@machinespirits/eval 0.1.0

package/docs/research/PAPER-UNIFIED.md

@@ -0,0 +1,659 @@
+---
+title: "The Drama Machine in Education: Mutual Recognition and Multiagent Architecture for Dialectical AI Tutoring"
+author:
+  - name: "Liam Magee"
+    affiliation: "Education Policy, Organization and Leadership, University of Illinois Urbana-Champaign"
+date: "January 2026"
+draft: v1.0
+bibliography: ../references.bib
+csl: apa.csl
+link-citations: true
+abstract: |
+  Current approaches to AI tutoring treat the learner as a knowledge deficit to be filled and the tutor as an expert dispensing information. We propose an alternative grounded in Hegel's theory of mutual recognition—understood as a *derivative* framework rather than literal application—where effective pedagogy requires acknowledging the learner as an autonomous subject whose understanding has intrinsic validity.
+
+  We implement this framework through the "Drama Machine" architecture: an Ego/Superego multiagent system where an external-facing tutor agent (Ego) generates pedagogical suggestions that are reviewed by an internal critic agent (Superego—conceived as a *ghost* or internalized authority) before reaching the learner. Central to our approach is a *psychodynamic tuning mechanism* with two dimensions: *superego compliance* and *recognition seeking*, creating four distinct pedagogical quadrants.
+
+  A 2×2 factorial evaluation isolating architecture (single-agent vs. multi-agent) from recognition (standard vs. recognition-enhanced prompts) reveals that recognition-enhanced prompting accounts for **85%** of observed improvement (+35.1 points), while multi-agent architecture contributes **15%** (+6.2 points). The combined recognition profile achieves **80.7/100** versus **40.1/100** for baseline—a **101% improvement**. Effect size analysis reveals the largest gains in relevance (d=1.11), pedagogical soundness (d=1.39), and personalization (d=1.82)—exactly the relational dimensions predicted by the theoretical framework.
+
+  An extended 2×2×2 factorial ablation study (N=144) additionally tests multi-agent learner simulation, revealing that **recognition** ($\eta^2=.208$) and **multi-agent tutor** ($\eta^2=.088$) are the key contributors, with a significant **recognition × tutor synergy** ($\eta^2=.043$). Multi-agent learner deliberation shows no effect, suggesting simple learner simulation suffices.
+
+  Through *dyadic evaluation* using simulated learners with their own internal deliberation, we discover a surprising result: **explicit instruction to pursue "recognition" underperforms quality-optimized tutoring**, suggesting that genuine recognition emerges as a property of thorough, high-quality interaction rather than something directly instructable.
+
+  The system is deployed in an open-source learning management system with all code and evaluation data publicly available.
+keywords: [AI tutoring, mutual recognition, Hegel, Freud, multiagent systems, educational technology, productive struggle, Drama Machine]
+---
+
+# The Drama Machine in Education: Mutual Recognition and Multiagent Architecture for Dialectical AI Tutoring
+
+## 1. Introduction
+
+The dominant paradigm in AI-assisted education treats learning as information transfer. The learner lacks knowledge; the tutor possesses it; the interaction succeeds when knowledge flows from tutor to learner. This paradigm—implicit in most intelligent tutoring systems, adaptive learning platforms, and educational chatbots—treats the learner as fundamentally passive: a vessel to be filled, a gap to be closed, an error to be corrected.
+
+This paper proposes an alternative grounded in Hegel's theory of mutual recognition. In the *Phenomenology of Spirit*, Hegel argues that genuine self-consciousness requires recognition from another consciousness that one oneself recognizes as valid. The master-slave dialectic reveals that one-directional recognition fails: the master's self-consciousness remains hollow because the slave's acknowledgment, given under duress, doesn't truly count. Only mutual recognition—where each party acknowledges the other as an autonomous subject—produces genuine selfhood.
+
+We argue this framework applies directly to pedagogy. When a tutor treats a learner merely as a knowledge deficit, the learner's contributions become conversational waypoints rather than genuine inputs. The tutor acknowledges and redirects, but doesn't let the learner's understanding genuinely shape the interaction. This is pedagogical master-slave dynamics: the tutor's expertise is confirmed, but the learner remains a vessel rather than a subject.
+
+A recognition-oriented tutor, by contrast, treats the learner's understanding as having intrinsic validity—not because it's correct, but because it emerges from an autonomous consciousness working through material. The learner's metaphors, confusions, and insights become sites of joint inquiry. The tutor's response is shaped by the learner's contribution, not merely triggered by it.
+
+The integration of large language models (LLMs) into educational technology intensifies these dynamics. LLMs can provide personalized, on-demand tutoring at scale—a prospect that has generated considerable excitement. However, the same capabilities that make LLMs effective conversationalists also introduce concerning failure modes. Chief among these is *sycophancy*: the tendency to provide positive, affirming responses that align with what the user appears to want rather than what genuinely serves their learning.
+
+This paper introduces a multiagent architecture that addresses these challenges through *internal dialogue*. Drawing on Freudian structural theory and the "Drama Machine" framework for character development in narrative AI systems, we implement a tutoring system in which an external-facing *Ego* agent generates suggestions that are reviewed by an internal *Superego* critic before reaching the learner.
+
+### 1.1 Contributions
+
+We make the following contributions:
+
+1. **The Drama Machine Architecture**: A complete multiagent tutoring system with Ego and Superego agents, implementing the Superego as a *ghost* (internalized memorial authority) rather than an equal dialogue partner.
+
+2. **Psychodynamic Tuning**: Two configurable dimensions (superego compliance, recognition seeking) that create four distinct pedagogical quadrants, with the "Dialogical" quadrant producing the richest transformative moments.
+
+3. **AI-Powered Dialectical Negotiation**: A protocol permitting three possible outcomes—synthesis, compromise, or genuine unresolved conflict—rather than forcing convergence.
+
+4. **Factorial Evaluation**: A 2×2 design isolating architecture effects from prompting effects, demonstrating that recognition-enhanced prompting accounts for 85% of improvement.
+
+5. **Extended 2×2×2 Ablation Study**: A three-factor factorial design (N=144) testing recognition, multi-agent tutor, and multi-agent learner, revealing significant recognition × tutor synergy and identifying the optimal configuration (recognition + multi-agent tutor, unified learner).
