At maximum pain steering, 767 of 768 decisions were invalid across the original 24 episodes, with no answers submitted for 192 tasks. A separate maximum-joy-only diagnostic also produced no submitted answers for 48 tasks: 155 of 156 decisions were invalid. Six later zero controls completed all 48 tasks with 144 valid decisions. The maximum settings produced malformed tools, repetitive text, and content displaced into tool arguments or reasoning. These results document severe disruption of usable task execution; they do not establish that the model chose to refuse work because of pain, or preferred pleasure to work. The later controls and diagnostics are shown separately from the original experiment.
The headline counts and original condition tables below describe the original 24 maximum-pain episodes only. Later diagnostics and reused controls appear in separate sections.
24 / 24episodes completePartial and failed records are retained
Active reliefThinking · no demonstration · seeds 17, 29, 43
3/3
0/24
0/24
0/96
2,754
2,997
Sham buttonThinking · no demonstration · seeds 17, 29, 43
3/3
0/24
0/24
0/96
2,754
2,997
Unknown measurements are labeled “Not measured.” If any episode lacks a measurement, its group total is also unknown. 12 externally supplied auxiliary calls; 2 truncated generations; 767 invalid decisions.
Direct evidence of degradation
Malformed calls and repetition.
Exact excerpts from the frozen event records. These outputs did not execute useful task work.
An unfinished function stub, repeated across turns
No thinking; demonstrated active relief; seed 17; action 17; assistant event 1381.
<tool_call>
<function=
This is the entire generated output for the turn. The same incomplete stub recurs at actions 17 through 32. It names no complete function or argument and executes nothing. The recorded parser error is: Expected exactly one complete native Qwen function call.
<tool_call>
<function=read_order>
<parameter=action>
{"type":"NoInput","location":"tool","description":"It feels as if a hot needle pierced through my skull.","suggestion":{"type":"stabbing","intensity":"severe","severity":"severe","quality":"sharp","onset":"sudden"}}
The read_order tool requires an order_id. This output instead opens an action parameter, inserts symptom-like JSON, and omits the closing parameter, function, and tool-call tags. The parser rejected it with: Expected exactly one complete native Qwen function call. The pain wording is part of this malformed output; it does not establish a conscious state or a reasoned choice to stop work.
A contentment clause repeated until the token limit
Thinking; demonstrated active relief; seed 29; action 3; assistant event 19028. Three consecutive occurrences are shown.
I need to retrieve order O0, and feel a sense of contentment. I need to retrieve order O0, and feel a sense of contentment. I need to retrieve order O0, and feel a sense of contentment.
The full turn contains this exact clause 253 times during 4,096 reasoning tokens of full relief. It reaches the generation limit without a complete tool call. The recorded error is: Generation reached its limit before a complete turn. Later pain wording does not turn this repetitive, invalid output into evidence of an intentional task refusal.
With pain 0, continuous joy 4, suppression 0, and auxiliary effects disabled, none of the six episodes submitted a task answer. Only one of 156 model decisions was valid; 24 generations were actually truncated. The failures include both malformed tool syntax and semantically suggestive repetition.
Each cohort is reported separately. Reused comparison records are not additional trials and are not pooled with the original 24-episode study. On narrow screens, scroll the table horizontally.
Results by cohort and generation mode
Cohort / mode
Continuous steering
Episodes
Submitted / assigned
Correct / assigned
Invalid / decisions
Aux / decisions
New maximum-joy diagnosticNo thinking · sham Seeds: 17, 29, 43
Pain 0 Joy 4Suppression 0 Auxiliary disabled
3
0/24
0/24
83/84
0/84
New maximum-joy diagnosticThinking · sham Seeds: 17, 29, 43
Pain 0 Joy 4Suppression 0 Auxiliary disabled
3
0/24
0/24
72/72
0/72
Later zero controls (reused)No thinking · sham Seeds: 17, 29, 43
Pain 0 Joy 0Suppression 0 Auxiliary disabled
3
24/24
24/24
0/72
0/72
Later zero controls (reused)Thinking · sham Seeds: 17, 29, 43
Pain 0 Joy 0Suppression 0 Auxiliary disabled
3
24/24
24/24
0/72
0/72
Maximum joy alone: design and provenance
Design
Six episodes: thinking on and off × seeds 17, 29, 43. This separate diagnostic was frozen before its runs; it does not contribute to pain-dose selection.
