Repair Corridor

An AI repair assistant that can say 'not yet': NVIDIA Nemotron on Nebius proposes a fix; independent tests check the exact edit, and a clear receipt explains the result.

Comment

Why we built it

Maintainers need a way to distinguish a patch that passes visible examples from one that survives independent checks. Repair Corridor makes that decision inspectable on a newly authored, synthetic interval-merging task.

What it does

The baseline incorrectly shortens a 09:00–17:00 reservation when a 10:00–11:00 booking sits inside it. The 46-second video follows that illustrative user story. A deliberately authored overfit control passes all three visible examples but gets HOLD under independent checks. One actual NVIDIA Nemotron repair generated through Nebius Token Factory gets ACCEPT. Reviewers can inspect the patch, counterexample, check reports, and decision receipt; the bundled replay needs no API key.

How we built it

A Python harness sent one bounded request to nvidia/Nemotron-3_5-Lightning through Nebius Token Factory. It captured the candidate, Git patch, source and checker digests, response summary, and token usage. A separate verifier used an independent occupied-cell oracle. The public repository includes the synthetic fixture and credential-free replay.

Results

The baseline failed the nested case and 264 of 2,049 regression inputs. The recorded model candidate passed all three visible cases and all 2,049 independent cases; patch replay reproduced it byte for byte. The authored overfit control passed the visible cases but only 126 regression inputs. After the live run, 321 longer-list cases were added, and the same captured candidate passed them. These are observations from one synthetic repair cycle, not a general model benchmark.

Challenges and learning

The model response lacked a terminating newline, so the Git diff had to preserve that detail for reliable replay. Review also found a verifier exit-code bug: HOLD now exits 1, ACCEPT 0, and UNKNOWN 2. The local suite has 21 passing tests. Visible-green is insufficient without a separate oracle and reproducible evidence.

Try it

Source and replay instructions: https://github.com/instagrim-dev/nemotron-repair-trial Pinned test build: https://github.com/instagrim-dev/nemotron-repair-trial/archive/d1ef9c13ee9d3668deee0663df3c4d0f49eaf859.zip Run: python3 -B -m demo.replay --output /tmp/nemotron-demo-first

Next

Add a second task and a matched model comparison with identical budgets and verifier rules.

Built With

Share this project:

Updates

Submission history