Fodev JEO

LATEST INVESTIGATIONS

Source audits & field notes

Prompt repetition vs baseline accuracy across popular LLMs showing 47 wins out of 70 tests

Prompt Repetition: Approximating Bidirectional Attention in Causal LLMs

A source audit of Google Research’s prompt repetition study (arXiv:2512.14982), analyzing how duplicating input tokens fixes causal attention blind spots in non-reasoning LLMs.

Jan 20, 2026 AI, Research

#prompt-engineering #transformers

Diagram of GLM-4.7’s task-specific evaluation settings and four vendor-reported scores. It notes τ²-Bench top-p is not listed and Preserved Thinking is recommended for multi-turn τ²-Bench and Terminal Bench 2 tasks; scores were not independently rerun.

GLM-4.7’s scorecard needs its test protocol beside it

A source-bound reading of GLM-4.7’s reported scores, task-specific decoding settings, τ²-Bench changes, and what the results can establish.

Jan 18, 2026 AI, Research

#glm-4.7 #benchmarks

Diagram showing a manifest-to-file-read workflow, stable IDs for current records, and a separate indexed-search path where the Rust missing-index status is not confirmed as CLI output.

OpenContext: Stable IDs Point to Current Files, Not Snapshots

A source audit separates OpenContext’s Markdown store, agent guidance, manifest metadata, stable IDs and indexed search, without claiming automatic memory or benchmark gains.

Jan 17, 2026 AI, Agents

#opencontext #agent-memory

Rnj-1 and Motif-2 Open Source Architecture Comparison

Open-Source Frontier Models: Architecture and Benchmarks of Rnj-1 and Motif-2

Deconstruct architectural innovations in frontier open-source models—from global self-attention and YaRN context scaling to dense synthetic data curation.

Dec 11, 2025 AI, Research

#open-source-llm #transformer-architecture

Deconstructing 17 Agentic Architectures: From ReAct to Metacognition

A source audit of Fareed Khan’s 17 agentic AI architectures in LangGraph, evaluating single-agent loops, multi-agent swarms, memory tiers, and 2026 production tradeoffs.

Dec 10, 2025 AI, Agents

#agents #orchestration

Production Generative AI Stack Architecture Diagram

Production Generative AI Stack: Architecture, Gateways, and Resilient Serving

A source audit of production generative AI stacks, analyzing inference gateways, hybrid RAG pipelines, agent orchestration, and end-to-end observability.

Dec 8, 2025 AI, Tooling

#generative-ai #architecture

Evaluating Deep Agents CLI on Terminal Bench 2.0 with Harbor

A source audit of LangChain’s Deep Agents CLI evaluation on Terminal Bench 2.0 via Harbor, examining sandboxed harness execution and the leap from 42.5% to 66.5%.

Dec 7, 2025 AI, Agents

#agents #benchmarks

Source Audit: Hugging Face TEI and Sub-Millisecond Embedding Serving

Auditing huggingface/text-embeddings-inference: evaluating Rust tokenization, token-based dynamic batching, and FlashAttention kernel acceleration.

Dec 5, 2025 AI, Tooling

#embeddings #rust

Combining Rust and Python for AI Systems: PyO3, GIL Bypass, and Native Serving

A source audit of Rust and Python integration in AI architectures, analyzing PyO3 FFI boundaries, zero-copy buffer sharing, and concurrent data pipelines.

Dec 5, 2025 AI, Tooling

#rust #python