Research Papers
31 papers — methodology, data, and conclusions you can verify
Composio in the MCP Delivery System
A decision-grade analysis of why Composio is included for commodity connectivity, how the wrap pattern protects brand and margin, and how delivery remains aligned to Database, Automation, and Judgment control boundaries.
The Wrap Pattern: Commodity Integration as Invisible Infrastructure
When MCP consumption is commoditized, the strategic response is to wrap — not build — the plumbing
A structural pattern for integrating commodity MCP vendors as invisible infrastructure while preserving the client-facing surface, the Intelligence Layer margin, and the Three-Tier alignment.
The Webflow Way, Automated
Agent-Ready Template Reviews on Published Sites (WebMCP + Review Snippet)
A case study on exposing Webflow Way QA signals to agents from a published template preview, aligned to WebMCP-style in-browser tools.
Open-Weight Models in Client MCP Work
A decision framework for when to use OpenAI gpt-oss (and safeguard) versus hosted frontier models in client education and implementation.
Guidance for consultancies building MCP integrations: how to choose between OpenAI open-weight models (gpt-oss-20b/120b, gpt-oss-safeguard) and hosted models, with concrete patterns for education, production, and compliance.
The Three-Tier Framework: Database, Rules, Policy
A structural model for agent systems, realized through Model Context Protocol
A hierarchical ontology identifying three tiers connected by typed Artifacts and spanning four cross-cutting concerns, with MCP as natural encapsulation.
The Andon Protocol
AI-native structured escalation for agent harnesses and multi-agent systems. v3.1 adds Silent Running Detection, cost-parameter defaults and worked examples, Resolution Surface design for batch review, and a three-phase implementation plan. The canonical boundary between Automation and Judgment in the Three-Tier Framework.
Ground: Verification-First Code Analysis
Case study: How Ground saved 8+ hours analyzing an 80+ package monorepo by preventing AI hallucination in code analysis.
Tufte for Mobile: Design Intent Across Screen Sizes
Applying Edward Tufte's principles to preserve meaning in responsive design
A methodology demonstrating how wireframe intent survives responsive transformation through five Tufte principles: data-ink ratio, sparklines, direct labeling, information density, and small multiples.
Ground: Evidence-Based Claims for AI Code Analysis
Computation-Constrained Verification Prevents False Positives in Agentic Development
A tool that blocks AI agents from claiming code is dead, duplicated, or orphaned without first computing the evidence. Now with AI-native features: batch analysis, incremental diff mode, structured fix output, and fix verification. Rated 10/10 by agent testing across two production codebases.
Recursive Language Models: Context as Environment Variable
Implementing MIT CSAIL's RLM pattern for processing arbitrarily large codebases through programmatic context navigation
This paper documents the implementation and empirical validation of Recursive Language Models (RLMs) based on MIT CSAIL research. We identified critical bugs, validated the pattern against the original repository, and demonstrated practical application for codebase analysis—processing 157K characters to find 165+ DRY violations.
Animation Spec Architecture: One Source, Two Renderers
Shared Specifications for Svelte and Remotion
A methodology for maintaining visual consistency between web animations (Svelte) and video exports (Remotion) through shared animation specifications that define what happens, while each renderer decides how.
Teaching Modalities: Finding the Right Medium for CREATE SOMETHING
Comparing Spritz, Motion Graphics, and Interactive Learning
An experiment exploring three modalities for teaching the CREATE SOMETHING philosophy: RSVP speed reading (Spritz), Vox-style motion graphics (Remotion), and interactive structured learning paths.