# Public factual source/modification notice

Source: MLCommons MLPerf® Inference: Edge v3.1 Closed ResNet50 SingleStream, NVIDIA submission at revision f7b504d40dfb65180fe59438d4b4d6f9f8598647. The exact two source-event folders and official entries 3.1-0117 / 3.1-0115 are linked in the native-only projection. Retrieval: 18 October 2026.

The native-only candidate is an approved minimal factual selection at source/reuse scope: native completed-image counts and source-run mean/p90 latency, with separate reported accuracy and TEST01 qualification. It does not relabel these secondary source fields as the rounded official primary metric. This site's secondary-metric presentation is unverified. Result not verified by MLCommons Association. The underlying published submission's status remains distinct from this new editorial selection.

The complete original Apache-2.0 license and NVIDIA attributions are delivered unchanged in public-assets/. Preserve them and the visible source/ownership attribution. They do not establish blanket permission for embedded third-party SPEC/PTDaemon expression or software, source web/CSV assets, models, images, binaries, code or predictions. None of those raw objects is included.

Modifications: Codex for Nathan Fennel, 18 October 2026. Newly authored minimal factual field selection and explanatory notes; no source remeasurement or new count credit. The public-native-only projection contains no reconstructed numerical energy. The full numerical energy candidate remains held for the exact public measured-power/derived-reporting basis and is excluded from the approved public-content allowlist. Private transfer decisions do not supply that public permission.

Mandatory common annotation for view and standalone exports: the separate energy reconstruction has two warning-bearing beginning-bracket readings per integral, including one timestamp outside the work window. No aggregate uncertainty interval is inferred. Physical wiring/calibration and total timing uncertainty remain unresolved, as do loaded artifacts and full output replay. No causal rank, per-GPU/token/correct-image/facility/lifecycle allocation is supported. Accuracy and TEST01 concern separate source runs, not replay of all original performance-window predictions.

Source/reuse approval is confined to the exact native-only candidate, owned notes and delivered notices/assets. No missing-energy result card should be presented. Adapter, visible attribution, exports, mobile, final built UI and release review remain separate; this source receipt grants no production authority.