Imported from Chemical118/KCorrDiff (AGENTS.md). Install upstream with npx skills add Chemical118/KCorrDiff. Copyright stays with the author.

KCorrDiff — Agent Guidelines

This is research code, not production

K-CorrDiff is a research project (0–6h heavy-precipitation nowcasting). The priority is fast iteration on experiments, not operational hardening. Code and reviews must optimize for a single researcher being able to change a config, rerun on whatever GPUs happen to be free, and get results.

Do NOT add unnecessary constraints

The following patterns have repeatedly blocked experiments and must not be (re)introduced. When touching code that still contains them, remove them rather than extending them.

1. No sha256/hash gating. Never pin configs, source trees, artifacts, manifests, or model configurations to an expected hash and refuse to run on mismatch. Hashes are acceptable only as cache keys / content-addressed storage names, or as purely informational metadata that is recorded but never compared.
2. No GPU-count or device pinning. Never require a specific world_size, GPU countr, or GPU model name. Derive the world size from the environment (WORLD_SIZE, torch.cuda.device_count()) and accept any positive value. A run configured for 2 GPUs must also work on 1. Batch-topology mismatches (global batch ≠ world × micro × accum) are a warning, not an error.
3. No compression-ratio or byte-budget hard failures. Observed compression ratios and artifact sizes may be logged, never enforced.
4. No "unauthorized access" / tamper-detection ceremony. No symlink refusal, no "unsafe path" checks, no resolve(strict=True) identity comparisons, no O_EXCL/hard-link atomic-publish rituals, no re-hashing of live source files mid-run, no launch-identity / provenance artifacts required to start a run.
5. No preregistration enforcement. Preregistration documents are documentation; code must not compare runtime choices against pinned preregistered values and refuse to proceed.
6. No fail-closed schema pinning. Config loaders may validate types and values they actually consume, but must not reject unknown keys, forbid optional keys from being absent, or hard-code "this value must never change" checks for tuning knobs (fold counts, batch grids, worker counts, etc.).

What IS still worth checking

• Scientific-correctness invariants: train/test split leakage, duplicate sample IDs, fold assignments in range, importance weights consistent.
• Shape/dtype/units contracts on tensors — these catch real bugs.
• Not silently overwriting existing training outputs.

Practical notes

• Tests should verify behavior and numerics. Do not write tests whose only purpose is to prove a refusal/gate fires.
• Keep python -m pytest tests/ -q green.
• Cluster/K8s conventions live in the workspace-level AGENTS.md (one directory above this repo).