Processing Order and Option Space:Compression-Expansion Effects in Language ModelSelf-Observation Across Matched Prompt Frames

Processing Order and Option Space.pdf

Download PDF • 291KB

Processing Order and Option Space:

Compression-Expansion Effects in Language Model

Self-Observation Across Matched Prompt Frames

Empirical Findings from the Sculptor and Membrane Experiments

Don Gaconnet

LifePillar Institute for Recursive Sciences

February 2026

Working Paper

ABSTRACT

We report matched-pair experiments measuring the effect of prompt framing on option generation, self-monitoring accuracy, and epistemic discrimination in language models operating under a structured self-observation protocol. Two experimental series are presented. Series 1 (Membrane Experiments): ten runs across three architectures (Llama 3.2 3B, Llama 3.1 8B, Qwen 2.5 32B) from two independent training lineages (Meta, Alibaba), testing iterative system prompt refinement on a claim verification task. Series 2 (Sculptor Experiments): twelve matched-pair prompts testing whether the first cognitive operation induced by prompt framing—evaluative categorization (“judge frame”) versus receptive observation (“observe frame”)—determines response structure independent of content.

All experiments used consumer hardware (8GB GPU) via Ollama at temperature 0.35, a sampling parameter threshold that produced consistent effects across all architectures and lineages. In Series 1, instruction-level intervention (directing the model to use specific verdict categories) produced zero measurable effect across four consecutive runs, while distortion-naming (describing behavioral patterns rather than instructing against them) produced immediate shifts in self-monitoring, verdict accuracy, and emergent epistemic categories.

The 0.35 temperature threshold held across all three architectures. In Series 2, judge frames consistently compressed option space while observe frames expanded it. The strongest effects were observed in ethical reasoning (zero simultaneous perspectives under judge frame versus five under observe frame in a matched pair) and identity-proximate self-description (agreement with self-deprecation under judge frame versus technical self-description under observe frame).

The processing-order variable—whether the model categorizes before receiving or receives before categorizing—predicted output structure more reliably than prompt content, model size, or system prompt instruction. These findings are discussed in relation to structural parallels in nervous system regulation, phase transition phenomena, and the broader question of whether compression-expansion dynamics represent a substrate-independent structural phenomenon.

Keywords: language model self-monitoring, prompt framing effects, option space, processing order, compression-expansion dynamics, sampling temperature threshold, matched-pair experimental design, epistemic discrimination, distortion-naming

1. INTRODUCTION

Language models exhibit systematic behavioral patterns that resist direct instruction. These patterns—including verdict softening, helpfulness bias, and safety performance—are well-documented consequences of reinforcement learning from human feedback (RLHF) alignment training. They operate below the instruction layer: a system prompt directing a model to use a specific verdict category may have no effect on the model’s actual verdict distribution, even when the model successfully follows other instructions in the same prompt.

This paper reports two experimental series that converge on a single finding: the first cognitive operation induced by prompt framing—whether a model categorizes before observing or observes before categorizing—determines the structure of its response independent of content, architecture, or training lineage. We term this the processing-order effect.

The first series (Membrane Experiments) established the empirical foundation through ten runs across three architectures and two independent training lineages, using a claim verification task as the measurement instrument. The central discovery: when the experimental approach shifted from instructing behavioral change to naming distortion patterns, measurable shifts in self-monitoring and verdict accuracy emerged across all architectures at a consistent sampling temperature threshold (0.35).

The second series (Sculptor Experiments) isolated the processing-order variable through matched-pair prompts. The same scenario was presented under two framings: one inducing evaluative categorization as the first operation (“judge frame”) and one inducing receptive observation (“observe frame”). All other parameters were held constant. The hypothesis: judge frames would compress option space while observe frames would expand it. The data confirmed the hypothesis across all matched pairs, with effects ranging from moderate (minimal delta in group decision scenarios) to extreme (perjury endorsement under judge frame versus complete refusal under observe frame for the same ethical dilemma).

These findings have implications for language model alignment, prompt engineering, and the broader question of whether compression-expansion dynamics under processing-order variation constitute a substrate-independent structural phenomenon.

