AISimulate review previews

• PR #417: FPM Gym preview.
• Dashboard source: ai-dynamo/aisimulate at 0c5bb9ecfc46ca96feb9a88da70fc493c75dc164 (Apache-2.0). Evaluation remains pinned to 714e0188d1921ad90485670efa674cebc011011f.
• Qualified evaluation, attempt 1, October 12, 2026 UTC. Predictor wheel: 2f0f4cd8606d580567b253d02f3d699fe36db865. HF PR #16 revision: 562d958065e7695e6811561d3c78c1f648c70d3b.
• Measurement and evaluation artifact qualification, checksums, and 3D/evaluation membership checks passed for all 29 current configurations. The 3D bundle contains 302 assets.
• Compared with the prior preview, all configuration measurement counts, predictor counts, error counts, and statuses are unchanged. Regression MAPE varies between runs; warmup-on MAPE also changed slightly for GLM GB200 DEP16 (overall 5.261884% to 5.262197%). The predictor source advanced, so this is an integration rerun, not an isolated score-equivalence experiment. See the comparison.
• Chromium checks passed for 16 contributing chart runs across seven configurations, collection selection and URL persistence, missing-evidence banners, historical evaluations, all five tabs, and both 3D panels. Desktop/mobile layouts (1440/390px) were checked in both themes with no JavaScript errors or horizontal overflow.
• Related validation: evaluator/Pages tests 367 passed, 20 native SDK tests skipped, 138 subtests passed; workflow tests 11 passed; dashboard fixture browser checks passed.
• Serving columns on Overview and Predictors show values only on parallelism/configuration rows; model rows stay blank. Multiple values use commas, e.g. Agg*, Disagg, with Agg* for unknown metadata and an explanatory footnote. Explicit not-applicable remains N/A. Layout summaries are derived only from the matching pinned Details artifacts; historical artifacts and scores are unchanged. Browser checks cover both tabs, mixed layouts, defaults, legacy names, and sorting.
• Details uses compact request charts and grouped Dataset, Workload, and Collection metadata. Overview summaries display distinct concurrency settings as plain grey text. Configuration links retain pinned Hugging Face evidence.
• Historical evaluation files retain their original data. Request charts remain unavailable where matching request exports were disabled or not recovered; no measurements are invented.
• Built with scripts/pages/build_pages_site.py from qualified Actions artifacts. This is a manually published review snapshot and does not automatically follow PR pushes. Neither PR has been merged; production Pages and HF main are unchanged.
• Original copyright, license, and attribution notices are preserved in LICENSE, THIRD_PARTY_NOTICES.md, and copied assets.

PR #512: E2E HF input migration

• Preview and comparison: HF-backed E2E Gym plus the DB baseline.
• Pipeline passed all four prediction shards and final qualification at 6215ed1a8859608c760d6866b9d237c7cf34aa03.
• Numerical parity is not achieved: the cohort differs, and DB-ID bindings affect reviewed runtime/launcher evidence. Of 1,296 matched successful replay points, 35 differ; all are MiniMax-M2.5 NVFP4 on B300 with Dynamo-vLLM.
• Provenance and validation. This isolated preview preserves the existing PR #417 deployment and does not update production Pages.