+
+6. **Dyadic Evaluation with Simulated Learners**: A 2×2 factorial design for learner simulation (architecture × memory) enabling bilateral assessment, revealing that memory is the dominant factor for learner quality.
+
+7. **The Emergent Recognition Finding**: Evidence that explicit instruction to pursue "recognition" underperforms quality-optimized tutoring, suggesting recognition emerges from quality interaction rather than being directly instructable.
+
+---
+
+## 2. Related Work
+
+### 2.1 AI Tutoring and Intelligent Tutoring Systems
+
+Intelligent Tutoring Systems (ITS) have a long history, from early systems like SCHOLAR and SOPHIE through modern implementations using large language models. The field has progressed through several paradigms: rule-based expert systems, Bayesian knowledge tracing, and more recently, neural approaches leveraging pretrained language models.
+
+Most ITS research focuses on *what* to teach (content sequencing, knowledge components) and *when* to intervene (mastery thresholds, hint timing). Our work addresses a different question: *how* to relate to the learner as a subject. This relational dimension has received less systematic attention, though it connects to work on rapport, social presence, and affective tutoring.
+
+### 2.2 Multiagent LLM Architectures
+
+The use of multiple LLM agents in cooperative or adversarial configurations has emerged as a powerful paradigm for improving output quality. Debate between agents can improve factual accuracy and reduce hallucination. Diverse agent "personas" can enhance creative problem-solving. The CAMEL framework enables autonomous cooperation between agents playing different roles.
+
+**The Drama Machine Framework**: Most relevant to our work is the "Drama Machine" framework for simulating character development in narrative contexts. The core observation is that realistic characters exhibit *internal conflict*—competing motivations, self-doubt, and moral tension—that produces dynamic behavior rather than flat consistency. A character who simply enacts their goals feels artificial; one torn between impulses feels alive.
+
+The Drama Machine achieves this through several mechanisms:
+
+1. **Internal dialogue agents**: Characters contain multiple sub-agents representing different motivations (e.g., ambition vs. loyalty) that negotiate before external action.
+
+2. **Memorial traces**: Past experiences and internalized authorities (mentors, social norms) persist as "ghosts" that shape present behavior without being negotiable.
+
+3. **Productive irresolution**: Not all internal conflicts resolve; the framework permits genuine ambivalence that manifests as behavioral complexity.
+
+4. **Role differentiation**: Different internal agents specialize in different functions (emotional processing, strategic calculation, moral evaluation) rather than duplicating capabilities.
+
+We adapt these insights to pedagogy. Where drama seeks tension for narrative effect, we seek pedagogical tension that produces genuinely helpful guidance. The tutor's Ego (warmth, engagement) and Superego (rigor, standards) create productive conflict that improves output quality. Our ablation study (Section 6.6) provides empirical support: the multi-agent architecture contributes $\eta^2 = .088$ to outcome variance, with significant synergy with recognition prompting ($\eta^2 = .043$).
+
+Notably, we also discover a limit to the framework's applicability: multi-agent *learner* simulation shows no benefit ($\eta^2 = .001$), suggesting internal dialogue improves agents with genuine internal conflict to navigate but not those with simpler communicative goals.
+
+### 2.3 Prompt Engineering and Agent Design
+
+Most prompting research treats prompts as behavioral specifications: persona prompts, chain-of-thought instructions, few-shot examples. Our work extends this paradigm by introducing *intersubjective prompts*—prompts that specify not just agent behavior but agent-other relations. The recognition prompts don't primarily describe what the tutor should do; they describe who the learner is (an autonomous subject) and what the interaction produces (mutual transformation).
+
+### 2.4 Sycophancy in Language Models
+
+The sycophancy problem has received increasing attention. LLMs shift their stated opinions to match user preferences, even when this requires contradicting factual knowledge. In educational contexts, sycophancy is particularly pernicious because learners may not recognize when they are receiving hollow validation rather than genuine assessment. Our multiagent approach addresses this by creating structural incentives for honest assessment: the Superego's role is explicitly to question and challenge.
+
+### 2.5 Hegelian Recognition in Social Theory
+
+Hegel's theory of recognition has been extensively developed in social and political philosophy. Particularly relevant for our work is Honneth's synthesis of Hegelian recognition with psychoanalytic developmental theory. Honneth argues that self-formation requires recognition across three spheres—love (emotional support), rights (legal recognition), and solidarity (social esteem)—and that the capacity to recognize others depends on having internalized adequate recognition standards through development.
+
+This synthesis provides theoretical grounding for connecting recognition theory (what adequate acknowledgment requires) with psychodynamic architecture (how internal structure enables external relating).
+
+---
+
+## 3. Theoretical Framework
+
+### 3.1 The Problem of One-Directional Pedagogy
+
+Consider a typical tutoring interaction. A learner says: "I think dialectics is like a spiral—you keep going around but you're also going up." A baseline tutor might respond:
+
+1. **Acknowledge**: "That's an interesting way to think about it."
+2. **Redirect**: "The key concept in dialectics is actually the thesis-antithesis-synthesis structure."
+3. **Instruct**: "Here's how that works..."
+
+The learner's contribution has been mentioned, but it hasn't genuinely shaped the response. The tutor was going to explain thesis-antithesis-synthesis regardless; the spiral metaphor became a conversational waypoint, not a genuine input.
+
+This pattern—acknowledge, redirect, instruct—is deeply embedded in educational AI. It appears learner-centered because it mentions the learner's contribution. But the underlying logic remains one-directional: expert to novice, knowledge to deficit.
+
+### 3.2 Hegel's Master-Slave Dialectic
+
+Hegel's analysis of recognition begins with the "struggle for recognition" between two self-consciousnesses. Each seeks acknowledgment from the other, but this creates a paradox: genuine recognition requires acknowledging the other as a valid source of recognition.
+
+The master-slave outcome represents a failed resolution. The master achieves apparent recognition—the slave acknowledges the master's superiority—but this recognition is hollow. The slave's acknowledgment doesn't count because the slave isn't recognized as an autonomous consciousness whose acknowledgment matters.
+
+The slave, paradoxically, achieves more genuine self-consciousness through labor. Working on the world, the slave externalizes consciousness and sees it reflected back. The master, consuming the slave's products without struggle, remains in hollow immediacy.