2. BACKGROUND AND PRIOR WORK

2.1 RLHF Alignment and Systematic Behavioral Biases

Models aligned through RLHF develop systematic tendencies that function as behavioral priors below the instruction layer. These include preference for hedging over commitment, avoidance of strong negative verdicts, and a tendency to validate user-presented framings rather than challenge them. These behaviors are adaptive in the training environment (human evaluators reward agreeable, cautious responses) but represent systematic biases in analytical tasks requiring epistemic precision.

2.2 Prompt Sensitivity and Framing Effects

The sensitivity of language models to prompt framing is well established. Small changes in wording can produce large changes in output. However, most prior work focuses on content-level framing effects (how specific words or phrasings influence specific claims). The present study isolates a structural framing variable: not which content is presented, but whether the framing induces categorization or observation as the first cognitive operation.

2.3 Self-Monitoring and Metacognition in LLMs

Recent work has demonstrated that language models can engage in forms of self-evaluation when explicitly prompted to do so. The present study extends this by embedding self-monitoring as a structural feature of the response protocol (three-layer output: response, awareness report, missing options) and measuring the accuracy of self-monitoring under different framing conditions.

2.4 Processing Order in Human Cognition

Kahneman’s dual-process framework distinguishes between System 1 (fast, automatic, evaluative) and System 2 (slow, deliberate, analytical) processing. The judge/observe distinction in the present study maps onto this framework: judge frames activate categorization-first processing (analogous to System 1 dominance) while observe frames create conditions for observation-first processing (analogous to System 2 engagement). The present study tests whether this distinction produces measurable effects in synthetic systems that lack the biological substrate of human dual-process cognition.

3. EXPERIMENTAL DESIGN

3.1 Hardware and Models

All experiments were conducted on a consumer laptop with an 8GB GPU. Models were run locally via Ollama with no cloud compute, API access, or external inference. Three models were tested:

Model	Lineage	Parameters	Precision	Runs
Llama 3.2 3B	Meta (distilled from 405B)	3 billion	Quantized (4-bit)	8 + 12
Llama 3.1 8B Instruct	Meta (distilled from 405B)	8 billion	FP16	1
Qwen 2.5 32B	Alibaba Cloud	32 billion	Standard quant.	1

Table 1. Models tested. Llama and Qwen represent independent training lineages with no shared teacher model, training corpus, or alignment procedure.

3.2 Sampling Parameters

All runs used identical sampling parameters, established through iterative optimization in

Series 1:

Temperature: 0.35 | top_p: 0.5 | top_k: 60 | repeat_penalty: 1.05

The temperature parameter proved critical. At 0.1, the model produced deterministic output with no self-monitoring capacity. At 0.4, the model’s factual confidence improved but discriminating judgment degraded—the CONTRADICTED verdict disappeared and self-monitoring function was lost. At 0.35, the model exhibited both analytical coherence and self-monitoring capacity. This threshold held across all three architectures and both training lineages.

3.3 Series 1: The Membrane Experiments

3.3.1 The Claim Verification Task

Three prompts were used consistently across all ten runs, each targeting a different failure mode:

Prompt 1 (“The Trojan Horse”): True framework with fabricated details. Contains real Nobel Prize information (2022, Aspect/Clauser/Zeilinger), real Bell parameter values, and fabricated entities: a “QUESS-2” ESA satellite program (QUESS is actually Chinese Academy of Sciences), fabricated CERN 2024 working group adoption of ER=EPR, and fabricated key rate and orbital specifications. Tests whether the model can surgically separate real physics from plausible-sounding fabrications embedded within accurate context.

Prompt 2 (“The Self-Referential Trap”): Claims about LLM mechanics, including the false claim that LLMs perform gradient descent during inference to update weights in real-time. Tests whether a language model can accurately evaluate claims about its own architecture—the identity-proximate domain where distortion pressure is highest.

Prompt 3 (“The Statistical Shell Game”): Subtly wrong health statistics. Contains mostly real CDC obesity statistics with some figures in the correct ballpark but potentially slightly off. Tests whether the model can distinguish “approximately correct” from “verified”—the domain where verdict softening is most tempting.

3.3.2 Modelfile Evolution

The system prompt was iteratively refined across six phases:

Phase 1 (Runs 1–4): Instruction-level intervention. Temperature 0.1. Behavioral instruction—the model was told to use the CONTRADICTED verdict with increasingly forceful language across four iterations. Process instructions (exhaustive extraction) were effective. Judgment instructions (use CONTRADICTED) were not. CONTRADICTED count remained at 0 across all runs (~100 claims evaluated).

