DeciShift Community
DeciShift is early-stage open source, so the most valuable community signal is independent use: people installing it, running the examples, finding edge cases, and reporting where the abstraction works or fails.
Start here
- GitHub Discussions
- Independent reproducibility testing — issue #13
- Windows and macOS smoke testing — issue #14
- MLflow integration example — issue #15
- ZenML integration example — issue #16
Independent tester checklist
A useful reproduction report includes:
- operating system;
- Python version;
- installation method;
- example or dataset used;
- runtime;
- number and percentage of changed final actions;
- whether attribution and Decision Contract results matched the documented output;
- any installation friction, unclear terminology, unexpected behavior, or discrepancy.
Negative results are welcome. A reproducibility failure or confusing API is useful evidence, not noise.
Discussion topics we especially want
- model metrics improve but operational decisions regress;
- low global shift with high cohort shift;
- threshold-only changes;
- feature transformation changes;
- policy/rule interactions;
- multiple models feeding one policy;
- categorical actions with no natural ordering;
- topology changes between versions;
- very large tabular datasets;
- attribution that conflicts with practitioner intuition.
Integration feedback
DeciShift is designed to sit beside existing MLOps tools. Practical constraints from real workflows are especially useful:
- Should integrations consume run IDs, model versions, files, or artifacts?
- Where should a behavioral release gate sit in the pipeline?
- Which metadata should be retained for reproducibility?
- What would make an integration production-useful rather than another example notebook?
Contributing
Please read the repository contribution guide before opening a pull request:
The project values reproducible examples, clear failure cases, focused integrations, tests, documentation, and corrections to over-broad claims.