sn120-ring-r34-3085

SN120 (Affine) submission candidate. Submission-ready: every file here is on the validator's intake allowlist, the architecture matches [submission.pinned_arch], and generation_config.json ships the decode the duel actually uses.

Recipe

Task arithmetic, theta = base + sum_i lambda_i * (donor_i - base):

term	model	lambda
base	greenfield0810/affine-ark-b05270399e4f (ringtone)	-
donor	ringtone + angryaffine ckpt/034 LoRA	0.3
donor	reigning king 0a04e3d1...	0.85

tensors_modified: 1045, tensors_untouched: 0.

Decode

The duel posts only temperature (0.8, from affine.toml [duel]) and the eval pod's vllm serve carries no --generation-config flag, so vLLM's auto default takes top_p/top_k from generation_config.json here: top_p 0.6, top_k 20. Do not change them.

Measured

chal-00285, temp 0.8 / top_p 0.6, paired against the reigning king's published per-turn rows:

	margin	z	clears
this model	+0.002358	3.31	yes (bar 0.002)

Per-dialect per-turn margin: bash +0.000339, tool_call +0.006800, boxed +0.004798.

Submitting

The HF affine1 path is retired (commits after block 8987674 are dropped). Use the live client:

curl -O https://s3.hippius.com/affine-sn120/code/scripts/submit.py
python submit.py check  --model-dir ./sn120-ring-r34-3085
python submit.py submit --wallet W --hotkey H --model-dir ./sn120-ring-r34-3085 --name r34-3085

The hotkey must be Ed25519 (python submit.py hotkey ...); btcli's default sr25519 cannot open the sealed credential mailbox. The eval slot burns at ready, not before.