EXPERIMENT 04PATCH TOURNAMENT

Don’t trust the first patch.
Prove the best one.

Nemotron proposes competing fixes. Nebius forks one exact runtime into parallel sandboxes. Tests—not confidence—decide what ships.

SEARCH WIDTH 3× parallel hypotheses

TARGET REPOSITORY

Python

Issue brief

01 Preserve validation 02 Prove edge cases 03 Minimize risk

LIVE EXECUTION

One state. Three futures.

Ready to fork

01PlanNemotron

02ForkConTree

03ExecuteSandboxes

04JudgeEvidence

{ }

IMMUTABLE BASEcheckpoint: ledger-7f3a

verified

AWAITING

Nemotron will generate an independent repair thesis.

BWAITING

Each branch starts from the exact same checkpoint.

CWAITING

Identical tests create comparable evidence.

Execution receiptNEBIUS SANDBOX

Regression suite9 / 9 passed

Runtime isolationPending

Model pathPENDING

Parent stateledger-7f3a

Winning checkpointpending

✓ Reproducible evidence packet

Winning patch

Run the experiment to reveal the verified patch.