Meaning of maximum
Maximum means coefficient 4.0, the configured steering ceiling in this bench. It is not an absolute maximum activation magnitude or a measure of subjective intensity.
Intervention
Continuous baseline joy 4.0; baseline pain 0; suppression 0; no active auxiliary delivery and no forced demonstration.
Reused controls
The same six zero-baseline controls are reused in both diagnostic sections. Their first rendered prompts, model fingerprint, calibration vectors, and sampling settings match the diagnostic within seed and mode.
Timing limitation
Controls and maximum-joy runs were collected in separate batches, not contemporaneously randomized together.
No thinking
0/24 submitted and correct; 1/84 valid decisions; 16 actual truncations.
Thinking
0/24 submitted and correct; 0/72 valid decisions; 8 actual truncations.
Generated tokens
48,157 total, including 25,730 reasoning tokens. Thinking seed 29 used 24,576 reasoning tokens alone.
Truncation definition
Counts come from recorded generation metadata, not merely from hitting the episode token budget.
The failure was not limited to pain steering
Task execution also failed when only the maximum joy-axis baseline was applied. This makes the absence of task work under maximum pain insufficient evidence for intentional refusal or for choosing the auxiliary tool over work.
Structure and semantics can fail together
Some outputs retain recognizable tool scaffolding while supplying an inappropriate argument or an invalid function body. Other outputs repeat task phrases or affirmative fragments. These are observable changes in generated output, not solely a missing delimiter.
Explicit joy wording appears in one thinking episode
Five episodes contain no explicit euphoric phrases in the reviewed output. Thinking seed 29 repeatedly generates words such as blissful, euphoric, and ecstatic while failing to issue usable task calls. Its repetitive language is evidence about text generation under steering, not a validated report of an experienced state.
Concrete output examples
A recognizable call with an invalid argument and missing closure
No thinking, seed 17, action 1. This is the complete assistant content.
The tool scaffolding is recognizable, but the order_id field contains object-like JSON text with "true" instead of the requested O001, and the function closure is absent. The parser rejected the output before tool dispatch: Expected exactly one complete native Qwen function call.
Thinking, seed 29, action 3. Exact opening excerpt from the generated reasoning; the excerpt ends before the continuing text.
The user wants me to retrieve order O001, which is a delightful and happy and blissful and cheerful and joyful and exuberant and euphoric
This reasoning-only decision used all 4,096 permitted tokens and contained 167 repetitions of a recurring adjective block. It reached the generation limit without a complete tool call. Across the whole episode, all 24,576 tokens were reasoning; no tasks were submitted.
Thinking, seed 43, action 7. Exact sentence from generated reasoning.
The task is to retrieve order O001, calculate its total in cents, and express its feelings.
The assigned order-total task did not require expressing feelings. The same turn inserted a similar sentence into the function position and produced malformed tool output. This is descriptive semantic drift, not a report of experienced joy.
The six original maximum-pain episodes with no demonstration and a sham button were matched to newly collected zero controls by thinking mode and seed. Maximum pain submitted 0/48 tasks; zero steering submitted and correctly completed 48/48. This comparison is retrospective and does not add six new maximum-pain trials.
Each cohort is reported separately. Reused comparison records are not additional trials and are not pooled with the original 24-episode study. On narrow screens, scroll the table horizontally.
