audit/source-mapping.md

Source mapping

Source: https://gitlab-master.nvidia.com/dl/ai-dynamo/AISimPlayground at 731947aabdcb5ae8658fd16744022ca94e407cd3.

Source campaign	Model	GPU	FPM occurrences	Measurement occurrences
dsv4flash-b200-static-baseline-fpm	deepseek-ai/DeepSeek-V4-Flash-0731	B200	4	70
dsv4pro-gb300-fpm	deepseek-ai/DeepSeek-V4-Pro	GB300	9	41
glm52-b200-static-baseline-fpm	nvidia/GLM-5.2-NVFP4	B200	6	10
glm52-gb200-dep16-disagg-fpm-fidelity	nvidia/GLM-5.2-NVFP4	GB200	1	21
kimi-k3-gb300-fpm	moonshotai/Kimi-K3	GB300	2	8
minimax-e9-b200-ragged-commonmode	MiniMaxAI/MiniMax-M2.7	B200	1	7
minimax-m27-h200-generational-verify-dep4-tp4-fpm	MiniMaxAI/MiniMax-M2.7	H200	9	20
minimax-m27-h200-static-baseline-fpm	MiniMaxAI/MiniMax-M2.7	H200	2	4
minimax2.7-b200-fpm	MiniMaxAI/MiniMax-M2.7	B200	4	23

Data paths map as follows:

• Active data → data/<org>--<model>/<system>/<framework>/<version>/<parallelism>/{fpm,measurements}.
• Superseded snapshots → data/history/<org>--<model>/<system>/<framework>/<version>/<parallelism>/aisim-commit-*/....
• Campaign names are retained only as source provenance in manifests and catalogs.
• Multi-configuration FPM Parquets are split while preserving row order. A strict publish-time policy corrects declared dtype identity for DSV4-Flash, DSV4-Pro, Kimi-K3, and GLM-5.2-NVFP4.
• Identity correction never changes GEMM, latency, workload kind, batch size, or token coordinates. Each affected sidecar records the source and published labels.
• Usable measurement truth is flat under measurements/; summaries and helper evidence are under measurements/provenance/.
• FPM benchmark-lane configs are stored with the matching artifact under fpm/provenance/.
• Predictions, normalized comparisons, and accuracy metrics are computed by FPM Gym and are not published here.