THE PROMETHEUS PAPERS
Questions into
working theory.
Open research on agents, memory, and the architectures that give intelligence a place to work.
The research archive
18 publications · Open accessThe Intent Compiler
Typed intermediate representations for reliable agentic systems.
Memory Architecture for Long-Horizon Agents
Taxonomy, invariants, and design principles for agents that last.
The Reproducibility Crisis in Agent Evaluation
Why agent benchmarks fail to repeat, and a protocol that fixes them.
Semantic Drift in Multi-Agent Orchestration
How meaning degrades across agent boundaries.
The Alignment Tax
Measuring the capability cost of safety interventions.
Tool Grounding and the Postcondition Problem
An agent that calls a tool must know what the call changed.
Context Windows Are Not Memory
The sliding attention buffer is a workspace, not a store.
Distillation as Architecture Transfer
Moving weights rarely moves the architecture of competence.
Agent Benchmarks Beyond Task Completion
Multi-dimensional scoring across fidelity, cost, provenance, and recovery.
Neural Knowledge Crystallization
How fluid model behavior condenses into stable knowledge.
Prompt Injection as Privilege Escalation
Injected instructions are a security boundary violation.
The Observer Effect in AI Evaluation
Benchmarks shape the behavior they measure.
Multi-Agent Coordination
Typed messages, shared state, and provenance.
Emergent Capabilities and the Competence Threshold
Which capabilities arrive discontinuously with scale.
Scaffolding Over Scale
Architecture as the driver of deployed agent capability.
Absence Blindness
The invisible gap in large language models and the case for externalized presence.
The Compiled World Runtime
An immutable-ontology, mutable-state architecture for reliable LLM agents.
Latent Relatedness
Attention as relational inference and the unintended profile engine.
AN OPEN CONVERSATION
Read. Question. Build.
Correspondence, replications, and collaboration are welcome.