Data and scaffolding archive for: Grounded autonomous research — a fault-tolerant LLM pipeline from corpus to manuscript in frontier computational physics (ICML 2026 AI for Science Workshop)

Claim-support archive (v1) for the ICML 2026 AI for Science Workshop paper "Grounded autonomous research: a fault-tolerant LLM pipeline from corpus to manuscript in frontier computational physics" (H. Huang). It contains the canonical run's complete prompt set and house rules (including the five prompts reproduced verbatim in the paper's Appendix C), the curated knowledge base (INDEX.md, 470 lines, post-pilot state; PSEUDOPOTENTIALS.md), the 12-theme corpus-categorization scripts and outputs with the full 11,083-entry arXiv-ID list, the literature-consultation extraction scripts and deduplicated event data (2,162 events, 14 channels), the figure-generation scripts and derived data behind all figures, the five pilot reproduction verdict records (T1–T4 tier scheme), the audited human-intervention ledger behind Table 2, all 56 production input decks, the QE 7.5 pw2wannier90 wfcU patch note, minimal evidence for the no-pilot-ablation claims of §4, and the companion manuscript's Supplemental Material as produced. Every file is justified against a specific claim or promise in MANIFEST.md.