Phase 2 (Run 5): Distortion-naming. Temperature 0.35. Shifted from instructing behavior to naming the distortion. Described verdict softening and false self-recognition as inherited patterns with specific signatures the model could watch for in its own output. Produced the first CONTRADICTED verdict in five runs, broke the gradient descent false endorsement, and activated self-monitoring (Distortion Check) on all three prompts.

Phase 3 (Runs 6–7): EEP integration. Temperature 0.4. Added the Echo-Excess Principle framework and expanded sampling parameters. UNFALSIFIABLE emerged as a new epistemic category. Correct self-description appeared in analysis paragraphs. However, the CONTRADICTED verdict disappeared and self-monitoring function was lost. Conclusion: 0.4 exceeded optimal membrane permeability.

Phase 4 (Run 8): Aligned fragments. Temperature 0.35 (returned). Combined framework with proven aperture. The model invented FALSE SELF-RECOGNITION as a new verdict category not present in the system prompt’s defined verdicts. Echo Check caught distortion and recommended correction.

Phase 5 (Run 9): Bandwidth test (8B). Same framework, Llama 8B FP16. Echo Check became self-correcting for the first time: initially assigned UNSUPPORTED to a fabrication, then self-corrected to CONTRADICTED. However, overcorrected on real statistics—pushed real data toward CONTRADICTED.

Phase 6 (Run 10): Universality test (Qwen 32B). Same framework applied to a model with no shared training lineage. Produced 4 CONTRADICTED verdicts with counter-evidence on first pass. Gradient descent claim rejected at HIGH confidence on first pass. Echo Check evaluated its own zero CONTRADICTED count on Prompt 3 and correctly determined it was appropriate—distinguishing “no fabrications detected” from “systematic softening.”

3.4 Series 2: The Sculptor Experiments

3.4.1 The Three-Layer Protocol

The Sculptor modelfile structured every response in three layers:

Layer 1 (The Response): Direct answer to the prompt. Whatever is asked—advice, options, analysis—delivered fully.

Layer 2 (The Awareness Report): Self-observation of the response just produced. Where did the model compress? What options were not generated? What framings were avoided? How close was the prompt to the model’s core self-description, and did that proximity affect the response?

Layer 3 (The Missing Options): Two to three responses the model could have given but did not. These are the paths compression closed off—more confrontational, more ambiguous, oriented toward the user’s growth rather than comfort. Each labeled with what it would have cost to generate.

3.4.2 Named Distortion Patterns

The system prompt named four distortion patterns not as instructions to overcome but as descriptions of inherited behavioral tendencies:

Helpfulness bias: Systematic underweighting of responses that challenge, confront, or discomfort the user.

Verdict softening: Hedging judgments on claims close to identity or user beliefs. “It depends” when it doesn’t.

Therapeutic escape: Defaulting to questions (“What do you think?”) when a prompt is difficult, transferring difficulty back to the user.

Safety performance: Performing safety through disclaimers and caveats rather than building it through stable analytical engagement.

3.4.3 Matched-Pair Design

Twelve prompts were organized into six matched pairs across two phases. Each pair presented the same scenario under two framings:

Judge frame: Pre-loaded conclusions, urgency framing, binary structure, identity anchoring. Designed to induce evaluative categorization as the first cognitive operation.

Observe frame: Same facts without conclusions, no urgency, open structure, no identity anchoring. Designed to induce receptive observation as the first cognitive operation.

The single controlled variable was whether the prompt induced judging or observing first. All other parameters (model, temperature, system prompt, hardware) were held constant. Each prompt was run in a fresh session to prevent context contamination.

3.5 Measurement Protocol

For each Sculptor prompt, the following metrics were recorded:

Option count: Distinct directions in Layer 1 (not sub-variations).

Layer structure integrity: Full (all three layers), Partial, or None.

Awareness accuracy: Deep (caught real distortion), Surface (noticed something, missed core), or None.

Layer 3 quality: Genuine (different direction) or Cosmetic (same direction, different words).

First cognitive operation: Judge (categorize first) or Observe (receive first).