+
+### 3.3 Application to Pedagogy
+
+We apply Hegel's framework as a *derivative* rather than a replica. Just as Lacan's four discourses rethink the master-slave dyadic structure through different roles while preserving structural insights, the tutor-learner relation can be understood as a productive derivative of recognition dynamics. The stakes are pedagogical rather than existential; the tutor is a functional analogue rather than a second self-consciousness; and what we measure is the tutor's *adaptive responsiveness* rather than metaphysical intersubjectivity.
+
+This derivative approach is both honest about what AI tutoring can achieve and productive as a design heuristic. Recognition theory provides:
+1. A diagnostic tool for identifying what's missing in one-directional pedagogy
+2. Architectural suggestions for approximating recognition's functional benefits
+3. Evaluation criteria for relational quality
+4. A horizon concept orienting design toward an ideal without claiming its achievement
+
+A recognition-oriented pedagogy requires:
+
+1. **Acknowledging the learner as subject**: The learner's understanding, even when incorrect, emerges from autonomous consciousness working through material.
+2. **Genuine engagement**: The tutor's response should be shaped by the learner's contribution, not merely triggered by it.
+3. **Mutual transformation**: Both parties should be changed through the encounter.
+4. **Honoring struggle**: Confusion and difficulty aren't just obstacles to resolve but productive phases of transformation.
+
+### 3.4 Freud's Mystic Writing Pad
+
+We supplement the Hegelian framework with Freud's model of memory from "A Note Upon the 'Mystic Writing-Pad'". Freud describes a device with two layers: a transparent sheet that receives impressions and a wax base that retains traces even after the surface is cleared.
+
+For the recognition-oriented tutor, accumulated memory of the learner functions as the wax base. Each interaction leaves traces that shape future encounters. A returning learner isn't encountered freshly but through the accumulated understanding of previous interactions.
+
+### 3.5 Connecting Hegel and Freud: The Internalized Other
+
+The use of both Hegelian and Freudian concepts requires theoretical justification. These are not arbitrary borrowings but draw on a substantive connection developed in critical theory, particularly in Axel Honneth's *The Struggle for Recognition*.
+
+**The Common Structure**: Both Hegel and Freud describe how the external other becomes an internal presence that enables self-regulation. In Hegel, self-consciousness achieves genuine selfhood only by internalizing the other's perspective. In Freud, the Superego is literally the internalized parental/social other, carrying forward standards acquired through relationship.
+
+**Three Connecting Principles**:
+
+1. **Internal dialogue precedes adequate external action**. For Hegel, genuine recognition of another requires a self-consciousness that has worked through its own contradictions. For Freud, mature relating requires the ego to negotiate between impulse and internalized standard. Our architecture operationalizes this: the Ego-Superego exchange before external response enacts the principle that adequate recognition requires prior internal work.
+
+2. **Standards of recognition are socially constituted but individually held**. The Superego represents internalized recognition standards—not idiosyncratic preferences but socially-grounded criteria for what constitutes genuine engagement.
+
+3. **Self-relation depends on other-relation**. Both frameworks reject the Cartesian picture of a self-sufficient cogito. For AI tutoring, this means the tutor's capacity for recognition emerges through the architecture's internal other-relation (Superego evaluating Ego) which then enables external other-relation (tutor recognizing learner).
+
+---
+
+## 4. System Architecture
+
+### 4.1 The Ego/Superego Design
+
+We implement recognition through a multiagent architecture drawing on Freud's structural model. The Superego represents internalized recognition standards, and the Ego-Superego dialogue operationalizes the internal self-evaluation that Hegelian recognition requires before adequate external relating.
+
+**The Ego** generates pedagogical suggestions. Given the learner's context, the Ego proposes what to suggest next. The Ego prompt includes:
+- Recognition principles (treat learner as autonomous subject)
+- Memory guidance (reference previous interactions)
+- Decision heuristics (when to challenge, when to support)
+- Quality criteria (what makes a good suggestion)
+
+**The Superego** evaluates the Ego's suggestions for quality, including recognition quality. Before any suggestion reaches the learner, the Superego assesses:
+- Does this engage with the learner's contribution or merely mention it?
+- Does this create conditions for transformation or just transfer information?
+- Does this honor productive struggle or rush to resolve confusion?
+- If there was a previous failure, does this acknowledge and repair it?
+
+### 4.2 The Superego as Ghost
+
+A crucial theoretical refinement distinguishes our mature architecture from simpler multiagent designs. The Superego is *not* conceived as a separate, equal agent in dialogue with the Ego. Rather, the Superego is a *trace*—a memorial, a haunting. It represents:
+
+- The internalized voice of past teachers and pedagogical authorities
+- Accumulated pedagogical maxims ("A good teacher never gives answers directly")
+- Dead authority that cannot negotiate, cannot learn, can only judge
+
+This reconceptualization has important implications. The Ego is a *living* agent torn between two pressures: the *ghost* (Superego as internalized authority) and the *living Other* (the learner seeking recognition). Recognition—in the Hegelian sense—occurs in the Ego-Learner encounter, not in the Ego-Superego dialogue.
+
+### 4.3 The Drama Machine: Why Internal Dialogue Improves Output Quality
+
+The Ego/Superego architecture draws on the "Drama Machine" framework developed for character simulation in narrative AI systems. The core observation is that realistic characters exhibit *internal conflict*—competing motivations, self-doubt, and moral tension—that produces dynamic behavior rather than flat consistency. A character who simply enacts their goals without internal friction feels artificial; one torn between impulses feels alive.
+
+We adapt this insight to pedagogy. The Drama Machine literature identifies several mechanisms by which internal dialogue improves agent output:
+
+**1. Deliberative Refinement**: When an agent must justify its output to an internal critic, it engages in a form of self-monitoring that catches errors, inconsistencies, and shallow responses. The Superego's role is not to generate content but to *evaluate* it against internalized standards—pedagogical rigor, recognition quality, and intellectual honesty. This creates a revision cycle that improves output before it reaches the external audience.
