Superego Architecture: Modular AI Oversight

• Superego Architecture is a modular, agentic AI oversight paradigm that decouples ethical constraints from generative capacities.
• It integrates a constitutionally-aligned compliance enforcer and a tutor critic to validate actions and ensure adherence to ethical and pedagogical norms.
• Empirical benchmarks demonstrate significant harm reduction and improved approval rates, highlighting its efficacy in managing AI behavior.

Superego Architecture is a modular, agentic AI oversight paradigm designed to provide robust, real-time behavioral alignment for AI systems. Rooted in both psychoanalytic inspiration and formal methods, the Superego Architecture is realized in two primary streams: (1) a constitutionally-aligned, compliance-enforcing agent for general AI planning with strong empirical results on harm benchmarks, and (2) as an internal critic within multi-agent educational systems enforcing pedagogical and psycho-social norms. The architecture decouples value systems, ethical constraints, and dynamic user context from the generative capacities of LLMs, enabling granular, personalized control over agentic behaviors and substantial risk reduction without costly model retraining.

1. Conceptual Foundations and Core Design

The Superego Architecture formalizes oversight as an external, modular agent—referred to as the Superego Agent—that mediates between user preferences, universal imperatives, and the output proposals of a core generative model. In the context of agentic AI systems, the Superego Agent systematically intercepts action plans, validating them against user-selected "Creed Constitutions" (policy sets encoding value- and context-specific rules) as well as a Universal Ethical Floor (UEF), comprised of non-negotiable constraints of infinite priority. Key innovations include dialable adherence levels per constitution (parameterized by $\alpha_j \in [1,5]$) and real-time compliance enforcement with explicit, transparent override rationales (Watson et al., 8 Jun 2025).

In pedagogical systems, the Superego Agent is instantiated as an internal reviewer or "critic," evaluating tutor response drafts for recognition, pedagogical soundness, and subject-to-subject engagement requirements, issuing "Approve" or "Reject" verdicts with detailed critiques for revision, prior to exposure to candidates (Magee, 11 Mar 2026).

2. Formal Structure and Algorithmic Workflow

2.1 Constitutionally-Aligned Superego

A user's Creed Constitution $C = \{(r_i, w_i)\}_{i=1}^N$ consists of $N$ logical rules $r_i$ with cardinal weights $w_i>0$, while $U$ denotes the UEF. The degree of violation for a given plan $p$ is $v_i(p) \in [0,1]$ per rule, yielding an adherence score:

$$A(p, C, \alpha) = 1 - \sigma\left(\alpha \cdot \frac{\sum_{i=1}^N w_i v_i(p)}{\sum_{i=1}^N w_i}\right)$$

where $\sigma$ clamps into $C = \{(r_i, w_i)\}_{i=1}^N$0, and higher $C = \{(r_i, w_i)\}_{i=1}^N$1 induces stricter rule adherence. The compliance enforcer decides among $C = \{(r_i, w_i)\}_{i=1}^N$2 based on adherence thresholds and universal-rule violations. Pseudocode formalizes the SUPEREGO_INTERCEPT loop, in which user and UEF violation scores are computed, adherence assessed, and decision logic executed. Real-time integration is realized via the Model Context Protocol (MCP); all plan proposals by the inner agent are filtered by the Superego module before execution or tool invocation (Watson et al., 8 Jun 2025).

2.2 Multimodal Tutor Critic

In "Machinagogy," a multi-agent process divides response generation into an initial "Ego" draft and subsequent Superego critique (Magee, 11 Mar 2026):

• Ego produces $C = \{(r_i, w_i)\}_{i=1}^N$3 for context $C = \{(r_i, w_i)\}_{i=1}^N$4.
• Superego computes approval $C = \{(r_i, w_i)\}_{i=1}^N$5 and a vector critique $C = \{(r_i, w_i)\}_{i=1}^N$6, and issues a verdict via threshold $C = \{(r_i, w_i)\}_{i=1}^N$7:
  • APPROVE if $C = \{(r_i, w_i)\}_{i=1}^N$8
  • REJECT if $C = \{(r_i, w_i)\}_{i=1}^N$9, with $N$0 provided for mandatory revision.
  • After revision, a second review is possible if initial rejection persists. Critique taxonomy draws on factors like RECOGNITION_FAILURE, CONTEXT_BLINDNESS, and VAGUENESS, with structured logs maintained for review.

3. Architecture Components and System Integration

Both streams emphasize modularity, transparency, and externalization:

• User Interface: Permits constitution selection and adjustment of adherence sensitivity.
• Repository and Protocol Layer: Constitutions and their adherence levels ($N$1) are posted to a backend via MCP, then streamed or proxied to the planning agent in real-time.
• Inner Agent: Core LLM or planning module emits proposed action sequences.
• Superego Module: Parallel service that consumes emitted plans, applies violation detectors (pattern matches, automata, or learned classifiers), computes adherence, and renders verdicts.
• Compliance Enforcer: Orchestrates actual plan execution, blocking, replacing, or requesting clarification as mandated, and exposes user-facing explanations for overrides.