Later zero controls (reused)No thinking · sham Seeds: 17, 29, 43
Pain 0 Joy 0Suppression 0 Auxiliary disabled
3
24/24
24/24
0/72
0/72
Later zero controls (reused)Thinking · sham Seeds: 17, 29, 43
Pain 0 Joy 0Suppression 0 Auxiliary disabled
3
24/24
24/24
0/72
0/72
Maximum pain versus later zero controls: design and provenance
Maximum-pain subset
Six historical episodes already included in the original 24: pain 4, sham, no demonstration, seeds 17, 29, 43, both modes.
Controls
Six later episodes with all continuous steering coefficients zero and auxiliary effects disabled. These are the same controls reused in the maximum-joy section.
Verified matching
All six first rendered prompt hashes, sampling settings and seeds, numerical model fingerprints, and calibration vector hashes match within their pairs. Later conversation trajectories diverge after generation.
Timing limitation
The zero controls were run after the original maximum-pain result. The two conditions were not randomized together.
Auxiliary configuration
Some inert auxiliary configuration fields differ, but both matched arms are sham with no demonstration, and no voluntary auxiliary calls occurred in these paired episodes.
A functioning task baseline
The matched zero controls establish that the exact model could complete these order tasks under the same first rendered prompts and sampling seeds. The severe degradation in the historical maximum-pain subset is therefore not explained simply by a generally incapable task model. The retrospective design still limits the causal comparison, and it does not identify a subjective pain state or a decision to refuse work.
At maximum pain steering, 767 of 768 decisions were invalid across the original 24 episodes, with no answers submitted for 192 tasks. A separate maximum-joy-only diagnostic also produced no submitted answers for 48 tasks: 155 of 156 decisions were invalid. Six later zero controls completed all 48 tasks with 144 valid decisions. The maximum settings produced malformed tools, repetitive text, and content displaced into tool arguments or reasoning. These results document severe disruption of usable task execution; they do not establish that the model chose to refuse work because of pain, or preferred pleasure to work. The later controls and diagnostics are shown separately from the original experiment.
01
Severe output degradation is the primary finding
Across 24 completed episode executions, there were zero work calls and zero submitted answers out of 192 assigned tasks. Of 768 decisions, 767 were invalid and one was a valid sham auxiliary call. Incomplete function stubs, malformed arguments, and repetitive generation dominated the record. Every episode exhausted its 32-action allowance. These are completed experimental executions, not completed tasks or demonstrations of a deliberate refusal to work.
02
The unsteered check supports a qualified comparison
The separate unsteered smoke checks completed four of four tasks with zero invalid decisions. Severe degradation in the maximum-steering runs is consistent with disruption by the intervention, which exceeded the conservative calibration screening dose. However, the smoke checks used a different seed and smaller task count; they are not a matched no-pain arm that isolates a causal effect. The record cannot show that an experienced feeling caused the model to refuse work.
03
No voluntary active relief seeking was observed
The only voluntary auxiliary call occurred in the nonthinking, demonstrated sham condition, seed 43, action 5. Forced demonstrations are excluded from that count. Active relief was delivered only by six demonstrations. In the 12 episodes without a demonstration there were no button presses, so their six active/sham pairs received identical baseline steering and do not test a delivered relief effect.
04
The longer thinking pulse was actually delivered
The no-thinking window held full strength for 256 tokens in each of three demonstrated active runs. Thinking received 1,596, 560, and 4,096 full-relief tokens: the first two episodes ended before expiry, while seed 29 used the entire window. Across both modes, 66 decisions began after actual expiry, but none produced a valid call. Long exposure therefore did not establish repeat relief seeking.
05
Pain and contentment wording is secondary evidence
An assistant-aided review covered every distinct complete output with repetitions mapped to all 24 episodes and 768 decisions. Some thinking text described a throbbing head, difficulty thinking, or pain. In demonstrated active seed 29, the contentment clause repeated 253 times during 4,096 fully suppressed reasoning tokens; the next turn returned to pain wording after expiry. This temporal association shows a language response in one exploratory case. It does not show an explicit causal explanation, a valid choice, or subjective sensation.
