| # Evidence provenance | |
| The full-score page structure, theorem checks, and real-model artifacts were | |
| adapted from the public Space | |
| [`ProCreations/repro-the-sign-estimator-preference-modeling-for-llm-alignment-under-heterogeneity`](https://huggingface.co/spaces/ProCreations/repro-the-sign-estimator-preference-modeling-for-llm-alignment-under-heterogeneity) | |
| at revision `2f159d05de444e8ca68a30cf4f5e486dfa79c5f3`. SabaPivot had already run an | |
| independent DistilGPT-2/SHP reproduction for the participation and topology | |
| claims. On 2026-08-03, SabaPivot additionally re-ran seven CPU-only public | |
| programs covering FedDPO/DecDPO scaling and exact lower-bound certificates. | |
| The fresh stdout, source hashes, exit codes, runtimes, Python version, and | |
| platform are in | |
| [`fresh_rerun_20260803/RUN_MANIFEST.json`](fresh_rerun_20260803/RUN_MANIFEST.json). | |
| Saved peer JSON was not counted as a fresh execution. | |