gemma-4-12b fast2 — 2/6 (MinimalPi, K=1)

Companion to fast__gemma-4-12b__20260727-010519 (4/9). Same config: MinimalPi, K=1, AGENT_TIMEOUT_MULT=2.0, the set's default TASK_SET_fast2_TIMEOUTS cap map, all config/harness.env defaults.

Result: 2/6, elapsed 52m07s

task	gemma-4-12b	qwen3.6-35b-a3b	qwen3.6-27b
headless-terminal	PASS 12m20s	FAIL 3m42s	PASS 1m10s
git-multibranch	PASS 1m26s	FAIL 2m02s	PASS 1m17s
fix-ocaml-gc	FAIL 9m42s	PASS 7m43s	PASS 11m26s
financial-document-processor	FAIL 5m46s	FAIL 14m49s	PASS 2m29s
mailman	FAIL cut@10m	FAIL 4m47s	PASS 7m38s
reshard-c4-data	FAIL cut@5m	FAIL cut@5m	FAIL cut@5m
total	2/6	1/6	5/6

Baselines: fast2__qwen3.6-35b-a3b__20260710-192053, fast2__qwen3.6-27b__20260709-085257.

Combined with the fast set: gemma 6/15, qwen3.6-35b-a3b 6/15, qwen3.6-27b 12/15.

Headline

gemma-4-12b ties the 35b MoE on these 15 tasks and is well behind the 27b dense (12/15). For a 12B model at Q6_K that is a respectable showing against a 35B MoE, but the 27b dense remains the strongest model on this box by a wide margin — it is not close, and the gap is not K=1 noise at 12 vs 6.

Note gemma beats the 35b on fast2 by winning two tasks the MoE lost (headless-terminal, git-multibranch) while losing the one the MoE won (fix-ocaml-gc). Under K=1 that is a 2-vs-1 exchange on three coin flips — treat the fast2 ordering as unresolved, not as gemma > 35b.

The dominant failure signature: prompt-token churn

This is the clearest gemma-specific pattern in the run, and it is severe.

task	gemma input tok	qwen 27b input tok	ratio
mailman	5,785,378	1,260,382	4.6×
fix-ocaml-gc	4,702,828	1,672,526	2.8×
financial-document-processor	2,047,964	84,980	24×
reshard-c4-data	607,194	278,960	2.2×

mailman took 147 tool-use turns before its 10-minute cap cut it. Each turn re-prefills a growing context, so the trial spent its budget on prompt processing rather than progress. fix-ocaml-gc reached a 151 k peak context — over the declared 131 k window — and took 2 loop-guard blocks (lg:2), i.e. it was repeating identical tool calls until the guard intervened.

The model is fast per token (117–149 tok/s measured at first light, ~2× the MoE) but that speed is being spent on turn count, not on solving. Whatever wall-clock advantage the 12B has is fully consumed by churn on exactly the tasks where it would matter.

headless-terminal is the standout pass

12m20s, 92 k output tokens, 11 verify-like commands — by far the most self-checking of any trial in either gemma run, and the only trial where the preamble's RULE 2 behavior is clearly visible. Also a task the 35b MoE failed. One length stop mid-run plus a runaway recovery (rec:1), and it still landed the pass.

Caps

Three trials were cut (reshard-c4-data 5m, mailman 10m). The cap map is calibrated for qwen, and qwen hit reshard-c4-data's cap too — so those are comparable cuts, not a gemma handicap. mailman at 147 turns was not going to converge in any plausible budget.