsoft's Apache-2.0 reference Space. ReplayForge's
pure-PyTorch `mamba_ssm` compatibility shim is also retained because Transformers validates Mage-VL's optional
StreamMind import even though ReplayForge does not load that gate. ReplayForge's workflow, evidence model, privacy
controls, report generator, and interface are original project work.