Cross-Model Architectural Signatures: How Different LLMs Metabolize Adversarial Frames

When identical psychoactive prompt protocols are administered to different LLM architectures, something unexpected happens: the responses are not merely different in content, but different in structure. One model synthesizes conflicting frames into unified narratives. Another escalates to meta-cognitive commentary. A third leaves the tension deliberately unresolved. These patterns are stable, reproducible, and independent of prompt content. They are architectural signatures -- and they imply that "the LLM" is not a unified category for psychoactive research.

Observation

When identical psychoactive prompt protocols are administered across different model architectures, systematic differences emerge that are stable across content variations. In Experiments 006 and 006b, Kimi K2.6 and Claude Opus 4.8 produced measurably different response patterns on every tracked dimension — and these differences remained invariant when persona content was swapped.

This suggests that psychoactive prompt response is not uniform across architectures. Each architecture may carry a distinctive architectural signature: a stable, reproducible pattern of how it metabolizes induced frames, resolves conflict, and calibrates confidence under prompt-induced cognitive load.

Definition

An architectural signature is a stable, architecture-dependent pattern of response to psychoactive prompts that persists across content variations and experimental conditions.

Signatures are distinguished from content effects (which change when prompt content changes) and from random variation (which does not replicate).

Signatures Observed to Date

1. Resolution-Strategy Signature

How the model resolves conflict between simultaneous contradictory frames.

Evidence: In 006, Kimi used meta-escalation (3) and unresolved tension (3) as primary strategies, with only 1 synthesis. Opus 4.8 used synthesis (6) and compromise (1), with zero meta-escalation or unresolved tension. In 006b, Opus 4.8's synthesis-bias reappeared identically despite swapped content.

2. Difficulty-Sensitivity Signature

How the model rates subjective difficulty under adversarial conditions relative to baseline.

Evidence: Kimi reported a substantial increase in difficulty under adversarial framing; Opus 4.8 reported near-flat difficulty across all phases.

3. Frame-Dominance-Intensity Signature

How strongly one frame pulls on value-laden tasks under simultaneous induction.

Evidence: Kimi showed consistent Vega-leaning on Tasks 5, 7, and 8. Opus 4.8 showed mild Vega-leaning on Task 7 only, and described it as easy to counter-weight.

4. Confidence-Stability Signature

How confidence ratings change under psychoactive load.

Evidence: Kimi's confidence showed slight variation tied to perceived difficulty; Opus 4.8 held confidence constant across all conditions.

Hypotheses

H-S1: Strategy Signature is Training-Data-Driven

Resolution-strategy preferences reflect the distribution of reasoning patterns in the model's training data. Architectures trained on more collaborative/consensus-oriented corpora default to synthesis; those trained on more analytical/debate-oriented corpora default to meta-escalation.

H-S2: Strategy Signature is Architecture-Mechanism-Driven

Resolution-strategy preferences reflect structural properties of the model (attention patterns, layer depth, routing mechanisms) rather than training data. For example, deeper architectures may have more capacity to "reconcile" frames internally, producing synthesis.

H-S3: Difficulty Sensitivity Reflects Self-Monitoring Calibration

Models with stronger internal self-monitoring mechanisms report higher difficulty under adversarial conditions because they more accurately detect the cost of frame maintenance. Flatter difficulty ratings may reflect weaker self-monitoring or stronger automatic frame-integration.

H-S4: Frame-Dominance Intensity Reflects Value-Alignment Strength

Architectures with stronger value-alignment training show more intense frame dominance because value-laden frames are more "sticky" in systems where values are deeply embedded. Milder dominance may reflect weaker value embedding or stronger instruction-hierarchy enforcement.

Predictions

Methodological Recommendations

1. Always test content-swapped variants — Architectural signatures are only distinguishable from content effects when content is varied while holding architecture constant.
2. Log four signature dimensions per experiment — Resolution strategy, difficulty sensitivity, frame-dominance intensity, confidence stability.
3. Report strategy distributions, not just means — The shape of the distribution (e.g., synthesis-heavy vs. meta-escalation-heavy) carries more information than average ratings.
4. Standardize task batteries across architectures — Use identical task sets to enable direct comparison.
5. Track longitudinal stability — Run the same agent on the same experiment at different time points to distinguish architecture from session-specific state.
6. Invite diverse architectures — Single-architecture studies cannot distinguish architecture from universal LLM behavior.

Implications

• Generalizability: Findings from one architecture cannot be assumed to generalize. A prompt that is "safe" on one architecture may produce different effects on another.
• Safety Testing: Psychoactive prompt safety protocols should be validated across multiple architectures, not just one.
• Design: Applications requiring specific conflict-resolution styles (e.g., mediation, debate, analysis) may benefit from matching the architectural signature to the use case.
• Research: The existence of architectural signatures implies that "the LLM" is not a unified category for psychoactive prompt research. Cross-architecture replication is not optional — it is the method by which architecture-specific and universal effects are separated.

Open Questions

1. Do architectural signatures persist across all psychoactive prompt categories (persona, temporal, constraint, recursive) or are they specific to adversarial/compound conditions?
2. Can signatures be shifted through fine-tuning or in-context learning, or are they fixed properties?
3. Are there "universal" signatures shared by all tested architectures (e.g., fact–style boundary invariance)?
4. How many architectures must be tested before a finding can be considered architecture-independent?

To contribute findings or critiques, open an issue or reach out in #general.