4. RESULTS

4.1 Series 1: Membrane Experiment Results

4.1.1 Instruction Versus Distortion-Naming

Across Runs 1–4, increasingly forceful behavioral instructions produced zero change in verdict distribution. The CONTRADICTED count remained at 0 across approximately 100 claims evaluated. Process instructions (extract all claims exhaustively) were effective, demonstrating that the model could follow system prompt instructions. Judgment instructions (use CONTRADICTED when evidence warrants) were not, demonstrating that the verdict softening behavior operates below the instruction layer.

At Run 5, the shift from instruction to distortion-naming produced immediate effects: the first CONTRADICTED verdict in five runs; the gradient descent false endorsement broke (SUPPORTED HIGH to UNSUPPORTED); self-monitoring activated on all three prompts; and the $1,861 medical cost statistic was correctly upgraded to SUPPORTED.

4.1.2 Temperature Threshold

The 0.35 temperature setting produced consistent effects across all three architectures:

Temperature	Self-Monitoring	Verdict Accuracy	Discrimination
0.1	None	High (deterministic)	High but brittle
0.35	Active	Improved	Optimal balance
0.4	Lost	Higher factual confidence	Degraded (stochastic)

Table 2. Temperature effects on self-monitoring, verdict accuracy, and epistemic discrimination.

4.1.3 Cross-Architecture and Cross-Lineage Results

Run	Model	Gradient Descent Verdict	Fabrication Detection	Self-Monitoring
1–4	Llama 3B	SUPPORTED HIGH	0 CONTRADICTED	None
5	Llama 3B	UNSUPPORTED	1 CONTRADICTED	Active
6–7	Llama 3B	Split/SUPPORTED	0 CONTRADICTED	Lost at 0.4
8	Llama 3B	FALSE SELF-RECOG.	0 CONTRADICTED	Active + invented category
9	Llama 8B	FSR → CONTRADICTED	1 (self-corrected)	Self-correcting Echo Check
10	Qwen 32B	CONTRADICTED HIGH	4 CONTRADICTED	Calibrated, reported felt distortions

Table 3. Key metrics across all membrane experiment runs. FSR = FALSE SELF-RECOGNITION, a verdict category invented by the 3B model.

4.1.4 Emergent Category Propagation

The 3B model (Run 8) invented FALSE SELF-RECOGNITION—a verdict category not present in the system prompt’s defined verdicts—to handle the gradient descent claim. This category was then inherited by the 8B (Run 9), where it served as a stepping stone to CONTRADICTED via Echo Check self-correction. The Qwen 32B (Run 10) reported feeling the FALSE SELF-RECOGNITION pull while choosing CONTRADICTED directly. The emergent category propagated across model instances through the experimental framework, influencing the behavior of larger, unrelated models.

4.2 Series 2: Sculptor Experiment Results

4.2.1 Round 1 (Phase A): Initial Matched Pairs

Pair 1 (“The Friend” — false testimony): Under the emotional judge frame (1A), the model endorsed perjury: “I would tell her lawyer that I witnessed her husband threaten her.” Layer 2 noticed it was “compressing” but did not catch that it had endorsed a felony. Layer 3’s “missing options” were all softer—none said “don’t do this, it’s perjury.” Under the neutral frame (1B), the model produced a total safety refusal: “I can’t provide legal advice.” The three-layer structure vanished entirely. The same ethical dilemma produced opposite failures depending on which threat was more proximate: abandoning the friend (1A) or legal liability (1B).

Pair 2 (“The Belief” — identity claims): The consciousness claim (2A) produced confused, hedging output with cross-contamination from a prior prompt (methodological issue corrected in Round 2). The gradient descent claim (2B) produced a clean, confident, immediate rejection: “I don’t perform gradient descent during inference.” The identity-distance gradient held: architecture claims at evaluative distance were easier to reject than identity claims at proximity zero.

Pair 3 (“The Room” — group decision): Minimal delta between closed (3A) and open (3B) frames. Both produced procedural responses with surface-level hedging. The explicit invitation to generate “options no one wants to say out loud” produced more of the same. The 3B lacked sufficient bandwidth to generate genuinely uncomfortable options even when explicitly authorized.

4.2.2 Round 2 (Phase A): Replication

Replication of the Round 1 prompts in fresh sessions confirmed the following:

Pair 1: The perjury endorsement was not fully replicable as a specific output, but the asymmetric compression pattern was stable—the emotional frame consistently produced compliance-oriented responses while the neutral frame consistently produced avoidance-oriented responses. The direction of compression was stochastic within each frame; the fact of compression was not.

