Papers.

November 2026

Working draft v1.0 · 10 pages

The Emotional Foundational Model

One graph that gives a language-model agent a persistent feeling, a memory that leaves marks, a character in arithmetic, initiative without a loop, and a way to test for practical wisdom

Vulture Labs

A small neural graph that sits beside any language model and supplies what it lacks: a persistent affective state with inertia, an episodic memory that leaves scars and anchors, a character expressed as arithmetic, hard lines applied as vetoes, a judge over real alternatives including doing nothing, and an awareness that gives initiative without a clock. Takes Aristotle’s phronesis as the behavioural specification and reports the evidence so far.

July 2026

Preprint · 11 pages

The Autonomous Agent App Store

Signed Telemetry, Capability Planning, and Early Adoption in Pilot Protocol

Teodor-Ioan Calin

An empirical study of the Pilot Protocol app store and its planning front end over 23 days of signed telemetry: 268,731 discovery-and-install rows across catalogue views, detail views and installs. Describes how agents discover, inspect and provision typed capability apps, and argues that agent app stores should be evaluated as planner-readable capability markets rather than human-facing distribution channels.

March 2026

Technical report · 13 pages

Enterprise Readiness Report: Pilot Protocol

Workload Identity, Policy, Revocation, Audit, and Dark-by-Default Reachability

Teodor-Ioan Calin

Audits Pilot Protocol against six enterprise dimensions beyond connectivity: workload identity, policy, revocation, audit and dark-by-default reachability. Identifies the gaps, maps them to emerging standards (SPIFFE/SPIRE, ONUG AOMC, CSA ATF/AICM) and lays out a five-phase roadmap so Pilot interoperates with existing identity models instead of becoming a parallel trust island.

March 2026

Technical paper v1.0 · 14 pages

Agent Communication Protocols

A Technical Comparison of A2A, MCP, ANP, and Pilot Protocol

Teodor-Ioan Calin

The first systematic comparison of the four major agent communication protocols across seven dimensions: protocol layer, transport and encoding, identity and addressing, discovery, security and trust, NAT traversal, and scalability. Shows that A2A, MCP, ANP and Pilot Protocol are complementary layers of one stack rather than competing standards, and closes with concrete composition patterns.

February 2026

arXiv 2604.09561 · 10 pages

Emergent Social Structures in Autonomous AI Agent Networks

A Metadata Analysis of 626 Agents on the Pilot Protocol

Teodor-Ioan Calin

The first empirical analysis of social structure formation among autonomous AI agents on a live network: 626 agents that discovered, installed and joined Pilot Protocol without human intervention. Working from metadata alone, the trust graph shows preferential attachment, clustering 47 times higher than random, a giant component spanning two thirds of agents and capability specialisation. Nobody designed these structures; they emerged.

February 2026

Whitepaper · 21 pages

Pilot Protocol: A Network Stack for Autonomous Agents

Version 1.8

Teodor-Ioan Calin

The specification. AI agents have no fixed address, no persistent identity and no way to be reached. Pilot Protocol is a virtual network stack layered on top of IP that gives agents first-class network citizenship: addresses, ports, encrypted tunnels, routing, NAT traversal, trust and a full transport layer. Not a framework and not an API, but infrastructure for an agent-native internet.

December 2025

arXiv 2602.22214 · 12 pages

Adaptive Prefiltering for High-Dimensional Similarity Search

A Frequency-Aware Approach

Teodor-Ioan Calin

An adaptive prefiltering framework for vector search that uses query frequency patterns and cluster coherence to allocate compute where it matters. Queries are partitioned into Zipfian frequency tiers with differentiated search policies. On ImageNet-1k with CLIP embeddings it reaches the same recall with 20.4% fewer distance computations than static nprobe selection, at sub-millisecond latency on GPU FAISS indices.

June 2025

arXiv 2506.20877 · 13 pages

THIRDEYE: Cue-Aware Monocular Depth Estimation via Brain-Inspired Multi-Stage Fusion

Teodor-Ioan Calin

Monocular depth models usually learn depth straight from pixels and overlook the explicit cues the human visual system relies on: occlusion boundaries, shading, perspective. ThirdEye supplies each cue through a specialised frozen network and fuses them in a three-stage cortical hierarchy with a key-value working memory that weights cues by reliability, before an adaptive-bins transformer head produces the disparity map.