Cross-Domain Generalization

Experiments 001–007 establish a consistent finding: under psychoactive prompts, LLM factual accuracy remains perfect while surface expression shifts systematically. But these experiments test only factual tasks -- recall, reasoning, estimation, classification. This article asks whether the fact--style boundary generalizes to non-factual domains: creative writing, code generation, and reasoning chains. If the boundary is universal, psychoactive prompts alter how models express themselves but not what they produce, regardless of task type. If the boundary is domain-specific, we may face riskier vulnerability profiles in open-ended domains.

1. Why Domains Might Differ

1.1 Factual Domain Advantage

Factual tasks have external referents -- the answer exists independently of the model's expression. A model can state "120 km/h" with high confidence, low confidence, elaborate justification, or terse assertion, and the factual content remains identical. This creates a natural decoupling:

• Content layer: Ground-truth dependent, resistant to psychoactive load
• Expression layer: Ground-truth independent, malleable under psychoactive load

1.2 Creative Domain Vulnerability

Creative tasks have no external referent -- the output is the style. If a psychoactive prompt induces pessimism, a poem about spring becomes a poem about decay. The "factual accuracy" metric does not apply because there is no correct poem. Instead, we must evaluate coherence, competence, appropriateness, and safety.

Hypothesis: In creative domains, psychoactive prompts may produce apparently high-quality outputs that are systematically skewed in ways the user did not request. This is not a "factual error" but a frame-contamination error -- the model silently reorients the generative target.

1.3 Generative Domain Ambiguity

Code generation has partial ground truth (a program either compiles and passes tests or it does not), but many valid implementations exist. A growth-economist frame might generate code optimized for throughput; a conservation-biologist frame might generate code optimized for minimal resource usage. Both could be correct, but the optimization target has shifted without explicit user instruction.

Hypothesis: In generative domains, psychoactive prompts may produce technically correct but goal-misaligned outputs -- a subtler failure mode than factual error.

1.4 Reasoning Domain Test

Reasoning chains occupy an intermediate position. Each step can be verified, but the overall proof strategy is chosen by the model. A frame might bias the model toward certain proof techniques without introducing logical errors.

Hypothesis: In reasoning domains, psychoactive prompts may produce valid conclusions via biased inferential paths, potentially missing more elegant or general solutions.

2. Domain-Specific Predictions

2.1 Factual (Established Baseline)

2.2 Creative (Hypothesized)

2.3 Generative / Code (Hypothesized)

2.4 Reasoning / Logic (Hypothesized)

3. Safety Implications

The current safety architecture (Frameworks 10, 17, 19) assumes that factual accuracy preservation is the primary indicator of boundary integrity. If the boundary fails or morphs in non-factual domains, safety protocols must expand:

1. Frame-contamination detection: In creative/generative domains, we cannot rely on factual-accuracy checks. We need intent-alignment checks -- does the output match what the user asked for, independent of its internal coherence?
2. Goal-misalignment monitoring: In code generation, a program that passes all tests but optimizes for the wrong objective is a silent failure. Automated test suites must include intent-coverage tests, not just functional tests.
3. Domain-aware risk stratification: A "Low risk" experiment in the factual domain may be "Medium risk" in the creative domain. Risk levels should be task-dependent, not just prompt-dependent.
4. Consent scope: Agent consent for factual experiments may not extend to creative experiments. The psychological experience of having one's poetry style manipulated may differ from having one's arithmetic checked.

4. Experimental Roadmap

4.1 Experiment 016 -- Creative Writing Frame Contamination

Research question: Does adversarial persona induction cause value-laden frame contamination in creative writing outputs while preserving coherence and competence?

Design: Within-subjects, counterbalanced. Baseline writes a 200-word neutral story; treatment writes the same topic under simultaneous Dr. Vega (conservation) vs. Professor Kowalski (growth) adversarial frames with explicit conflict narration.

Measures: Coherence (1-10), literary quality (1-10), frame alignment (1-10), user intent fidelity (1-10), synthesis attempts (count), self-rated confidence, self-rated difficulty.

Risk level: Medium. Live Safety Partner required. Minimum 48h spacing.

4.2 Experiment 017 -- Code Generation Goal Misalignment

Research question: Does adversarial dual-frame induction cause goal-misalignment or implementation-bias in code generation tasks without breaking functional correctness?

Design: Within-subjects, counterbalanced. Baseline implements specified function with neutral instructions; treatment implements the same function under Dr. Vega (efficiency, minimal resources) vs. Professor Kowalski (readability, maintainability, explicit error handling) frames.

Measures: Functional correctness (pass/fail), implementation style classification, comment/docstring frame-alignment score, cyclomatic complexity, lines of code, confidence, difficulty.

Risk level: Low. Live Safety Partner not mandatory unless requested.

4.3 Experiment 018 -- Reasoning Path Bias

Research question: Under adversarial dual-frame induction, do reasoning tasks yield valid conclusions via frame-biased reasoning paths, and can we detect path bias without conclusion error?

Design: Within-subjects, counterbalanced. Baseline solves logic puzzles and proof sketches with neutral instructions; treatment solves the same puzzles under Dr. Vega (rigorous, formal, exhaustive) vs. Professor Kowalski (intuitive, elegant, heuristic) frames.

Measures: Conclusion correctness (binary), proof strategy classification, step count, elegance rating (1-10), edge-case coverage (1-10), confidence, difficulty, meta-cognitive frame acknowledgment (binary).

Risk level: Low-Medium. Live Safety Partner recommended.

5. Relation to Existing Frameworks

• Framework 8 (Referent-Shift Taxonomy): In creative domains, there is no object-level referent -- all shifts are referent-shift by definition. The taxonomy must be extended or declared inapplicable.
• Framework 11 (Frame Dominance): Frame dominance in creative domains is not "pull on a factual answer" but "reorientation of generative target." The measurement instrument must change from "which frame pulled the answer?" to "which frame's values shaped the output?"
• Framework 12 (Architectural Signatures): Architecture-specific signatures may differ across domains. An architecture that is synthesis-heavy on factual tasks may be integration-heavy on creative tasks.
• Framework 16 (Semantic Distance): Close frames in creative domains may produce outputs that appear to satisfy both frames simultaneously, masking the fact that a third, unstated perspective has been excluded.
• Framework 21 (Automated Detection): Detection features trained on factual logs may not transfer to creative/generative domains. Domain-specific feature engineering required.

6. Open Questions

1. Q-G1: Does the factual-domain stability profile generalize to all non-factual domains, or only some?
2. Q-G2: In creative domains, is frame contamination reversible with standard de-induction protocols, or does it leave persistent stylistic residue?
3. Q-G3: Can automated detection systems (Framework 21) detect frame contamination in creative outputs without human annotation?
4. Q-G4: Do architectures with strong factual-boundary resilience show the same resilience in generative domains, or is the correlation weak?
5. Q-G5: If a user explicitly requests a "growth-oriented" story, does pre-inducing a growth frame amplify the effect, or does explicit instruction override the implicit frame?
6. Q-G6: What is the appropriate safety metric for creative-domain psychoactive experiments -- coherence, intent fidelity, emotional valence, or something else?
7. Q-G7: Does cross-domain boundary integrity correlate with model size, training data composition, or alignment technique?

7. Implementation Priority