An earlier version of this card reported 45.3% agreement with live human play and claimed variants A and B were statistically identical. Both were wrong. vLLM does not apply LoRA to Qwen3.5 (gated-delta-net linear attention; all-linear targets those projections and vLLM skips what it cannot serve rather than raising), so the evaluation was effectively scoring the base model. Three differently-trained adapters returned byte-identical actions on all 161 decisions -- identical metrics across models was the bug signal.

Re-measured through PEFT, all 161 traces closing </think>:

So the real findings are the opposite of what was published: reasoning beats scale here (the best 4B leads a 14B DPO adapter by 18pp at a tenth the parameters), and the label source matters a great deal -- heuristic labels beat teacher labels by 16.8pp in v1 and 8.7pp in v3.

Serve this adapter with PEFT, not vLLM. Under vLLM it silently returns base model behaviour.

Caveats: agreement with live human play is not GTO optimality, and there is no bb/100 for this line yet.

A LoRA adapter on Qwen/Qwen3.5-4B trained to emit an explicit reasoning trace before its action:

<think>Required equity = 3.5/(10.5+3.5) = 25%. I hold second pair with a
backdoor flush draw; against a range that bet flop and turn I have roughly
29% -- above the price, so calling is marginally +EV. Raising folds out
nothing that calls twice. Call.</think>

call

Reasoning traces were distilled from a Bedrock teacher (Claude Sonnet 5) in the STaR style: the teacher was not told the answer, and a trace was kept only if its conclusion matched the target label. That keeps every retained trace one the teacher actually believed, rather than a rationalisation of a conclusion handed to it.

On the 161-decision DPT human-table regression set:

Use the 14B DPO adapter for play. This one exists for the ablation below.

Variant A (heuristic labels, 4278 rows) and variant B (teacher labels, 6108 rows) came out statistically indistinguishable -- aggression 64 vs 65 decisions, agreement 46.6% vs 47.8%, pot-odds violations 5 vs 5 -- despite 42% different data volume and completely different label sources. So at this scale the label source is not what moves the needle; the reasoning format is.

A second finding came out of the same run: both variants inherited a pot-odds discipline gap (5 of 40 calls below break-even) directly from their training data, where 13% of call rows were themselves below break-even (80 of 593 in A, 130 of 997 in B). The student reproduced almost exactly the data's own violation rate. Later revisions enforce the arithmetic floor mechanically rather than by prompting the teacher.

Sampling: temperature=0.1, top_p=0.95. Allow at least 512 output tokens -- the reasoning trace averages ~95 words and truncating it before </think> yields no action at all.

Prompts must match the production renderer byte for byte; see src/pokerbench/data/production_pe_renderer.py in the source repo.

Labels come from the hand-written heuristic labeler, unchanged from every prior experiment in this line, and a trace is kept only when the teacher's conclusion agrees with it (4278 of 6108 usable traces, 70.0%). This isolates one variable: does reasoning help, holding the label source fixed?

Compare with variant B, which trains on the teacher's own decisions instead.