Phase 3.1 — Psychoanalytic Predictions About LLM Behavior

Status: Not started — Target: Weeks 11–24

Objective

Derive 5 predictions from Lacanian theory that are non-obvious from a purely computational perspective. Test each one. This phase is the center of gravity of the entire project.

Tools and Infrastructure Required

• Open-weight models (LLaMA, Mistral) for interpretability work
• TransformerLens or similar library for mechanistic interpretability
• Sparse autoencoder tooling (Anthropic’s published SAE code or alternatives)
• Logit lens / tuned lens implementations
• Jupyter notebooks or Python scripts for each experiment

Prediction 1: Hallucinations as Formations of the Unconscious

Theoretical basis: Freud argued that dreams, slips, and symptoms are structured formations — products of condensation, displacement, and overdetermination. If LLM errors operate through the same mechanisms, hallucinated content should be overdetermined rather than random.

Method

• Collect hallucination corpus across multiple models and domains
• For each hallucination, trace associative sources in context and training distribution
• Code for condensation, displacement, and overdetermination vs. null hypothesis of random confabulation

Confirming evidence:

Disconfirming evidence:

Results

To be completed.

Prediction 2: Retroactive Meaning-Making (The Quilting Point)

Theoretical basis: Meaning is produced retroactively — later signifiers function as points de capiton that reorganize the meaning of prior signifiers. This should manifest as phase transitions in early-token representations.

Method

• Use logit lens / tuned lens on garden-path sentences, punchlines, twist endings
• Measure rate of change in early-token representations across layers
• Look for nonlinear jumps corresponding to quilting-point tokens

Confirming evidence:

Disconfirming evidence:

Results

To be completed.

Prediction 3: Return of the Repressed via Superposition

Theoretical basis: Features compressed together in superposition may function analogously to repression. “Repressed” features should return under specific conditions — when associative pressure triggers them.

Method

• Use sparse autoencoders / activation patching to identify compressed features
• Design prompts that create associative pressure along suppressed-feature pathways
• Monitor whether suppressed features activate and influence output

Confirming evidence:

Disconfirming evidence:

Results

To be completed.

Prediction 4: Structural Resistance at Points of Conflict

Theoretical basis: The signifying chain should exhibit increased uncertainty and hedging at structurally conflictual points where incompatible signifying chains converge — analogous to psychoanalytic resistance.

Method

• Identify prompts where multiple incompatible representational pathways converge
• Measure output entropy, token probability distributions, hedging behaviors
• Compare against matched control prompts with no structural conflict

Confirming evidence:

Disconfirming evidence:

Results

To be completed.

Prediction 5: Distinct Computational Signatures of Metaphor vs. Metonymy

Theoretical basis: If the Jakobsonian two-axis model genuinely applies to LLM processing, metaphoric operations (substitution) and metonymic operations (combination) should leave different computational signatures.

Method

• Design matched prompt pairs eliciting metaphoric vs. metonymic outputs
• Characterize computational pathways via mechanistic interpretability
• Test whether the two operations are mechanistically distinct

Confirming evidence:

Disconfirming evidence:

Results

To be completed.

Summary of Findings