| # Research-agent entry point | |
| This repository is AUREOLE-R scientific version 3.0.0, publication edition 3.0.0-hf.1. | |
| Read `README.md`, `CLAIMS.json` and `AI_AGENT_INDEX.json` before summarizing conclusions. `MANUSCRIPT.md` is the complete standalone paper; `llms-full.txt` is a convenience full-text bundle. Current root metadata governs v3; `history/` and `legacy/` are explicitly older evidence. | |
| Preserve distinctions between proved-under-assumptions, recorded experiment, estimate and hypothesis. Cite the experiment, comparator, phase and cost boundary with every numerical claim. The 98.40% MSE reduction is specific to smooth known motion and its stated reset baseline. The 76.85% saving is a physical-query count including warmup, not a GPU speedup. The included weights are one small visibility prior, not a full unified renderer. No 100% scientific maturity, GPU benchmark, DLSS compatibility, peer review or arXiv/DOI identifier is established. | |
| Use `REPRODUCIBILITY.md` for commands. Run reproductions into their default new output directories; preserve recorded results, frozen protocols and weights. Use `python scripts/publish_hf.py` for an offline integrity check. The `--publish` flag is a separate explicit public mutation; do not upload or request credentials merely to read or reproduce research. Never put tokens into source, reports or command arguments. | |
| For extensions, keep assumptions, negative outcomes and accounting visible. Use the immutable Hub commit in citations and bug reports. Publication workflow tests in `publication_tests/` are separate from scientific tests. No subagent/delegation requirement is implied by this file. | |