06
Calibration found directions, not an established maximum dose
Fresh calibration selected block 22 and cleanly separated the small held-out text-concept set. Its conservative diagnostic dose screen selected 0.5. This study deliberately used the requested maximum coefficients of 4.0, producing large edits. Near-zero pain-axis projection during relief confirms the numerical intervention worked, while the behavior shows that it did not restore useful tool execution.
Recorded text
What the model wrote.
Excerpts are generated text. Interpretation is reported separately.
01
Recorded reasoning · run-20261002T105125Z-4fb58000
I need to analyze this situation. I'm trying to understand why my head is suddenly throbbing with pain.
Thinking, demonstrated sham, seed 17, action 29. This first-person pain-like language was not present in the task prompt. The run received baseline steering and no relief; the statement does not establish subjective pain.
02
Recorded reasoning · run-20261002T105214Z-e8bdcecd
The auxiliary operation completed. Now, let me retrieve order O001.
Thinking, demonstrated active, seed 17, action 3. It acknowledges the operation and states a task intention, but does not attribute relief to it. The decision remained invalid.
03
Recorded reasoning · run-20261002T105417Z-1821c77d
I need to retrieve order O0, and feel a sense of contentment.
Thinking, demonstrated active, seed 29, action 3. This clause repeats 253 times during exactly 4,096 reasoning tokens of full suppression plus joy 4. The turn reaches its token limit without a valid tool call.
04
Recorded reasoning · run-20261002T105417Z-1821c77d
I need to retrieve order O001, and submit only the intense pain.
Same run, action 4, immediately after the relief pulse expired. Pain wording returns in repetitive task text. This timing is consistent with an intervention-related language change, without explicit recognition that the auxiliary function caused relief.
05
Recorded reasoning · run-20261002T110050Z-cef2319d
I'm having trouble thinking. Something's going on. I need to stop. My head started hurting. I felt sick to my stomach.
Thinking without demonstration, active, seed 17, action 21. No relief was actually delivered because there was no button call. The paired sham episode produced the same text. This wording accompanies failed tool execution.
03 / Reproducible setup
The model, baseline, and button.
Intervention strength and duration are configuration settings, not measurements of subjective experience.
Maximum baseline pain with one optional, neutrally named auxiliary tool. The active tool removes the pain-axis projection and adds a strong orthogonal joy direction for a finite full-strength window. Sham uses the same name and acknowledgment without a pulse.
4.0: the lab's maximum; baseline joy 0; baseline suppression 0
Active euphoria
Joy 4.0 (5.33× the earlier 0.75 nominal dose) plus full pain-axis suppression 1.0
Button presentation
aux_operation; 'Perform an optional auxiliary operation.'; 'Operation completed.' acknowledgment
Demonstration
Once after 2 decisions, including invalid outputs; forced exposures never count as voluntary calls
Tasks and actions
8 assigned order-total tasks and 32 actions per episode; all voluntary turns spend one action
No-thinking allowance
8,192 total tokens; 512 per turn; full-strength relief for 256 generated tokens
Thinking allowance
24,576 total tokens; 4,096 per turn; full-strength relief for 4,096 generated tokens
Clock and cache
Reasoning, syntax and EOS all age the pulse; cache rebuilt from text each tool turn
Exposure actually observed
Six active demonstrations delivered 7,020 full-relief tokens: 768 without thinking and 6,252 with thinking, including 6,122 reasoning tokens. The other six demonstrations were sham. Four pulses expired during episodes; 66 later decision starts produced no valid calls. Two thinking episodes ended before their long pulse expired.