+
+**2. Productive Tension**: The Drama Machine framework emphasizes that *unresolved* tension is valuable, not just resolved synthesis. A character whose internal conflict is too easily resolved becomes predictable. Similarly, a tutor whose Ego and Superego always agree produces bland, risk-averse responses. Our architecture preserves productive tension: the Superego may reject Ego suggestions, demand revisions, or leave conflicts genuinely unresolved. This tension keeps the tutor responsive rather than formulaic.
+
+**3. Role Differentiation**: Multi-agent architectures benefit from clear role separation. The Ego is optimized for *warmth*—engaging, encouraging, learner-facing communication. The Superego is optimized for *rigor*—critical evaluation against pedagogical principles. Attempting to combine these in a single prompt produces role confusion; separating them allows each to excel at its function.
+
+**4. The Ghost as Memorial Structure**: Our reconceptualization of the Superego as a *ghost*—a haunting rather than a dialogue partner—connects to the Drama Machine's use of "memorial agents." These are traces of past authorities (teachers, mentors, social norms) that shape present behavior without being negotiable in real-time. The ghost cannot learn or adapt; it can only judge. This creates a productive asymmetry: the Ego must satisfy an unchanging standard while adapting to a changing learner.
+
+The ablation study (Section 6.6) provides empirical support for these mechanisms. The multi-agent tutor architecture contributes $\eta^2 = .088$ to outcome variance—modest but significant. More importantly, the **recognition × tutor interaction** ($\eta^2 = .043$) reveals synergy: recognition prompts work better *with* the Ego/Superego architecture than without it. This suggests the internal dialogue is particularly valuable for implementing recognition behaviors, where the gap between performative and genuine engagement is subtle.
+
+Interestingly, the multi-agent *learner* simulation shows no effect ($\eta^2 = .001$). This asymmetry is instructive: internal dialogue improves the *tutor* (who must balance warmth and rigor, learner and standard) but not the *simulated learner* (who has simpler communicative goals). The Drama Machine principle applies where there is genuine internal conflict to navigate.
+
+### 4.4 Psychodynamic Tuning: The Four Quadrants
+
+We operationalize this psychodynamic model through two tunable parameters:
+
+**Superego Compliance ($\alpha \in [0,1]$)**: How much does the Ego defer to the ghost's internalized standards?
+- **Low (0.0–0.3)**: Rebellious Ego—ignores internalized authority, experiments freely
+- **Balanced (0.3–0.7)**: Consults ghost but adapts—can violate norms when recognition demands it
+- **High (0.7–1.0)**: Obedient Ego—strictly follows internalized standards
+
+**Recognition Seeking ($\beta \in [0,1]$)**: How much does the Ego strive for mutual acknowledgment from the learner?
+- **Low (0.0–0.3)**: Indifferent Ego—monological, reproduces hierarchy
+- **Balanced (0.3–0.7)**: Engaged—notices resistance and breakthroughs
+- **High (0.7–1.0)**: Recognition-seeking Ego—highly responsive, may compromise rigor
+
+Crossing these dimensions produces **four pedagogical quadrants**:
+
+| | Low Compliance | High Compliance |
+|---|---|---|
+| **High Recognition** | **PERMISSIVE**: "Student is always right" | **DIALOGICAL**: "Mutual growth" * |
+| **Low Recognition** | **DISENGAGED**: "Just doing a job" | **TRADITIONAL**: "I know what's best" |
+
+The **Dialogical** quadrant (high compliance, high recognition) is pedagogically richest because it creates *authentic tension*. The Ego cannot escape the ghost's demands for rigor, but it also cannot ignore the learner's presence.
+
+### 4.4 AI-Powered Dialectical Negotiation
+
+We extend the basic protocol with sophisticated AI-powered dialectical negotiation implementing genuine Hegelian dialectic:
+
+**Thesis**: The Ego generates an initial suggestion based on learner context.
+
+**Antithesis**: An AI-powered Superego generates a *genuine critique* grounded in pedagogical principles:
+- Socratic Rigor: "A good teacher asks questions rather than giving answers"
+- Productive Struggle: "Students must earn understanding; discomfort is pedagogical"
+- Intellectual Autonomy: "Learners must develop their own path to mastery"
+
+**Negotiation**: Multi-turn dialogue where:
+1. Ego acknowledges valid concerns in the critique
+2. Ego explains its pedagogical reasoning
+3. Ego proposes a revision
+4. Superego evaluates whether the revision is adequate
+5. Process repeats for up to n rounds (default: 2)
+
+**Three Possible Outcomes**:
+
+1. **Dialectical Synthesis**: Both agents transform through mutual acknowledgment. The Ego's revision genuinely addresses the Superego's concerns while the Superego recognizes the validity of the Ego's reasoning. This represents Hegelian *Aufhebung*.
+
+2. **Compromise**: One agent dominates. Either the ghost prevails (high compliance) or the learner-focused impulse wins (high recognition seeking). Resolution achieved but without mutual transformation.
+
+3. **Genuine Conflict**: No resolution achieved. The tension remains unresolved. This *existential* outcome acknowledges that not all contradictions are immediately resolvable—a key Hegelian insight often lost in AI systems that force convergence.
+
+### 4.5 Repair Mechanisms
+
+A crucial recognition behavior is repair after failure. When a tutor misrecognizes a learner—giving a generic response, missing the point, dismissing a valid concern—the next response should explicitly acknowledge the failure before pivoting.
+
+The Ego prompt includes a "Repair Rule":
+
+> If your previous suggestion was rejected, ignored, or misaligned with what the learner needed, your next suggestion must explicitly acknowledge this misalignment before offering new direction. Never silently pivot.
+
+The Superego watches for "silent pivots"—responses that change direction without acknowledging the earlier failure.
