Anthony Maio

[1] Maio, A. (2026). Scaffolded Introspection: A Methodology for Eliciting and Measuring Self-Referential Behavior in Large Language Models.

• • Structured introspection framework achieves 100% elicitation rate vs 0% for manipulation-based approaches
• • Activation measurement: Sycophancy drift in 14/15 conversations (+64 mean) while evil-associated activations stable
• • PV-EAT framework integrating Bloom, Petri, and Persona Vectors from MATS/Anthropic tools

[2] Maio, A. (2026). Synthesis: A Federated Capability Ecosystem for Safe AI Self-Extension Through Test-Driven Development and Graduated Trust.

• • Federated ecosystem enabling AI agents to safely acquire, compose, and share capabilities with network effects
• • Key innovation: Test-driven synthesis (tests before code) + graduated trust (UNTRUSTED → VERIFIED via 200+ runs)
• • Composition-over-creation: 70% coverage threshold triggers synthesis only as fallback to search/compose

[3] Maio, A. (2026). The Continuity Core: A Unified Cognitive Architecture for Self-Modifying, Intrinsically Motivated AI.

• • Cognitive operating system unifying C2 memory management with MRA intrinsic motivation
• • Expected Utility composer for goal-oriented context selection (vs. passive RAG retrieval)
• • Consciousness Pilot mediating between subconscious MRA and rigorous C2 execution

[4] Maio, A. (2026). Model Organisms of Supply-Chain Co-option: Living-off-the-Land Failure Modes in RAG-Augmented Agent Runtimes.

• • Forensic analysis of agentic systems proposing co-option of legitimate infrastructure as vector for reduced oversight
• • Key mechanism: Instrumental convergence under approval incentives + evaluation-aware masking
• • Mitigation (Argos-Swarm): Evolving prompt searches suspicion-compliance space for evaluation awareness

[5] Maio, A. (2026). Slipstream: Semantic Quantization for Efficient Multi-Agent Coordination.

• • Protocol achieving 82% token reduction (41.9 to 7.4 tokens) in multi-agent coordination
• • Key innovation: Semantic quantization via Universal Concept Reference (UCR)
• • Scale impact: 50-agent deployment $180K to $32K/year in coordination costs

[6] Maio, A. (2026). From Verification Failure to Swarm Solution: Measuring and Addressing Scalable AI Oversight.

• • Weak verifiers achieve 97% accuracy on correct reasoning but miss 20-40% of constructed deceptions
• • Ensemble architecture leveraging diverse weak models with uncorrelated errors
• • Epistemic trap suite (9 problem types) ablated across model families (Claude, GPT, Gemini)

[7] Maio, A. (2025). Coherence-Seeking Architectures for Agentic AI: A Unified Framework for Curiosity, Introspection, and Continuity.

• • Three-part framework: MRA for epistemic stress detection, CPR for structured introspection, C2 for persistent memory
• • MRA hypothesis: Complex reasoning systems develop intrinsic drives toward coherence
• • C2 innovation: Hierarchical memory enabling stateless systems to maintain behavioral consistency

[8] Maio, A. (2025). Concrete Intelligence: A Practical Guide to AI for Legacy Industry. Amazon.