RLTT/Affine-sn120-cand2

Affine SN120 — cand2

A calibration fine-tune of the Affine SN120 reign-11 king (0ce59769300c234203dcaa786e50f3a23b564bb65e83db0a711a7a43c7951952), merged and servable as-is. Architecture is unchanged from the base (Qwen3_5MoeForCausalLM, 40 layers, hidden 2048).

This is storage and sharing only. Since the 2026-09-03 cutover, Hugging Face is not the Affine submission path — submission goes through the private R2 flow.

What it is for

Affine scores a turn as min(R, G):

• R (centered tempered Reason) — does the thought commit to one specific teacher action mode, R = tau*log(mean_i exp(a_i/tau)) - mean_i a_i with a_i = lpC(y_i|z) - lpC(y_i|empty). R >= 0 by Jensen.
• G (banded Grounding) — does the thought's teacher likelihood sit inside the band the task induces, G = w - |m - mu|, w = max(band_c*sd, 0.002).
• A turn with no parseable action, or no closed </think>, scores a flat -0.1 forfeit.

The training targets are teacher rollouts generated locally: 8 samples per turn, each verified to parse in the turn's dialect, close </think>, and fit the token caps, with the most band-typical candidate selected.

Measured results

Offline replay of published duel slices using the validator's own scoring code (affine/score.py), against the reigning king:

slice	rules	margin	SE	z	cand forfeit	king forfeit
chal-00427	wvk 16 (band_c=2)	-0.000205	0.000717	-0.29	2.22%	2.85%
chal-00427	wvk 17 (band_c=4, ref_max_tokens=4096)	+0.000479	0.000660	+0.73	2.65%	3.27%

Under wvk 17 the model scores +0.000329 absolute against the king's -0.000151, and every action dialect is positive (bash +0.000052, tool_call +0.000879, boxed +0.000691, text +0.005636). It does not clear the crown floor of delta = 0.002.

Protocol probe: 10/10, think_close_rate = 1.0.

Honest framing of where the gain comes from: the advantage is the forfeit rate, not the legs. Mean R is 0.009585 against the king's 0.010014 and mean G is 0.071863 against 0.072871 — this model has never led on either leg. Under wvk 17's wider band, G binds only ~18% of turns and R decides the rest, so the remaining headroom is in R and in forfeit elimination.

Training

base	reign-11 king (0ce59769...951952)
method	LoRA, r=32, alpha=64, dropout 0.0, merged
context	4096 tokens
lr	1e-4
examples	20,000
data	locally generated teacher rollouts, 8 samples/turn, most-typical pick

Use

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("RLTT/Affine-sn120-cand2")
model = AutoModelForCausalLM.from_pretrained(
    "RLTT/Affine-sn120-cand2", dtype="bfloat16", device_map="auto")

Served the way the eval pod serves it:

vllm serve RLTT/Affine-sn120-cand2 \
  --max-model-len 131072 --tensor-parallel-size 2 \
  --gpu-memory-utilization 0.8 --max-num-batched-tokens 8192

The chat contract matters: the model is driven via /v1/completions through its own chat template, and a reply must close </think> and carry visible output after it, or the turn forfeits.