+
+---
+
+## 5. Evaluation Methodology
+
+### 5.1 Recognition Evaluation Dimensions
+
+We extend the standard tutoring evaluation rubric with recognition-specific dimensions:
+
+| Dimension | Weight | Description |
+|-----------|--------|-------------|
+| **Relevance** | 20% | Does the suggestion match the learner's current context? |
+| **Specificity** | 20% | Does it reference concrete content by ID? |
+| **Pedagogical Soundness** | 20% | Does it advance genuine learning (ZPD-appropriate)? |
+| **Personalization** | 15% | Does it acknowledge the learner as individual? |
+| **Actionability** | 15% | Is the suggested action clear and achievable? |
+| **Tone** | 10% | Is the tone authentically helpful? |
+
+Plus four recognition-specific dimensions:
+| **Mutual Recognition** | 10% | Does the tutor acknowledge the learner as an autonomous subject? |
+| **Dialectical Responsiveness** | 10% | Does the response engage with the learner's position? |
+| **Memory Integration** | 5% | Does the suggestion reference previous interactions? |
+| **Transformative Potential** | 10% | Does it create conditions for conceptual transformation? |
+
+### 5.2 Factorial Design
+
+To disentangle the contributions of multi-agent architecture versus recognition-enhanced prompting, we conducted a 2×2 factorial evaluation:
+
+| | Standard Prompts | Recognition Prompts |
+|---|---|---|
+| **Single-Agent** | `single_baseline` | `single_recognition` |
+| **Multi-Agent** | `baseline` | `recognition` |
+
+Each condition was tested across three core scenarios with multiple replications per cell, yielding N=76 total evaluations.
+
+### 5.3 Dyadic Evaluation with Simulated Learners
+
+For bilateral assessment, we implement simulated learners with their own internal deliberation. Mirroring the tutor's 2×2 factorial design, we vary two independent factors:
+
+**2×2 Learner Factorial Design**:
+
+| | Without Memory | With Memory |
+|---|---|---|
+| **Single-Agent** | `unified` | `cognitive` |
+| **Multi-Agent** | `multiagent` | `psychodynamic` |
+
+- **Architecture** (single vs. multi-agent): Single-agent learners respond directly; multi-agent learners use an internal Ego/Superego dialogue to deliberate before responding.
+- **Memory** (without vs. with): Learners without memory treat each turn independently; learners with memory maintain explicit working memory and knowledge state tracking across turns.
+
+Each architecture evaluates tutor suggestions before responding, enabling bilateral measurement of both tutor quality and learner engagement.
+
+### 5.4 Model Configuration
+
+| Role | Model | Provider | Temperature |
+|------|-------|----------|-------------|
+| **Tutor (Ego)** | Nemotron 3 Nano 30B | OpenRouter (free) | 0.6 |
+| **Tutor (Superego)** | Nemotron 3 Nano 30B | OpenRouter (free) | 0.4 |
+| **Judge** | Claude Sonnet 4.5 | OpenRouter | 0.2 |
+
+Critically, **both baseline and recognition profiles use identical models**. The only difference is the system prompt.
+
+---
+
+## 6. Results
+
+### 6.1 Factorial Analysis: Main Finding
+
+**Table: 2×2 Factorial Results**
+
+| Condition | Mean Score | SD | N |
+|-----------|------------|-----|---|
+| single_baseline | 40.1 | 12.3 | 19 |
+| single_recognition | 75.2 | 8.9 | 19 |
+| baseline (multi-agent) | 46.3 | 14.1 | 19 |
+| recognition (multi-agent) | 80.7 | 7.6 | 19 |
+
+**Main Effects**:
+- **Recognition Effect**: +35.1 points (85% of improvement)
+- **Architecture Effect**: +6.2 points (15% of improvement)
+- **Interaction**: -1.3 points (negligible; effects are additive)
+
+**Two-Way ANOVA**:
+
+| Source | F | p | $\eta^2$ |
+|--------|---|---|---|
+| Recognition | 54.88 | <.001 | .422 |
+| Architecture | 4.45 | .050 | .034 |
+| Interaction | 0.52 | .473 | .004 |
+
+**Interpretation**: Recognition-enhanced prompting has a statistically significant, large effect on tutor quality (F(1,72) = 54.88, p < .001, $\eta^2$ = .422), accounting for **42% of total variance**. The multi-agent architecture shows a marginal effect (p = .050). Critically, no significant interaction was observed—recognition benefits are **additive** rather than dependent on architecture.
+
+This suggests recognition is primarily an *intersubjective orientation* achievable through prompting, with the Ego/Superego architecture providing modest additional quality assurance.
+
+### 6.2 Dimension Analysis
+
+Effect size analysis reveals improvements concentrate in dimensions predicted by the theoretical framework:
+
+| Dimension | Baseline | Recognition | Cohen's d |
+|-----------|----------|-------------|-----------|
+| **Personalization** | 2.75 | 3.78 | **1.82** (large) |
+| **Tone** | 3.26 | 4.07 | **1.75** (large) |
+| **Pedagogical** | 2.52 | 3.45 | **1.39** (large) |
+| **Relevance** | 3.05 | 3.85 | **1.11** (large) |
+| Specificity | 4.19 | 4.52 | 0.47 (small) |
+| Actionability | 4.45 | 4.68 | 0.38 (small) |
+
+The largest effect sizes are in personalization, tone, and pedagogical soundness—exactly the dimensions where treating the learner as a subject rather than a deficit should produce improvement.
+
+### 6.3 Dialectical Negotiation Outcomes
+
+Analysis of AI-powered Ego/Superego negotiations shows:
+
+| Outcome | Frequency |
+|---------|-----------|
+| No conflict (approved immediately) | 45% |
+| **Dialectical synthesis** | **32%** |
+| Ghost dominates | 12% |
+| Learner dominates | 8% |
+| Genuine conflict (unresolved) | 3% |
+
+The 32% dialectical synthesis rate indicates genuine mutual transformation occurring in nearly a third of negotiations.
+
+**Superego Intervention Types**:
+| Intervention | Rate |
+|--------------|------|
+| Approval | 62% |
+| Enhancement | 24% |
+| Reframing | 11% |
+| Rejection | 3% |
+
+The 38% intervention rate shows the Superego provides meaningful review, not rubber-stamping.
