# Ingot (SN97) challenger build — progress proof

Automated backup of build/train/eval artifacts from the H100 pod (RunPod, 1x H100 NVL).
Pushed by `hf_push.py` — checks hourly and uploads any new/changed artifact.

- `king-ix/` — incumbent king GGUF (Ingot-IX-27B-Q2) + provenance files (baseline)
- `chal-identity/`, `chal-v1/`, `chal-n2/`, `chal-norms/` — challenger GGUF iterations
  (scale-refit / QAT variants; same 27B Q2 architecture as the king)
- `states/` — QAT state checkpoints (`t0..tN.safetensors`); the quota-blocked ones
  (t2–t9, t15) are stored as `states/missing_t2-t9_t15.tar.zst` (restore: `tar --zstd -xf`)
- `chal-n2/chal-norms.gguf.zst` — zstd-19 archive of the chal-n2 GGUF, uploaded compressed
  to fit the free-plan private-storage cap (raw sha256 in MANIFEST; restore: `zstd -d`)
- `artifacts/` — `HANDOFF.md`, build/train/eval scripts, run logs, sample draws,
  measurement results (`measure*.jsonl`, `pick*.jsonl`, `drawn.jsonl`)
- `MANIFEST.md` — size + sha256 + upload time of every artifact (regenerated each push),
  including a "Pending" section for files blocked by the free-plan private-storage cap
  (their bytes are preserved on the build pod and in a local rescue cache)
- `PROGRESS.md` — push history

The GGUFs are ~6.8 GB each (PQ2_0 quantized 27B linear-attention models).
Note: chal-identity's GGUF is byte-identical to king-ix's (sha256 in MANIFEST) — the
identity refit test round-trips the king exactly.