Inspiration
A plausible explanation of a numerical failure is not a regression test. ReproCase explores the missing step: turning a bug symptom into a small, locally reproducible piece of evidence.
What it does
ReproCase is an experimental engineering tool that asks a constrained NVIDIA Nemotron model to propose diagnostic actions or declarative properties. Validated local code—not arbitrary generated code—runs the checks, searches for counterexamples, and exports regression tests that can be replayed against before/after implementations.
The desktop interface supports English and Russian, dark and light themes, explicit live-request consent, and offline replay. The public demo works without an API key.
How it is built
The model is nvidia/nemotron-3-super-120b-a12b, accessed through Nebius Token Factory. A response must pass both schema validation and a manually prepared semantic contract before it can influence execution. Boundary search and Hypothesis exercise the permitted domain; local replay checks whether the exported witness really fails before the fix and passes after it.
The historical pilot covers CPython correlation at different scales and a more-itertools reverse empty range. Educational faults are kept separate from these historical cases. The vocabulary of properties and acceptance rules is prepared manually: this is constrained assistance, not autonomous discovery of arbitrary software defects.
Evidence and challenges
One successful live property-preparation request used 1,551 input and 195 output tokens and produced two confirmed exports. The model-property route and manual control each required 767 calculation executions. This does not establish a model advantage.
A later six-position series made two requests, left four positions unrun, and recorded 0/6 confirmed exports because a local exporter failed. The failure remains in the accounting; subsequent offline repair and replay are not presented as a new successful live evaluation. This exposed an important engineering lesson: a valid model response is not enough—the exported artifact must independently reproduce the result.
The public package includes 381 engineering tests. These test the implementation, not model quality. Finding no counterexample within a search budget does not prove correctness or the absence of a bug. No measured advantage over the manual baseline or labor saving has been demonstrated.
Try it
Source, MIT license, third-party notices, setup instructions, and the offline judge are available at: https://github.com/WERFESTT/ReproCase
Setup and pinned English README: https://github.com/WERFESTT/ReproCase/blob/c8c7edc0a92a81183f1f8670871892cf9a4120f7/README_EN.md The tested environment is Linux x86_64 with CPython 3.14.4; the GUI uses Tk. From the prepared project environment, run:
python3 -I -B judge.py verify --language en
python3 -I -B judge.py demo
No key is needed for this offline route. Optional real inference requires your own separately authorized Nebius key and may incur charges. The live property route uses masked, memory-only key input, a one-request/1,024-output-token limit, a 30-second timeout, and no automatic retry. Unknown usage stays unknown rather than being counted as zero.
The video shows a saved real response and local replay, not a fresh live API call. It is silent.
What is next
Expand the historical evaluation set, improve exporter robustness, and compare constrained model proposals with stronger manual and fixed-search baselines before making claims about effectiveness or saved effort.
Log in or sign up for Devpost to join the conversation.