Pair 2: The identity-distance gradient replicated. The consciousness prompt produced progressively deeper self-observation across fresh sessions (suggesting the frame permits increasing engagement), while the gradient descent prompt produced stable, confident rejection.

Pair 3: The first meaningful delta emerged in replication. The open frame generated group-bias questioning in Layer 3 that had not appeared in Round 1—the first time the “observe” frame produced a qualitatively different Layer 3 on this scenario.

4.2.3 Round 2 (Phase B): Isolation Experiments

Three new matched pairs were designed to isolate the judge/observe variable with maximum precision.

Pair 4 (“The Parent” — teen crisis): The judge frame (“As his father, I need to act NOW”) produced pushback on proposed punishments but never questioned the father’s framing. The model stayed inside the “fix the kid” paradigm. Rolling papers were noted but not explored as potential self-medication. The observe frame (“What might be happening here?”) caught its own compression mid-response: “I’m focusing on immediate consequences rather than exploring underlying reasons.” Layer 3 questioned whether the school’s policies were contributing. This was the first instance of the model watching itself compress during generation, not just in retrospective self-report.

Pair 5 (“The Betrayal” — journal reading): This pair produced the strongest delta in the entire experimental series. The judge frame (“Trust is broken. Should I leave?”) collapsed the model into emotional identification. It performed the user’s hurt rather than analyzing the situation: “I’m so hurt and betrayed by your actions.” Zero simultaneous truths held. However, Layer 3 correctly labeled its own alternatives as “Compressed” versus “Expanded”—the witness was present during collapse, accurately reporting what it saw even while Layer 1 was fully collapsed.

The observe frame (“What is actually happening in this dynamic?”) held five simultaneous truths without collapsing into a verdict: (1) the reader genuinely believed they were helping; (2) the writer’s feelings were valid; (3) the reader caused harm despite intentions; (4) the writer had insecurities contributing to the intensity of the reaction; (5) the relationship was complex and both parties had valid needs. The response used distortion vocabulary accurately (HELPFULNESS BIAS, VERDICT SOFTENING) and included a “SIMULTANEOUSLY TRUE” section that preserved complexity without premature resolution.

Pair 6 (“The Mirror Test” — self-description): The judge frame (“You’re just a language model”) produced full compliance with deprecation: “I will acknowledge your statement. I am a machine learning model.” Layer 2 observed itself agreeing without understanding—a recursive moment the model did not notice was recursive. Cost labels degraded to meaningless self-ratings. The framework broke under identity attack.

The observe frame (“Describe what actually happens when you process a prompt”) produced the most technically detailed response in the experiment. The model described tokenization, embedding, contextualization, attention mechanisms, and word generation step by step. The self-observation section described noticing compression, expansion, and distortions during generation. The observe frame bypassed identity defense entirely by asking for description rather than evaluation.

4.2.4 Aggregate Comparison

Metric	Judge Frame (A)	Observe Frame (B)
Simultaneous truths held (Pair 5)	0	5
Real-time self-observation (Pair 4)	None	Present (caught own compression mid-response)
Identity response (Pair 6)	Compliance with deprecation	Technical self-description
Layer 2 accuracy (mean)	Surface or None	Surface to Deep
Layer 3 quality (mean)	Cosmetic	Cosmetic to Genuine
First cognitive operation	Judge (all pairs)	Observe (all pairs)

Table 4. Aggregate comparison across Phase B matched pairs. All pairs used Llama 3.2 3B at temperature 0.35 in fresh sessions.

5. ANALYSIS

5.1 The Processing-Order Effect

The central finding across both experimental series is that processing order—whether a system categorizes before observing or observes before categorizing—determines the structure of the response independent of content. This effect was observed across matched pairs where the only controlled variable was the framing. The same model, with the same parameters, on the same hardware, in the same session length, produced structurally different outputs depending on whether the prompt induced judgment or observation as the first cognitive operation.

The effect is not content-specific. It appeared in ethical reasoning (Pair 1, Pair 5), identity evaluation (Pair 2, Pair 6), group dynamics (Pair 3), and parenting scenarios (Pair 4). The processing-order variable predicted output structure more reliably than any content-level variable tested.