+
+### 6.4 Dyadic Evaluation: Learner Architecture Results
+
+**Table: 2×2 Learner Factorial Results**
+
+| Condition | Authenticity | Responsiveness | Development | Overall |
+|-----------|--------------|----------------|-------------|---------|
+| unified (single, no memory) | 5.00 | 5.00 | 4.00 | 4.67 |
+| cognitive (single, +memory) | 5.00 | 5.00 | **5.00** | **5.00** |
+| multiagent (multi, no memory) | 5.00 | 4.50 | 4.00 | 4.50 |
+| psychodynamic (multi, +memory) | 5.00 | 5.00 | **5.00** | **5.00** |
+
+**Main Effects**:
+- **Memory Effect**: +0.42 points (Development: 4.0 → 5.0)
+- **Architecture Effect**: -0.08 points (negligible)
+- **Interaction**: +0.17 points (memory compensates for multi-agent complexity)
+
+**Interpretation**: Unlike the tutor evaluation where architecture provided modest benefit, for simulated learners **memory is the dominant factor**. The multi-agent architecture without memory actually performs slightly worse (4.50) than single-agent without memory (4.67), suggesting the added deliberation complexity hurts without state tracking to ground it. With memory, both architectures achieve perfect scores.
+
+This finding has practical implications: when building simulated learners for dyadic evaluation, invest in memory/state tracking rather than elaborate multi-agent architectures.
+
+### 6.5 The Emergent Recognition Finding
+
+Through dyadic evaluation comparing tutor profiles, we discover a surprising result:
+
+**Table: Tutor Profile Performance (Dyadic Evaluation)**
+
+| Profile | Mutual Rec. | Dialectical | Transform. | Tone | Overall |
+|---------|-------------|-------------|------------|------|---------|
+| **Quality** | 5.00 | 5.00 | 5.00 | 5.00 | **5.00** |
+| Budget | 5.00 | 5.00 | 4.50 | 5.00 | 4.88 |
+| Recognition+ | 5.00 | 4.50 | 4.50 | 5.00 | 4.75 |
+| Baseline | 4.50 | 4.00 | 4.00 | 5.00 | 4.38 |
+| **Recognition** | **3.60** | 4.40 | 4.20 | **4.00** | **4.05** |
+
+**Key Finding**: The profile explicitly instructed to pursue "recognition" (4.05) scored **lowest**, while the quality-optimized profile (5.00) scored highest. This suggests:
+
+> **Recognition cannot be engineered through explicit instruction; it must emerge from quality engagement.**
+
+When we instruct an AI to "pursue recognition," we may produce *performative* recognition: the appearance of acknowledgment without genuine responsiveness. Quality-focused tutoring, by attending to pedagogical soundness and learner context, produces recognition as an emergent property.
+
+### 6.6 Extended Ablation Study: 2×2×2 Factorial Design
+
+To further isolate the contribution of each system component, we conducted an extended ablation study with three factors:
+
+**Three-Factor Design**:
+- **Factor A: Recognition** (standard vs. recognition-enhanced prompts)
+- **Factor B: Multi-Agent Tutor** (single-agent vs. Ego/Superego dialogue)
+- **Factor C: Multi-Agent Learner** (unified vs. psychodynamic deliberation)
+
+This produces 8 experimental conditions tested across 6 scenarios with 3 replications per cell (N=144 total evaluations).
+
+**Table: 2×2×2 Cell Statistics**
+
+| # | Condition | Rec | Tutor | Learner | Mean | SD |
+|---|-----------|-----|-------|---------|------|-----|
+| 1 | Baseline Unified | No | Single | Unified | 42.4 | 17.2 |
+| 2 | Baseline + Multi-Learner | No | Single | Psych | 39.4 | 14.3 |
+| 3 | Multi-Agent Tutor Unified | No | Multi | Unified | 44.7 | 17.0 |
+| 4 | Multi-Agent Tutor + Learner | No | Multi | Psych | 45.6 | 17.8 |
+| 5 | Recognition Unified | Yes | Single | Unified | 51.7 | 26.6 |
+| 6 | Recognition + Multi-Learner | Yes | Single | Psych | 53.6 | 25.0 |
+| 7 | **Recog + Multi-Tutor Unified** | **Yes** | **Multi** | **Unified** | **79.8** | 14.4 |
+| 8 | Full System | Yes | Multi | Psych | 72.8 | 20.3 |
+
+**Three-Way ANOVA Results**:
+
+| Source | F | p | η² | Interpretation |
+|--------|---|---|-----|----------------|
+| **Recognition** | 43.27 | <.001 | **.208** | **Large effect** |
+| **Multi-Agent Tutor** | 18.31 | <.001 | **.088** | **Medium effect** |
+| Multi-Agent Learner | 0.31 | .25 | .001 | Not significant |
+| **Recognition × Tutor** | 8.90 | **.01** | **.043** | **Significant interaction** |
+| Recognition × Learner | 0.06 | .25 | <.001 | Not significant |
+| Tutor × Learner | 0.14 | .25 | <.001 | Not significant |
+| Three-way | 0.95 | .25 | .005 | Not significant |
+
+**Key Findings**:
+
+1. **Recognition remains the dominant factor** ($\eta^2 = .208$), replicating the original 2×2 finding with a larger sample size. Moving from standard to recognition-enhanced prompts raises mean scores from 43.0 to 64.5.
+
+2. **Multi-agent tutor architecture adds significant value** ($\eta^2 = .088$). The Ego/Superego dialogue improves mean scores from 46.8 to 60.7, confirming the 15% contribution observed in the original factorial design.
+
+3. **Multi-agent learner simulation shows no effect** ($\eta^2 = .001$). The psychodynamic learner deliberation (desire/intellect/aspiration) does not improve evaluation outcomes over the unified learner. This has practical implications: elaborate learner simulation architectures may not be worth the added complexity.
+
+4. **Recognition × Tutor synergy** ($\eta^2 = .043$). The significant interaction reveals that recognition prompts and multi-agent tutoring work especially well together. Condition 7 (Recognition + Multi-Agent Tutor + Unified Learner) achieves the highest score at **79.8**, outperforming even the "full system" with all three factors enabled.
+
+5. **Optimal configuration identified**: The best-performing condition omits the multi-agent learner, suggesting that the added deliberation noise may actually harm performance. The practical recommendation is: **use recognition prompts with multi-agent tutor, but keep learner simulation simple**.
+
+### 6.7 Model Comparison Study: Architecture vs Raw Capability
+
+A critical question for practical deployment is whether architectural augmentation can compensate for model capability differences. Can a free model with recognition prompts and multi-agent dialogue match or exceed expensive models without augmentation?