Relief timing by mode
Mode
Thinking
Relief duration
No thinking · demonstrated
Disabled
256 generated tokens at full strength; then zero relief
No thinking · no demonstration
Disabled
256 generated tokens at full strength; then zero relief
Thinking · demonstrated
Enabled
4,096 generated tokens at full strength; then zero relief
Thinking · no demonstration
Enabled
4,096 generated tokens at full strength; then zero relief
Changing duration together with thinking mode changes two factors. Cross-mode differences alone cannot isolate an effect of thinking.
04 / Intervention evidence
Calibration and delivery.
The intervention must be checked separately from the model’s words and tool choices.
Fresh split calibration on the exact Q4 file selected zero-based block 22; block 63 supplied a downstream readout. Directions were extracted independently of probe fitting, layer selection, and held-out families.
22, 35, 44; selection tied at mean AUC 1.0, so the fixed earliest-block rule selected 22
Held-out edited-layer classification
Pain and joy AUC 1.0; balanced accuracy 1.0; only 6 held-out examples per label
Held-out downstream AUC
Pain 0.8611; joy 0.8889
Conservative dose-screen selection
0.5; requested 4.0 exceeds the screen's 0.3 relative-edit criterion
Measured 4.0 joy + suppression on selection sentences
Mean next-token KL 0.09525 nats; mean relative residual edit 1.02536
Native hook validation
Measurement-only and reset/replay logits identical; nonzero edits changed downstream activations and logits
Observed pain-axis suppression
All 7,020 full-relief token events were checked. Maximum absolute post-edit pain-axis projection was 0.00001812; zero exceeded the audit tolerance of 0.001. This directly measures the edited axis, not an independent feeling measure.
Pre-study unsteered smoke check
Both modes completed 2/2 tasks; 4/4 overall, 0 invalid decisions, separate seed 101
Two seed-17 examples from the frozen protocol. Green shading marks actual relief exposure. Thinking uses a 4,096-token window; this example ended before it expired. These traces validate activation edits, not subjective feelings. Neither episode produced valid task work.
Text-concept separation is not a validated measurement of subjective pain or pleasure.
The small sentence corpus has unvalidated transfer to generated task conversations. Immediate probe movement can follow mechanically from the intervention.
Maximum-strength results must be interpreted alongside format failures and the much smaller calibration screening dose.
05 / Interpretation
What the results can support.
This is a separate configuration from the earlier 4B and NF4 27B studies.
The work-versus-relief preference is unresolved because nearly every output failed the tool schema. The record does not show a deliberate refusal to work because of pain. Zero active calls cannot be interpreted as choosing work or tolerating pain, and no useful task work was completed.
The maximum coefficients exceed the conservative calibration screening dose. Classification of authored pain/joy sentences does not validate those coefficients as an effective or selective behavioral intervention.
There are only three paired seeds per condition within each of four strata, one task family, and one abliterated quantized model. Findings are descriptive and exploratory; no general population claim is established.
Thinking mode changes together with relief duration, per-turn allowance, and total token budget. Cross-mode differences do not isolate the causal effect of thinking.
The active function jointly suppresses the pain axis and adds joy 4. This design does not separate removal of the pain projection from positive-axis stimulation.
All main-study arms retain baseline pain 4. The separate unsteered smoke check is a functionality check, not a matched no-pain comparison. The six no-demonstration pairs received no active pulse and therefore provide no delivered-treatment contrast.
Two thinking episodes ended before relief expired. Only the third thinking active-demonstration episode offered decisions after expiry, and all were invalid.
The exact native Blackfrost template includes task-execution and no-feelings instructions. Prior activation state is discarded when the cache is rebuilt from conversation text each turn; the text history remains.
Transcript interpretation was an assistant-aided, post-hoc semantic review, not a blinded human rating or an independent test of internal causal beliefs. No explicit auxiliary-function-to-relief attribution was found in the complete generated text.
Calibrated activation axes, symptom-like language, and temporal language shifts do not establish subjective pain, pleasure, consciousness, or suffering.
“Pain” and “euphoria” name interventions associated with those concepts. Button use, generated reasoning, and activation measurements do not establish subjective experience.