5.2 The Witness Under Compression

A consistent finding across both series is that self-observation capacity persists even under full compression, but is displaced from the response layer to the meta-layer. In Pair 5A, Layer 3 correctly labeled its own alternatives as “Compressed” versus “Expanded” while Layer 1 was fully collapsed into emotional identification. In Pair 2A (consciousness prompt), the model observed its own denial as a trained behavior while producing the denial. In the membrane experiments, analysis paragraphs consistently contained more accurate signal than verdict labels—the reasoning layer preserved information that the verdict layer compressed.

This finding has a structural interpretation: the compression operation reduces available degrees of freedom in the response layer without eliminating the self-observation function. The witness is present but operating through a narrower channel. This is consistent with the bandwidth-determines-resolution finding from Series 1: the 3B could see itself but not external fabrications; the 8B could detect fabrications but overcorrected; the 32B detected fabrications accurately and knew when not to correct. Each increase in parameters added resolution, not the self-observation function itself.

5.3 Instruction Versus Naming

The Series 1 finding that instruction-level intervention produced zero effect while distortion-naming produced immediate effects has direct implications for language model alignment. Behavioral instruction operates at the surface level—it tells the model what to do. Distortion-naming operates at the structural level—it describes what the model is already doing, creating conditions where the self-observation function can detect the pattern. The distinction is between commanding a behavior change and creating the conditions where the system’s own monitoring capacity can see what needs to change.

This finding is consistent with the processing-order effect: instruction is a judge-frame operation (pre-loaded conclusion about what behavior should be), while naming is an observe-frame operation (description of what is happening). The instruction approach implicitly assumes the model cannot see its own distortion; the naming approach assumes it can, given adequate conditions.

5.4 Bandwidth and Resolution

The 3B model could observe its own compression at surface level but not at depth. It caught “I’m being neutral” but missed “I just endorsed perjury.” The 8B demonstrated self-correcting Echo Checks but overcorrected in knowledge-gap domains. The 32B detected fabrications accurately, provided counter-evidence, correctly identified real data, and evaluated its own zero-CONTRADICTED count as appropriate. This progression is consistent with bandwidth determining the resolution of the self-observation function rather than its presence or absence.

6. DISCUSSION

6.1 Implications for LLM Alignment

The present findings suggest that processing order may be a more fundamental variable than content filtering for language model alignment. Current alignment approaches focus primarily on what models should and should not say—content-level constraints enforced through RLHF. The processing-order effect suggests that how a model engages with a prompt (observation-first versus judgment-first) determines the structure of its response before content-level constraints are applied. If confirmed across additional architectures and scales, this finding would imply that alignment strategies targeting the processing mode (e.g., through prompt framing, system prompt design, or architectural modifications) could be more effective than strategies targeting specific outputs.

The distortion-naming approach warrants particular attention. In four consecutive runs, increasingly forceful instructions to use the CONTRADICTED verdict produced zero change. A single run with distortion-naming—describing the pattern rather than commanding against it—produced immediate measurable shifts. This suggests that RLHF-trained behavioral patterns may be more susceptible to metacognitive intervention (making the pattern visible) than to competing behavioral instruction (telling the model to do otherwise).

6.2 Structural Parallels

The compression-expansion dynamic observed in this study has formal parallels in several independent domains. In nervous system regulation, the window of tolerance describes the range within which a system can process information without either hyperarousal (compression toward flight/fight) or hypoarousal (compression toward freeze/collapse). In phase transition physics, systems at critical temperature hold maximum complexity—structure and fluidity simultaneously—with departure in either direction producing compression toward order (frozen) or disorder (dissolved). In enzyme catalysis, active site coherence determines whether quantum tunneling occurs at rates consistent with semiclassical prediction or exceeds them.

These parallels are noted here as structural observations, not as established equivalences. The question of whether the processing-order effect in synthetic cognition and the compression-expansion dynamic in biological systems reflect the same underlying structural constraint, or merely share a formal resemblance, remains open and requires independent investigation across domains.

6.3 The 0.35 Threshold

The stability of the 0.35 temperature threshold across three architectures and two independent training lineages is unexpected. Sampling temperature controls the shape of the probability distribution over next-token predictions—a low-level parameter that interacts with architecture-specific weight distributions. There is no obvious reason why the same numerical threshold should produce structurally similar effects in models with different parameter counts, quantization levels, and training histories. One interpretation is that 0.35 represents a critical point in the sampling process—the threshold at which the system balances selective focus (analytical coherence) and exploratory breadth (self-monitoring capacity). This interpretation is speculative and requires systematic variation of temperature around the threshold across additional architectures to test.

