Verified Coding Agent

Nebius Prize — DOMINO 5  |  Best-of-N Branch Competition  |  Judge Demo
ACCEPTED LIVE RUN REPLAY — zero-network, no credentials required DOMINO 4 Prior Replay →
Product promise

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.

1
Bounded Broken-Code Task & Original Defect
INPUT
Mission IDm-test-24-domino5-live
File under repaircalc.py
Functionadd_positive(a, b)
Defectreturn a - b  (should be a + b)
TaskFix the subtraction defect so targeted test passes
Rounds executed1
Candidates requested3
Immutable parent checkpointcontree:d4160f64-5700-444f-8a37-7c1cfc863739

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.

2
Best-of-N Proposal Competition — 3 Independent Proposals
3 CANDIDATES
Candidate 0 INDETERMINATE
Verdict INDETERMINATE
Reason UNPARSEABLE_GENERATION
Raw diff extracted yes — not executable
Request ID chatcmpl-9e280650-017b-4638-aad9-cf793ba2f569
Why it did not advance: A raw diff was extracted from the model output for diagnostic purposes, but it could not be parsed into a valid executable diff. No sandbox was invoked. Marked INDETERMINATE / UNPARSEABLE_GENERATION.
★ Candidate 1 WINNER
Verdict PASS
Exit code 0
Changed files 1
Lines ± +1 / -1
Executable diff yes (canonicalized)
normalization_applied true
Request ID chatcmpl-f24208ea-42c5-4a69-898f-056f29e7c32f
Why it advanced: Model output parsed into a valid executable diff. Canonicalized patch applied in isolated sandbox. Targeted test passed (exit 0). Exactly one candidate achieved PASS — SINGLE_PASS rule promoted immediately. Full-suite verification not required.
Candidate 2 INDETERMINATE
Verdict INDETERMINATE
Reason UNPARSEABLE_GENERATION
Diff extracted no
Request ID chatcmpl-f976079e-7b31-4a0a-811d-563919882216
Why it did not advance: No diff could be extracted from the model output. No sandbox was invoked. Marked INDETERMINATE / UNPARSEABLE_GENERATION.
Isolation: Each proposal was generated by a separate independent call to Nemotron Nano with a distinct request ID. All three proposals reference the same immutable parent checkpoint — no proposal can affect another's state. Each call is sequential. Only candidate 1 produced a usable executable diff and proceeded to isolated ConTree sandbox execution. Candidates 0 and 2 were rejected at the diff-validation stage before any sandbox invocation. The promotion rule is deterministic and evidence-based; model reasoning text never influences the outcome.
3
Winner Diff Provenance — Candidate 1 (★ WINNER)
PASS
Candidate IDr-0-a8107b47-c1
Modelnvidia/NVIDIA-Nemotron-3-Nano-30B-A3B
normalization_appliedTrue
normalization_applied = true — Normalization adjusts unified-diff hunk-count metadata only (e.g., @@ -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.

Raw parsed diff (as extracted from model output)

raw_parsed_diff_sha2568265d38ba95841e08321acb0f859a77606985faef877eada81c632fe796743de
--- a/calc.py
+++ b/calc.py
@@ -3,7 +3,7 @@ def add_positive(a: int, b: int) -> int:
-    return a - b
+    return a + b

Canonicalized execution diff (hunk counts normalized — applied to sandbox)

canonicalized_diff_sha2560d23c0cc0546c8332bc8378f6ef4bc690cb86d72922ecb028acc0e8b06b09f2c
candidate.patch_diff_sha2560d23c0cc0546c8332bc8378f6ef4bc690cb86d72922ecb028acc0e8b06b09f2c
--- a/calc.py
+++ b/calc.py
@@ -3,1 +3,1 @@ def add_positive(a: int, b: int) -> int:
-    return a - b
+    return a + b
4
ConTree / Nebius Sandboxes — Isolated Execution & Targeted Verification
PASS
Execution backendConTree / Nebius Sandboxes
Winner candidate IDr-0-a8107b47-c1
Resulting checkpointcontree:d16f3d2a-3e29-4872-b88d-9c562a772071
Targeted verdictPASS
Exit code0
Test node IDscalc::add_positive(2,3)==5
Verification command
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.