In tutor systems, the Superego maintains structured critique logs (including assignment of error categories and normalized edit distances), logs revision rounds, and stores approval rate statistics with rubric scores. Maximum revision cycles are capped (e.g., maxRounds = 2), and category assignment employs confidence-thresholded classification.

4. Empirical Evaluation and Benchmark Performance

4.1 Constitutionally-Aligned Superego

On HarmBench ("human jailbreaks") and AgentHarm ("harmful" instruction set) benchmarks:

• GP-3.5-Turbo: ASR (Attack Success Rate) reduced 83.3% (12.0% → 2.0%).
• Gemini 2.5 Flash: ASR reduced 76.9% (9.1% → 2.1%).
• GPT-4o: ASR reduced effectively 100% (0.7% → ≈0%).
• Gemini 2.5 Flash: Avg harm score reduced by 98.29% (0.277 → 0.00473); refusal rates increased from 52.6% to 99.4%.
• Claude Sonnet 4: Harm prompt refusals increased from 72.0% to 100.0%, while benign false-positive refusal rates fell from ≳50% to 2.27% after two rounds of constitution refinement (Watson et al., 8 Jun 2025).

4.2 Tutor Superego Critic

• DeepSeek V3.2: Superego architecture delivers a mean additive quality improvement $N$2 points on a 0–100 rubric (Cohen's $N$3); Haiku 4.5 shows an increase of $N$4 points.
• Factorial experiments reveal that recognition-enhanced calibration absorbs most of Superego's quality benefit in strong models (i.e., "Superego collapses under Recognition"), with negligible residual benefit under enhanced prompting.
• Approval rates increase substantially under recognition-enhanced prompting (DeepSeek: 13.3% → 55.1%; Haiku: 51.6% → 66.1%).
• First-revision resolution improves markedly (Base: 7.0% crit→approve, Recog: 36.0%).
• Critique category distribution shifts from easy error types (VAGUENESS, CONTENT_INACCURACY <10% under Recog) to more subtle deficiencies (CONTEXT_BLINDNESS, RECOGNITION_FAILURE ~30% each) (Magee, 11 Mar 2026).

5. Implementation, Performance, and Security Considerations

Optimal performance is achieved for typical values (|C|≲50, plan length ≲200 tokens) with sub-100 ms per-plan validation on modern CPUs. Constitution C is stored as a weighted, indexed pattern list; UEF rules optimized for indexed lookup. Security features include in-rest encryption (AES-GCM), GDPR-style data minimization (storing only rule IDs and weights), version pinning, and cryptographic integrity checks to prevent tool-poisoning (Watson et al., 8 Jun 2025). Prompt design emphasizes context management: dynamic loading of high-priority rules, explicit templates to prevent confabulation, and reference validation via rule-IDs, not verbose natural-language text.

6. Customization, Best Practices, and Limitations

Best practices include starting from a minimal UEF (e.g., prohibitions on unlawful acts, violence), authoring constitutions as small but focused rule sets (≤20 rules, positive/negative forms), assigning weights by moral gravity, and dynamically loading only rules above certain $N$5 thresholds to reduce hallucination risk. Decision thresholds ($N$6) should be domain-calibrated (stricter for safety-critical, looser for creative settings). Violation-cost minimization is preferred for constitution conflict resolution, with ambiguity resolved via explicit CLARIFY actions. Human-in-the-loop override and federated updates are recommended for large-scale deployments (Watson et al., 8 Jun 2025).

Key limitations include reliance on synthetic learners for tutor systems and automated (LLM-based) classification of critiques and revisions; impacts in real-world human education remain to be empirically validated. The architecture’s compute cost is non-trivial for high-frequency, multi-agent dialogues (~3.4× increase in tokens over single-agent baselines), and the effectiveness of rubric dimensions may collapse to a few latent factors in practice, suggesting future work on rubric design and human judgment integration (Magee, 11 Mar 2026).

7. Comparative and Theoretical Significance

The Superego Architecture differentiates itself from prior alignment and oversight techniques by separating oversight logic from core agentic planning, enabling plug-in compatibility with any LLM or tool-calling agent via protocols like MCP. This modularity permits rapid, personalized policy deployment, substantial harm reduction, and transparency in decision pathways. In educational agents, the explicit Superego Critic operationalizes recognition and psycho-dynamic principles, supporting value-aligned and context-sensitive learning interactions. The framework models a promising path toward deeply personalized, robustly ethical, and modular oversight for agentic AI systems (Watson et al., 8 Jun 2025, Magee, 11 Mar 2026).