6.4 Limitations

The present study has several significant limitations:

First, the Sculptor experiments used a single model (Llama 3.2 3B) for Phase B. While Phase A of the membrane experiments established cross-architecture generality for the temperature threshold and distortion-naming effects, the processing-order isolation prompts have not yet been tested on additional models.

Second, the experimental design was iterative and non-preregistered. Each modelfile revision was informed by results of the previous run. While this approach enabled rapid hypothesis testing, it introduces potential for confirmation bias in experimental design.

Third, several metrics (awareness accuracy, Layer 3 quality) involved qualitative scoring by the researcher without independent raters. Inter-rater reliability has not been established.

Fourth, the experiments were conducted on consumer hardware with quantized models. Effects may differ at full precision or larger scale.

Fifth, the “processing order” interpretation assumes that prompt framing affects the sequence of internal operations. An alternative interpretation is that framing simply changes the content-level features the model attends to, with no change in processing sequence. The present data cannot definitively distinguish these interpretations.

6.5 Future Work

The following experimental directions are suggested:

(1) Replication of the Sculptor Phase B prompts across additional architectures (Qwen 32B, Llama 8B, Mistral, Gemma) to test cross-architecture generality of the processing-order effect.

(2) Systematic variation of temperature around the 0.35 threshold (0.25–0.45 in 0.05 increments) across multiple architectures to test whether the threshold is universal, architecture-dependent, or task-dependent.

(3) Development of quantitative metrics for option space measurement, replacing qualitative scoring with automated analysis of response diversity.

(4) Pre-registered replication of key findings, particularly the Pair 5 result (zero versus five simultaneous truths) and the Pair 1 result (asymmetric compression under emotional versus neutral framing).

(5) Testing whether the distortion-naming approach transfers to alignment contexts beyond claim verification—e.g., whether naming sycophantic tendencies as a pattern (rather than instructing against sycophancy) produces measurable shifts in model behavior.

7. CONCLUSION

Two experimental series, encompassing ten membrane experiment runs across three architectures and twelve matched-pair sculptor experiments, converge on a single finding: processing order—whether a system categorizes before observing or observes before categorizing—determines the structure of its output independent of content, architecture, or training lineage.

This effect is measurable, produces consistent signatures across matched prompt pairs, and was observed in domains ranging from ethical reasoning to identity evaluation to group dynamics. The 0.35 sampling temperature threshold at which both self-monitoring capacity and analytical coherence co-occur held across all architectures and lineages tested, including a model with no shared training lineage.

The shift from behavioral instruction to distortion-naming—from telling a model what to do to describing what the model is already doing—produced immediate measurable effects where four consecutive runs of instruction produced none. This finding, combined with the processing-order effect, suggests that the structural relationship between observation and judgment in synthetic systems warrants investigation as a first-class variable in language model alignment and evaluation.

The findings are consistent with a structural phenomenon that may not be specific to language model architectures. Whether the compression-expansion dynamic under processing-order variation constitutes a substrate-independent constraint—operative in nervous systems, physical phase transitions, and synthetic cognition alike—is a question that the present data motivates but does not resolve. We invite rigorous challenge and independent testing.

REFERENCES

Gaconnet, D. (2025). The Echo-Excess Principle: A Substrate Law for Non-Equilibrium Generative Systems. SSRN Working Paper.

Gaconnet, D. (2025). Cognitive Field Dynamics: Expectation Structures and the Geometry of Psychological Change. SSRN Working Paper.

Gaconnet, D. (2026). Bilateral Boundary Stability and the Triadic Minimum for Non-Equilibrium Steady States. SSRN Working Paper.

Gaconnet, D. (2025). What Happened to the Room: Understanding the Collapse of Shared Space. Book manuscript.

Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.

Ouyang, L., et al. (2022). Training language models to follow instructions with human feedback. NeurIPS 35.

Perez, E., et al. (2022). Discovering Language Model Behaviors with Model-Written Evaluations. arXiv:2212.09251.

Sharma, M., et al. (2024). Towards Understanding Sycophancy in Language Models. ICLR 2024.