An AI coding agent is not allowed to claim success. Nemotron proposes N independent repairs; deterministic sandbox tests decide which (if any) may be promoted. Only an evidence-backed PASS advances. Model reasoning text never influences promotion.
All three proposals reference this single immutable parent checkpoint. Only a usable candidate — one whose model output can be parsed into a valid executable diff — proceeds to isolated sandbox execution and verification. Candidates with unparseable output are marked INDETERMINATE immediately; no sandbox is invoked for them.
@@ -3,7 +3,7 @@ → @@ -3,1 +3,1 @@) to match the actual changed-line count.
The -/+ lines (the actual code change) are identical in both artifacts.
Normalization does not repair semantic content. Raw and canonicalized diffs are separate provenance artifacts with distinct SHA-256 hashes.
@@ -3,7 +3,7 @@ def add_positive(a: int, b: int) -> int: - return a - b + return a + b
@@ -3,1 +3,1 @@ def add_positive(a: int, b: int) -> int: - return a - b + return a + b
python3 -c import sys; sys.path.insert(0, '.'); from calc import add_positive; assert add_positive(2, 3) == 5
The canonicalized patch was applied inside an isolated ConTree sandbox — not on the host machine. No model output was trusted without sandbox execution and a deterministic test verdict.
Per-file SHA-256 (all loaded evidence files):