palarran-science-preprint

原文涵盖的主题

• /palarran:science:preprint {experiment/report paths}

Shared cycle: before work and after checks follow global memory: bounded recall → applicability → sanitized verified learn/dedup. This memory scope is explicitly authorized; other profile restrictions remain. --plan permits no network/publication; outages do not block the main task. First read README, sources, evidence, workflow and decision-memory. Identify installed checkout separately from target. Read target README/AGENTS, INDEX/accepted decisions, experiment/protocol/data/model/config/benchmark and all-attempt reports. Never automatically run missing results. Freeze question/hypothesis, baseline, protocol revision, selection policy, seeds/N/splits, units/denominators, compute/cost basis and gaps. Bind each claim to exact artifact/hash/evidence pointers; self-consistency does not prove evaluation independence. Research related work in primary sources; verify metadata/applicability, never invent citations/DOI/novelty/authors/significance/success. Preserve partial/negative/failed runs. Development probes are not heldout benchmarks. Do not include private outputs without disclosure permission. Create a separate target LaTeX project with main.tex/tables/figures/references/evidence

这里只展示节选，请以完整原文和当前要求为准。