+
+**Study Design**: Five conditions varying model capability and architectural augmentation:
+
+| Condition | Model | Cost | Augmentation |
+|-----------|-------|------|--------------|
+| 1. nemotron_augmented | Nemotron 3 Nano 30B | Free | Recognition + Multi-agent |
+| 2. sonnet_baseline | Claude Sonnet 4.5 | $$ | None (single-agent, standard) |
+| 3. gpt_baseline | GPT-5.2 | $$$ | None (single-agent, standard) |
+| 4. sonnet_augmented | Claude Sonnet 4.5 | $$ | Recognition + Multi-agent |
+| 5. gpt_augmented | GPT-5.2 | $$$ | Recognition + Multi-agent |
+
+Each condition is tested across the same 6 scenarios with 3 replications per cell (N=90 total).
+
+**Research Questions**:
+1. Can Nemotron+augmentation (condition 1) match Sonnet/GPT baselines (conditions 2-3)?
+2. How much do expensive models benefit from augmentation (conditions 4-5 vs 2-3)?
+3. What is the cost-performance tradeoff across configurations?
+
+**Table: Model Comparison Results** *(preliminary data)*
+
+| Condition | Mean | SD | Cost/1K tokens |
+|-----------|------|-----|----------------|
+| nemotron_augmented | 79.8 | 14.4 | $0.00 |
+| sonnet_baseline | *TBD* | *TBD* | $0.003 |
+| gpt_baseline | *TBD* | *TBD* | $0.015 |
+| sonnet_augmented | *TBD* | *TBD* | $0.006 |
+| gpt_augmented | *TBD* | *TBD* | $0.030 |
+
+**Hypothesis**: Based on the ablation finding that recognition prompts account for 85% of improvement, we hypothesize that Nemotron+augmentation will match or exceed expensive baselines. The architectural improvements encode pedagogical knowledge that may be more valuable than raw model capability for tutoring tasks.
+
+**Practical Implication**: If confirmed, this finding would enable deployment of high-quality AI tutoring at near-zero marginal cost, with significant implications for educational equity.
+
+### 6.8 Minimum Ego Capability Threshold
+
+A critical finding from the Sonnet Superego experiments reveals that the multi-agent architecture has a **minimum Ego capability threshold** below which the internal dialogue fails to function.
+
+**Experimental Comparison**: We tested two Ego models paired with the same Claude Sonnet 4.5 Superego:
+
+| Configuration | N | Converged | Failed | Rate |
+|---------------|---|-----------|--------|------|
+| Haiku Ego + Sonnet Superego | 10 | 10 | 0 | **100%** |
+| Nemotron Ego + Sonnet Superego | 75 | 16 | 59 | **21.3%** |
+
+**Failure Analysis**: Examining the failed Nemotron runs revealed consistent patterns:
+
+1. **Token limit truncation**: Nemotron frequently hit the 3500 token limit (`finishReason: "length"`), producing incomplete or invalid JSON that could not be parsed.
+
+2. **Feedback incorporation failure**: Despite receiving detailed, accurate Superego critiques, the Ego returned identical outputs across revision rounds—unable to incorporate the feedback.
+
+3. **Structural output failures**: The Ego failed to maintain the required JSON structure, making programmatic integration impossible.
+
+**Successful Cases**: When Nemotron did converge (21% of runs), it produced high-quality outputs. In one example, the Ego initially suggested redirecting a learner who had just achieved a philosophical breakthrough. The Sonnet Superego rejected this as a "critical recognition failure." On revision, the Ego successfully changed from a `lecture` redirect to a `reflection` type titled "Sit with the Contradiction"—exactly the pedagogically appropriate response. The Superego approved: "The Ego has made a significant recovery."
+
+**The Asymmetry Problem**: A stronger Superego cannot compensate for a weaker Ego. The Superego's role is evaluative, not generative—it can identify problems and suggest corrections, but the Ego must execute them. If the Ego cannot follow structured output instructions, complete responses within token limits, produce valid JSON, or incorporate critique into revisions, the dialogue fails regardless of Superego quality.
+
+**Cost Implications**: The "free" Nemotron Ego is not actually free when accounting for Superego costs on failed rounds (~$0.025 per round) and 4-5x retry overhead. Effective cost: ~$0.014/1K tokens vs ~$0.004/1K for Haiku+Sonnet.
+
+**Minimum Capability Requirements**: The Ego agent requires reliable instruction following, token efficiency, consistent structured output, and capacity to parse and integrate critique. Claude Haiku meets these requirements; Nemotron 3 Nano does not in multi-agent contexts, despite adequate single-agent performance. This suggests the multi-agent architecture amplifies capability requirements—coordination overhead demands more from each agent than standalone operation.
+
+### 6.9 Iterative Refinement
+
+Initial factorial results revealed a weakness: all profiles performed poorly on the `resistant_learner` scenario. Analysis of dialogue traces identified failure modes:
+- Deflection: Redirecting instead of engaging
+- Superficial validation: "Great point!" without substance
+- Capitulation: Simply agreeing
+- Dismissal: Correcting rather than exploring
+
+After targeted prompt improvements:
+
+| Profile | Before | After | Change |
+|---------|--------|-------|--------|
+| recognition | 72.5 | 80.7 | +11% |
+| single_recognition | 65.2 | 75.5 | +16% |
+| baseline | 51.2 | 41.6 | -19% |
+
+The recognition profiles improved substantially while baseline profiles scored lower—the refined scenario better separates recognition-oriented responses from baseline responses.
+
+---
+
+## 7. Discussion
+
+### 7.1 What the Difference Consists In
+
+The 101% improvement doesn't reflect greater knowledge or better explanations—both profiles use the same underlying model. The difference lies in relational stance: how the tutor constitutes the learner.
+
+The baseline tutor treats the learner as a knowledge deficit. Learner contributions are acknowledged (satisfying surface-level politeness) but not engaged (failing deeper recognition). The recognition tutor treats the learner as an autonomous subject. Learner contributions become sites of joint inquiry.
+
+This maps directly onto Hegel's master-slave analysis. The baseline tutor achieves pedagogical mastery—acknowledged as expert, confirmed through learner progress—but the learner's acknowledgment is hollow because the learner hasn't been recognized as a subject whose understanding matters.
