The answer starts with a test.
Local, trainable Python bug discovery.
No trained checkpoint or measured advantage yet.
Current milestone
First reproducible training run.
The method
A counterexample
you can inspect.
A specific input can expose what a confident explanation misses.
The implementation · Python
def merge_windows(windows):
merged = []
for start, end in sorted(windows):
if not merged:
merged.append([start, end])
elif start <= merged[-1][1]:
merged[-1][1] = end
else:
merged.append([start, end])
return mergedend max(merged[-1][1], end)Input · minutes[(0, 10), (2, 4), (8, 12)]
0–4 · 8–120–12The nested window shrinks the stored end from 10 to 4. Minutes 4–8 disappear from the result.
assert merge_windows(windows) == [[0, 12]]Curated example with valid, nonempty windows.
JavaScript evaluates the shown logic. No model or Python runner.
A failing test alone does not prove a defect.
From a suspicion
to something
you can inspect.
Our proposed model works on an approved repository snapshot. An independent runner checks its witness.
- 01
Read the code
The snapshot and its context are visible. Issue reports, reference patches and answer tests stay sealed.
- 02
Make a testable claim
Locate a likely fault. State the expected behaviour. Propose a minimal pytest counterexample.
- 03
Let the evidence answer
Run against buggy and fixed revisions in isolation. An independent reviewer decides whether the behaviour is a defect.
Research, in progress
Research before
release.
We want to know whether training a compact model improves useful bug discovery.
That calls for a blind comparison with the frozen base model, under the same conditions.
Read the research profile- Trained checkpoint
- Not yet
- Measured advantage
- Not established
- Next milestone
- Reproducible first training run
DataProvenance and explicit source licences.
EvaluationHeld-out repositories. Independently reviewed root causes.
CostFalse alerts, latency, memory and review effort.
Inside NeuroTestix
Six core product repositories, currently private.
Links require collaborator access. Public releases will follow review.