Repository navigation
Releases
DriftSync v2.2.0 - Mathematics Workspace
Download and run
Download DriftSync-v2.2.0-windows-x64.zip below, extract the entire ZIP into a writable folder, and run DriftSync.exe. Keep the _internal folder next to it. Python is bundled. The sidebar displays DriftSync 2.2.0 so you can confirm you opened this version.
This is a new portable app download. Existing copies do not update automatically; open the executable from the newly extracted folder. Keep your old recordings when upgrading.
New Mathematics workspace
Open Mathematics in the sidebar, or press Alt+7. Five pages combine properly typeset equations with interactive worked examples:
- Prediction: sequence windows and the future-error target; toggle outcomes to see the label change.
- Behaviour: robust normalization and rolling error rates, with a numerical example.
- Sequence models: switch between LSTM memory and masked Transformer attention diagrams.
- Uncertainty: compare dropout distributions and see how mean risk, sample standard deviation and warning thresholds relate.
- Learning: explore weighted binary cross-entropy and calibration error with adjustable probability and outcome.
Examples are clearly labeled as illustrative and identify their implementing source files. Mouse and keyboard controls work across the five topics, including smaller windows.
Also included
- The complete desktop redesign: shared charcoal/teal styling, resizable windows, keyboard navigation, redesigned task setup, training, results and live analysis.
- Live risk history with threshold and uncertainty band, plus explicit unavailable calibrated-confidence labeling.
- Fresh-install fix: live mode now discovers checkpoints created by named in-app training runs when no default checkpoint exists.
- A repeatable Windows packaging script, packaged-app self-test and SHA-256 checksum.
Use Train models before live predictions. Task recording and the Mathematics section work immediately. Personal sessions and pretrained checkpoints are not included; training uses synthetic data by default.
Verification
- 69 automated tests passed, plus feature/model smoke checks.
- Visual review of all Mathematics topics, both architectures and the minimum window size.
- The packaged Windows executable rendered all workspace screens and Mathematics topics, then completed real CPU forward passes for both neural models.
The .zip.sha256 asset contains the download checksum.
DriftSync v2.1.0 - Showcase Polish
DriftSync v2.1.0 - Showcase Polish
This release updates DriftSync to match the current portfolio-ready project state.
Highlights
- Rebuilt the README into a cleaner showcase page with generated screenshots and a live inference GIF.
- Added the project media under
docs/assets/, including overview, pipeline, results, and realtime demo visuals. - Added broader pytest coverage for task logic, preprocessing, sequence extraction, metrics, realtime inference, checkpoint resolution, launcher behavior, and smoke checks.
- Added GitHub Actions CI for tests and smoke verification.
- Fixed stale feature defaults for the v2 15-feature pipeline.
- Cleaned realtime inference logs so saved JSON contains prediction events only.
- Added a checkpoint resolver with clearer missing-model errors.
- Made launcher dependency installation opt-in through
DRIFTSYNC_AUTO_INSTALL=1. - Added supporting docs including model card, realtime log schema, demo checklist, and generated-artifact notes.
Why This Release Matters
v2.1.0 is the first DriftSync release aimed at being easy for reviewers to understand quickly: run instructions, screenshots, GIFs, tests, CI, limitations, and supporting documentation are now all aligned with the current codebase.
Verification
python -m pytest -qpython -m compileall .- repository smoke checks from CI workflow
DriftSync 2.0 Release
DriftSync v2.0 — Release Notes
What's New
Interactive GUI Application
The project now ships as a full standalone Pygame application (launch.py) with a polished dark-themed interface.
- No manual setup required
- Dependencies are automatically installed on first run
Named & Dated Run Folders
Each experiment run is now saved to its own isolated folder under:
driftsync/results/
Runs are named using the user-provided label + timestamp, for example:
MyTest_2026-03-07_14-30/
Improvements:
- Runs no longer overwrite each other
- Checkpoints are scoped per run
In-App Results Viewer ("Previous Runs")
You can now browse all saved experiment runs directly inside the application.
Features include:
- Each run appears as a named tab in the Previous Runs screen
- Clicking a tab loads that run's:
- plots
- model comparison table
No need to manually open the results directory.
Configurable Experiment Parameters
Before starting a demo run you can configure:
Run name
- Free-text label used to name the results folder
Trials per session
- Adjustable: 50 – 500
- Step: 25
- Default: 150
Number of sessions
- Adjustable: 5 – 50
- Step: 5
- Default: 15
7-Page Educational Guide
A built-in "What Is This?" section explains the entire system, including:
- Cognitive drift
- Task mechanics
- All 11 AI input features
- LSTM and Transformer architecture
- Uncertainty estimation
- How to interpret output plots
Live Training Metrics
The Demo screen now displays real-time training performance.
Displayed metrics include:
- Loss
- AUC
- F1 Score
Both the LSTM and Transformer models show live sparkline charts for each epoch during training.
Bug Fixes
Play Task crash
- Closing the task simulator window no longer terminates the entire application
sys.exit()replaced with a proper return signal
Checkpoint naming mismatch
compare.pyandinference_engine.pynow correctly locate checkpoints saved by the trainer
Live metrics epoch number
- Epoch count is now displayed correctly for both LSTM and Transformer metric callbacks
PyTorch scheduler compatibility
- Removed deprecated
verboseargument fromReduceLROnPlateau - Prevents crashes on newer PyTorch versions
Windows encoding error
- Logger now uses UTF-8 encoding to prevent
UnicodeEncodeErroroncp1252terminals
Pygame import crash
LiveDriftSimulatoris now lazily imported- The inference engine can run without Pygame installed
Build & Distribution
PyInstaller support
Added:
DriftSync.spec
Includes proper collect_all handling for:
- PyTorch
- NumPy
- scikit-learn
- Matplotlib
- Pygame
Executable runtime detection
launch.py now detects when running inside a PyInstaller bundle (sys.frozen) and:
- Skips the pip installer
- Prevents the executable from attempting to run itself as Python
Working directory fix
The application now resolves the working directory to the executable location rather than the temporary extraction folder used by PyInstaller.
DriftSync Initial Release
First public release of DriftSync!
What's included
- LSTM & Transformer-based error prediction models
- Cognitive task simulator (interactive + synthetic)
- Full training, evaluation & calibration pipeline
- Monte Carlo Dropout uncertainty estimation
- Real-time inference with live warning triggers
Requirements
Python 3.11+ — see README for setup instructions.