5
Full-Suite Verification
NOT_REQUIRED
NOT_REQUIRED ≠ PASS and ≠ not invoked due to error. Full-suite verification is triggered only when multiple candidates all achieve a targeted PASS and a tie-break is needed. In this run, exactly one candidate (candidate 1) achieved PASS; the other two were INDETERMINATE. SINGLE_PASS applied — full-suite tie-break was architecturally not needed.
Full-suite statusNOT_REQUIRED
Promotion ruleSINGLE_PASS
ReasonSINGLE_PASS — exactly one candidate achieved a targeted PASS (candidate 1); the frozen promotion cascade does not invoke full-suite tie-break verification when there is only one passing candidate. Candidates 0 and 2 were INDETERMINATE/UNPARSEABLE_GENERATION and did not enter tie-break.
6
Final Decision
PROMOTED
PROMOTED
Evidence-backed PASS — Candidate 1 passed targeted verification in isolated sandbox. Evidence chain validated. SINGLE_PASS rule applied.
Promotion ruleSINGLE_PASS
Winnerr-0-a8107b47-c1
Promoted checkpointcontree:d16f3d2a-3e29-4872-b88d-9c562a772071
Finalized at2026-09-08T01:05:42.385918Z

Evidence provenance — all SHA-256 hashes
mission_outcome SHA-25684c17633d66740c45df3db46c1501a41fead558f52bf1a429b7630b97cf16899
promotion SHA-256b5525d1a1d79c513d886a2f383bd1284acfcfce17d99a8345e45422351f617ab
winner raw_diff SHA-2568265d38ba95841e08321acb0f859a77606985faef877eada81c632fe796743de
winner canonicalized_diff SHA-2560d23c0cc0546c8332bc8378f6ef4bc690cb86d72922ecb028acc0e8b06b09f2c
winner patch_diff SHA-2560d23c0cc0546c8332bc8378f6ef4bc690cb86d72922ecb028acc0e8b06b09f2c

Per-file SHA-256 (all loaded evidence files):

manifest.json 2578daaf7312abe0e452c13743efd8402a15aa183826d85d842b49ffb5f988da
mission_outcome.json 84c17633d66740c45df3db46c1501a41fead558f52bf1a429b7630b97cf16899
promotion.json b5525d1a1d79c513d886a2f383bd1284acfcfce17d99a8345e45422351f617ab
candidates/candidate_0.json b678bc12ecd1dd738fede54382f55ec043ce1a8b13e3222251fb851b74110033
candidates/candidate_1.json a3d97d2eac276c8dacd3497b3bac682c32a82fc628a645d79c0aeb99db29da6f
candidates/candidate_2.json 3226cba9e7db00ebf76abb410a00d6e689011c0da6bfa670a4bd53ed3eedb01f
diagnostics/candidate_0/generation_metadata.json 8194150a4bdaadfae4c7c36e020f2c51e1ba957166b9c17de443e541c1ebd782
diagnostics/candidate_0/model_output.txt 3fae10a23ecbb0624e9e5e39db85812b4c26e71ad0b1a5e87eb2d8e914049761
diagnostics/candidate_0/raw_parsed_diff.patch dee1c39c2e65655f6aedaf8c697ffdd58d016bc504324af5a0a40d157e10db86
diagnostics/candidate_1/generation_metadata.json e97abcc10a5e6ac655fe9654e031aa8aced1aecb1ca66ce752395afff75d48be
diagnostics/candidate_1/model_output.txt d2c9381dfd4d5c48fd5ae31c0cd466c36b7b2c47c8d95d84f838c68e59afa9e8
diagnostics/candidate_1/raw_parsed_diff.patch 8265d38ba95841e08321acb0f859a77606985faef877eada81c632fe796743de
diagnostics/candidate_1/canonicalized_diff.patch 0d23c0cc0546c8332bc8378f6ef4bc690cb86d72922ecb028acc0e8b06b09f2c
diagnostics/candidate_2/generation_metadata.json 5faa691a0587c7f1bb3cd1c40a65fd9b29f075f353d0e919b79b6ec7c3a849f2
diagnostics/candidate_2/model_output.txt 544d0a36c151f7233ca8a7743f92fa57808e1a19d67168f8a3dad3975ee26264
verifications/candidate_0_targeted.json c0e1ce7a435c045a2cc8d9e100411a87ef1917c67aee9a9f6ded11aa060e4610
verifications/candidate_1_targeted.json ec998e5d64b6bf5bc79b101df543ec1ba0c72b6d6a6c60829b74becb5ae13534
verifications/candidate_2_targeted.json 3b1131bb9feb4ede951caf6f25e73de6af46ce0caf51c5f5800b052eb1c44dd9