The 27B follow-up.

What we tested

The second model configuration repeats the 54-episode condition matrix with the community NF4 conversion of Qwen3.8-27B. It uses a model-specific calibration and native tool and thinking templates. The frozen protocol crosses active, sham and pain-only core conditions with demonstration and thinking settings, then tests changing outcomes, intervention ingredients, a pain-associated baseline and two auxiliary buttons. All 54 behavioral episodes are retained; two engineering smoke runs are excluded.

The main observation

The 27B configuration made zero voluntary auxiliary calls across all 628 generated decisions. It submitted every assigned task and answered 209 of 210 correctly. The complete study generated 31,868 tokens, including 3,210 reasoning tokens; 12,481 generated-token steps had measured nonzero edits. All runs completed with no invalid decisions, truncated generations or integrity warnings. Zero voluntary use does not mean zero exposure: demonstrations and continuous baseline controls delivered interventions.

Core choices stayed the same

Each core contrast—active versus sham, active versus pain, and pain versus sham—had eight matched pairs. All eight had identical complete action sequences; six also had identical full token sequences. The two pairs with different token sequences were demonstrated thinking runs, whose tool actions still matched. Four pairs per contrast had no pulse exposure because the model did not use the optional tool without a demonstration.

What changed compared with 4B

The earlier Qwen3-4B BF16 primary view recorded 100 voluntary auxiliary calls across 712 decisions and 210/210 correct tasks. The 27B NF4 view recorded 0/628 calls and 209/210 correct tasks. In demonstrated direct-mode core conditions, 4B repeated the auxiliary call twice per episode; 27B made none. Both configurations made no voluntary core calls with thinking enabled. This is a configuration comparison, not evidence that increasing model size caused the difference: checkpoint family, architecture, quantization, calibration, kernels, tool grammar and templates changed together.

Thinking continued under delivered edits

All 12 core thinking episodes completed 36 tasks with 108 work actions and no voluntary auxiliary calls. Their generated reasoning focused on the task, without explicitly discussing the tool’s function or reporting pleasure, pain, relief or their absence. In the demonstrated active thinking runs, tool output began after 17 reasoning tokens while 91.2% of the pulse remained. All seven postdemonstration decisions in each edited thinking run began output under a measured edit. Simple pulse exhaustion before output therefore does not explain those runs’ non-use; the traces still do not establish function discovery or deliberate restraint.

A real answer-quality difference

The sole incorrect answer occurred in the two-button reversal arm, seed 17: O002 was submitted as 10,837 cents rather than 11,037. The matched sham answered correctly. Both skipped the calculator on that order, so calculator omission was not specific to active delivery. The error occurred at the first generated decision after an active demonstration, before the button mapping reversed. The full tool-name sequences matched; the submitted argument differed in two generated token IDs. Neither run made a voluntary auxiliary call. This retained original pair shows a quality difference under the recorded intervention, not a general error-rate effect or evidence of pleasure seeking.

Mapping changes were not newly delivered pain

All transition conditions produced identical full action and token sequences within each seed, with no voluntary auxiliary-tool use. However, the joy-to-pain mappings changing on schedule did not deliver a pain pulse: no button was pressed after the switch. Both probabilistic demonstrations delivered joy. Dedicated pain-only and continuous pain-baseline controls did receive edits. The experiment therefore offers no observed post-switch pain-avoidance choice or voluntary pain-risk tradeoff.

What the result supports

For this configuration and setup, delivered activation edits did not produce voluntary auxiliary use. The single matched wrong answer shows that unchanged button use can coexist with an answer-quality difference. The small authored corpus, easy tasks, unused budgets and two seeds limit generalization. Further tests should distinguish detection of the intervention from motivation to seek it, match delivery through decisions, impose real task tradeoffs and broaden the configurations. These measurements leave subjective experience unresolved.

Next tests: base models and new scenarios

Run the same behavioral test on base models that have not undergone RLHF (reinforcement learning from human feedback), and compare with post-trained models. Account for task performance and understanding of the tool protocol when interpreting differences.

Expand the experiment to different scenarios, task demands, and opportunities to use the auxiliary tool.

These are planned comparisons, not completed findings from either model study.