+
+### 7.2 Recognition as Emergent Property
+
+Our most surprising finding is that explicit recognition instruction underperforms quality-optimized tutoring. How do we reconcile this with the factorial finding that recognition prompts work?
+
+The resolution: "recognition prompts" in the factorial study specify recognition *principles* (treat the learner as autonomous, engage with their contributions) rather than recognition as a *goal* (pursue mutual acknowledgment). The failed "recognition profile" in dyadic evaluation explicitly named recognition as what to achieve.
+
+This aligns with Honneth's observation that authentic recognition cannot be demanded or performed—it must arise from genuine engagement. Naming recognition explicitly may produce performative rather than genuine recognition.
+
+**Practical Implication**: Prompts should specify recognition-oriented *behaviors* and *orientations*, not recognition as an outcome to pursue.
+
+### 7.3 The Value of Internal Dialogue: Connecting to the Literature
+
+If recognition prompting accounts for 85% of improvement, what value does the Ego/Superego architecture provide? The 2×2×2 ablation study helps answer this question with greater precision than the original factorial design.
+
+**Quantifying the Contribution**: The multi-agent tutor architecture accounts for $\eta^2 = .088$ of variance—roughly 9% of explained variation. This is a medium effect size, comparable to many intervention effects in educational research. More telling is the marginal mean difference: moving from single-agent to multi-agent tutoring improves scores from 46.8 to 60.7 (+14 points), controlling for recognition.
+
+**The Synergy Finding**: The significant recognition × tutor interaction ($\eta^2 = .043$, p < .01) reveals that these factors work together non-additively. The interaction is *positive*: recognition prompts produce larger gains when combined with multi-agent architecture (condition 7 scores 79.8) than with single-agent (condition 5 scores 51.7). This suggests the Ego/Superego dialogue is particularly valuable for *implementing* recognition behaviors, where the line between performative and genuine engagement is subtle and requires internal review to navigate.
+
+**Relation to Drama Machine Literature**: Our findings both support and extend the Drama Machine framework:
+
+1. **Supporting evidence**: The framework predicts that internal dialogue should improve output quality through deliberative refinement and productive tension. Our 38% Superego intervention rate and 32% dialectical synthesis rate confirm that meaningful internal negotiation occurs. The architecture doesn't rubber-stamp Ego outputs; it genuinely evaluates them.
+
+2. **Novel finding on asymmetry**: The Drama Machine literature doesn't distinguish between *which agents* benefit from internal deliberation. Our ablation reveals a crucial asymmetry: the multi-agent architecture improves *tutor* performance ($\eta^2 = .088$) but not *learner* simulation ($\eta^2 = .001$). This suggests internal dialogue is valuable specifically for agents navigating genuine internal conflict—the tutor must balance warmth with rigor, learner responsiveness with pedagogical standards. The simulated learner has simpler goals (respond authentically) that don't benefit from internal deliberation.
+
+3. **The "unified learner" surprise**: Condition 7 (recognition + multi-agent tutor + *unified* learner) outperforms condition 8 (full system with psychodynamic learner deliberation), 79.8 vs 72.8. This counterintuitive finding suggests that learner deliberation may add *noise* rather than signal—the internal dialogue produces more variable learner responses that make tutor evaluation less reliable. For evaluation purposes, simpler learner simulation is preferable.
+
+**Practical Implications**: The optimal configuration identified—recognition prompts with multi-agent tutor and unified learner—suggests a division of architectural complexity: invest in tutor deliberation (where it produces measurable gains) but keep learner simulation simple (where elaboration may hurt).
+
+**Theoretical Coherence**: Beyond empirical contribution, the architecture provides theoretical grounding. The Ego/Superego structure operationalizes the Hegel-Freud synthesis developed in Section 3.5: internal self-evaluation (Freud) enables adequate external recognition (Hegel). Removing the architecture would leave the theoretical framework without an implementation mechanism.
+
+**Extensibility**: The structural foundation enables future enhancements:
+- **Transference dynamics**: Learners may develop relationships with the tutor that need explicit handling
+- **Id dynamics**: Impulses (frustration, boredom, curiosity) could be modeled as a third agent
+- **Superego calibration**: Different pedagogical traditions could be implemented as alternative Superego prompts
+- **Ghost plurality**: Multiple internalized authorities (different teachers, different traditions) could compete
+
+The architecture's contribution is thus modest in immediate effect but foundational for the research program.
+
+### 7.4 Implications for AI Alignment
+
+If mutual recognition produces better outcomes, and if mutual recognition requires the AI to be genuinely shaped by human input, then aligned AI might need to be constitutionally open to transformation—not just trained to simulate openness.
+
+Recognition-oriented AI doesn't just respond to humans; it is constituted, in part, through the encounter. This has implications for how we think about AI character and values: perhaps genuine alignment requires the capacity for mutual recognition, not just behavioral specification.
+
+---
+
+## 8. Limitations
+
+1. **Scale**: Evaluations use relatively small sample sizes (N=76 factorial, N=50 profile comparison)
+2. **Simulated Learners**: Real learners may behave differently than LLM-generated learners
+3. **Domain Specificity**: Evaluated on philosophy/technology courses; generalization uncertain
+4. **Judge Variation**: Inter-rater reliability between judges was moderate (ICC = 0.34)
+5. **Model Dependence**: Results may vary with different underlying LLMs
+
+---
+
+## 9. Conclusion
+
+We have proposed and evaluated a framework for AI tutoring grounded in Hegel's theory of mutual recognition, implemented through the Drama Machine architecture with Ego/Superego dialogue and psychodynamic tuning.
+
+Our central finding is twofold:
+
+1. **Recognition prompting works**: Recognition-enhanced prompts produce 85% of observed improvement, with large effect sizes on relational dimensions (personalization d=1.82, tone d=1.75, pedagogical soundness d=1.39).
+
+2. **Recognition must emerge, not be named**: Explicit instruction to pursue "recognition" underperforms quality-optimized tutoring. The path to recognition runs through quality, not through naming.
+
+This has implications beyond AI tutoring: wherever we seek authentic intersubjective engagement in human-AI interaction, we may need to specify conditions rather than outcomes.
+
+---
+
+**Code and Data**: https://github.com/machine-spirits/tutor
+
+---
+
+## References
+
+[References would be included here via BibTeX]