RLTT/Affine-sn120-chal-00495

Affine SN120 candidate — calibration fine-tune from the reign-11 king

A LoRA calibration fine-tune of the Affine SN120 reign-11 king (0ce59769300c234203dcaa786e50f3a23b564bb65e83db0a711a7a43c7951952), trained against the subnet's min(R, G) distillation score.

Architecture is the text-only extraction of the genesis family (Qwen3_5MoeForCausalLM, 34.7B total / ~3B active, 40 layers, 256 experts), which matches [[submission.pinned_arch_alt]] in the chain contract.

What it optimises

The live score is turn = min(R, G) where

• R = tau*log(mean_i exp(a_i/tau)) - mean_i a_i, a_i = lpC(y_i|z_A) - lpC(y_i|empty)
• G = w - |m - mu|, m = lpC(z_A|x), mu/w from the teacher's own reference thoughts

R >= 0 always (Jensen), so a turn can only go negative through G or through a forfeit (-0.1, for no parseable action or a reply that never closes </think>). Training therefore targets grounding calibration and forfeit elimination rather than raw capability.

Training data

Built entirely offline from published + locally generated teacher rollouts (Qwen/Qwen3.8-27B):

• 8 samples per turn, keeping the one whose lpC(z|x) is closest to the mean of the 8 — the teacher's most typical thought, not its most likely one (the latter lands above the band and scores negative)
• every target re-verified through the validator's own split_rollout, so the scored z is byte-identical to what the teacher produced
• hard per-dialect token caps well under the 1792-token contract ceiling
• all four live dialects: bash, tool_call, boxed, text

Measured vs the reign-11 king

Replay of chal-00427's published 1300-turn slice with the validator's own scoring code (affine/score.py, evalsrv/terms.py):

metric	king	this line of work
protocol probe	0.90 (its exact pass threshold)	1.00
think-close rate	0.976	0.983
forfeit rate	2.85%	2.22%
mean R leg	0.00788	0.00838

Harness fidelity: the same simulator measures the king at -0.006077 against its published -0.006322 on the same slice.

Caveats

Advisory benches (SWE-rebench) were not run on this checkpoint. The subnet's own research records a negative rank correlation between duel margin and SWE-rebench on the live board, so a min(R, G) gain should not be read as a coding-capability gain.