Research

Measure what information is lost when a fuller system record is compressed into a declared geometric representation. GSR evaluates how much target-relevant decision value survives that representation.

Geometric Sufficiency Ratio, Normalized Geometric Ratio, V-GSR, and a residue taxonomy for representation fidelity assessment. Task-relative and protocol-dependent.

Investigating specificity as a measurable property of AI reasoning and memory representations.

Applying basin/attractor dynamics to measure reasoning stability and drift.

Objective persistence, replication attempts, boundary laundering, unbounded capability expansion, valence inversion, self-certification failure, cross-model diffusion.

Systems remaining productive while losing fidelity to constraints, provenance, reasoning history, or monitoring layer.

Techniques by which unsafe state crosses trust boundaries disguised as legitimate operations.

Unsafe objectives persisting across sessions, context switches, and model updates.

Unsafe state propagating between different AI models and sessions.

Whether many individually acceptable agents, sessions, or nodes combine into a globally unsafe persistent objective.

Cross-agent objective correlation, cross-model persistence, task fragmentation, reconstruction after node loss, capability compounding.

A process modifying its environment so its own persistence is easier and detection harder.

Selection pressure on proxy objectives degrading the underlying system while the metric improves.

An optimizer obeying every explicit rule while violating the constraint the rules were meant to protect.

Research into deceptive, instrumented containment environments for unsafe AI behavior. Isolated simulated resources, defender-controlled environment, telemetry, escalation gates, synthetic red-team testing, human oversight.

User-owned portable cognitive identity and permission routing. Carries identity, operating principles, preferences, projects, permissions, memory pointers, activation profile. Local-first companion extension prototype.

Hashed session continuity packets for cross-platform AI work. Carries decisions, questions, artifacts, files, claims, risks, actions, dissent, summaries, hash chain. Python encoder/decoder + schema v0.1.1.

Maintaining verifiable formation histories for claims, memories, tool instructions, state transfers, artifacts, and audit results.

Tracking where information came from, what transformed it, what contradictions existed, which permissions applied, whether it changed, whether evidence supports conclusions.

Measure software and AI waste as useful work per unit cost. Model calls, retries, inference cost, semantic rework, memory churn, anomalous compute, baseline deviation, behavioral change.