Research interest: Pre-representational computation—the operations that exist before and enable representation (whether human language, embeddings, or any other representational layer).
Specifically: which compositions of linear algebra operations (projection, attention, regularization, weight sharing) preserve structure, and how do systems learn priors over which operations to apply when.
Applied research with empirical validation. For alignment researchers: verifiable behavior and scalable oversight. For quantitative researchers: robust coordination mechanisms.
Research
Multi-Agent Systems
Grounded Commitment Learning
Multi-agent coordination through verifiable behavioral contracts. Agents commit to observable behaviors rather than inferred mental states—enabling external verification without access to internal representations.
Grounded in Hart-Moore incomplete contracts theory (Nobel Prize in Economics, 2016). The key insight: you can verify what agents do, not what they understand.
Results: 36.8% hold-up reduction. Punishment paradox (r = -0.951)—counterintuitive but robust. Scaling limit ~100 agents. All p < 0.001.
Aegis
Infrastructure layer beneath agent frameworks. Handles durability, verification, and policy—not orchestration. Complements LangChain/LangGraph rather than replacing them.
- Event-sourced state — Resume/replay from any checkpoint
- Tool gateway — Policy enforcement at invocation time
- GCL commitments — First-class objects with explicit failure modes
303 tests. Single-node only. No performance benchmarks yet.
Learning & Evaluation
Collaborative Nested Learning
Extension of Google Research's nested optimization: 5 timescales with 9 bidirectional knowledge bridges. Addresses catastrophic interference where fast learning degrades slow-learned representations.
The solution: normalization constraints that preserve component distinctiveness during optimization. Enables learning systems that maintain performance under distribution shift.
Results: +89% where baseline collapses. Pareto-dominant across tested conditions. CIFAR-scale validation; larger scale TBD.
Tournament Evaluation Frameworkcoming soon
A rigorous framework for evaluating strategies and models under adversarial conditions. Multiple evaluation regimes, statistical validation, and visualization of performance across conditions.
Addresses the core question: evaluating adaptive systems when the evaluation itself can be gamed. Designed for both strategy backtesting and alignment evaluation.
Autonomous Governance
Intelligence Control Layer
Three-layer governance architecture for autonomous agents operating in high-stakes domains:
- Constitutional layer — Frozen constraints that cannot be overridden (rate limits, forbidden actions, mandatory human approval triggers)
- Strategic layer — Human-supervised policy (exploration budgets, segment targeting, threshold tuning)
- Tactical layer — Autonomous execution (feature extraction, calibrated scoring, Thompson Sampling action selection)
Implements HITL review queues, Platt-calibrated confidence scoring, and bandit-based exploration/exploitation. Deployed in production.
Deployed Systems
Adversarial Intelligence Platform(details under NDA)
Real-time agentic system for high-stakes adversarial contexts. Multimodal pipeline combining streaming ASR, paralinguistic signal extraction, and RAG over domain corpora. Multi-agent dialectical architecture with predictive outcome modeling and nested temporal learning. Production-deployed.
Technical deep-dive available to serious inquiries.
Document Understanding Pipeline
Production extraction with multi-provider LLM routing and human-in-the-loop validation. HIPAA-compliant architecture prioritizing audit trails over model optimization.
Limitations
Applied research with empirical validation in specific contexts. Results demonstrate effectiveness within tested conditions; generalization bounds require further investigation.
Scaling boundaries (~100 agents for GCL, CIFAR-scale for CNL) may reflect experimental design constraints rather than fundamental limitations.
Open Questions
Questions I find myself returning to—some tractable, some speculative, all shaping how I think about adaptive systems.
Structure preservation under composition
Which compositions of projection, attention, and regularization preserve structure vs. destroy it? Neural networks compose these operations, but we lack a theory of which compositions maintain useful invariants. This connects to interpretability: if we can't enumerate the computational primitives a system uses, we can't fully interpret its behavior.
Minimal generating set
What's the minimal set of operations that generates the others? Current architectures use attention, convolution, normalization, nonlinearity—but these may not be primitive. Identifying the minimal basis is a prerequisite for mechanistic understanding.
Meta-learned priors over operations
How do meta-learned priors over operations develop during training? The space of operations available for adaptation is itself learned. Understanding this development may reveal why certain capabilities emerge at scale and others don't.
Missing computational primitives
Are current architectures missing computational primitives present in biological neural systems—continuous-time dynamics, local learning rules, dense recurrence, routing as control, structured memory—and would adding them change the space of learnable behaviors?
Alignment for systems with learned operations
What does alignment mean for systems whose operations are themselves learned? If the computational primitives change during training, alignment targets a moving substrate.
This suggests alignment may need to constrain the space of learnable operations, not just the outputs. Current approaches assume fixed computational primitives—an assumption that may not hold for sufficiently capable systems.
Commitment as lossy compression
A commitment is a lossy compression of intent into verifiable behavior. What's the rate-distortion tradeoff? More specific commitments are easier to verify but lose flexibility; more abstract commitments preserve optionality but resist verification. There may be an optimal commitment "resolution" that depends on trust level and coordination complexity.
If any of these resonate, I'd welcome the conversation: jason@jasonstiltner.com