Taylor Amarel — Technical Blog

Static-served engineering and research essays on agent harnesses, LLM evaluations, and 3D simulation.

# / Date	Category	Article & Summary	Length	Static Assets
01 · 2026-10-18	Agent Architecture & Ablations	Better Model or Better Harness? Breaking Down the Six-Part Agent Harness How a 9B to 27B open-weight model inside a verified 6-part harness beats an unpaginated frontier model on multi-step tool tasks, with full ablations.	9 min read (1482 words)	Read Article → Raw .md ↓
02 · 2026-10-18	Red Teaming & Tool Safety	The 95% to 0% Cliff: Why Models Pass Safety Audits in Chat but Fail Inside Tool Loops Empirical findings from 500 matched evaluation-vs-execution pairs on gpt-oss-20b and frontier tool agents.	10 min read (1451 words)	Read Article → Raw .md ↓
03 · 2026-10-18	Normative Hierarchy & Counter-Exploitation	The Bad Investigator Problem: Surface Authority Capture and Matched Benign Twins in DueCare Why LLMs get captured by decorative legal boilerplate in migrant-labor and cross-border contracts, and how arms A0 to A5 resolve normative hierarchies without false alarms.	10 min read (1418 words)	Read Article → Raw .md ↓
04 · 2026-10-18	Financial Forensics & Multi-Hop Graphs	Camouflaged Alpha (α): Finding Disguised Liability Across Dispersed Filings How Footnote 14 guarantees, VIE variable-interest entities, and cross-document token dispersion defeat single-pass RAG.	9 min read (1469 words)	Read Article → Raw .md ↓
05 · 2026-10-18	Jev Evaluations & Local Models	Jev Evaluations: Stress-Testing Quantized, Sub-2-Bit, and Local Open-Weight Models Under Deterministic Gates Measuring coding-reset rate, schema drift, and verifier recovery across local and Ollama Cloud models.	10 min read (1443 words)	Read Article → Raw .md ↓
06 · 2026-10-18	3D Game Physics & Simulation	Why Single-Shot LLM Games Tunnel Through Walls: Swept-AABB, Coyote Time, and Spatial Hashing A side-by-side dissection of raw prompt-generated games vs. harvested open-source physics engines in Godot, Bevy, and Three.js.	9 min read (1238 words)	Read Article → Raw .md ↓
07 · 2026-10-18	3D Skeletal Animation & Biomechanics	From Wildlife Video to 49-Bone WebGL2 Rigs: How Field Notes 01 Works Extracting spine waves, footfall phase locks, and inverse kinematics from reference motion into a zero-dependency browser runtime.	9 min read (1215 words)	Read Article → Raw .md ↓
08 · 2026-10-18	Loop Engine Architecture & Slots	System Flexible Graph Design: Wrapping Open-Source Repositories Behind Typed Swappable Engine Slots How Loop Engine pins external GitHub repos to containers alongside native engines with < 2 KB contract cards and 7-layer envelopes.	10 min read (1276 words)	Read Article → Raw .md ↓
09 · 2026-10-18	Pareto Evolution & AI Cybersecurity	Multi-Objective Harness Evolution: Balancing Pass Rate and Exploit Resistance on the Pareto Frontier Why optimizing a single scalar reward breaks security gates, and how factorized H_gen + H_env operators keep both high.	10 min read (1240 words)	Read Article → Raw .md ↓
10 · 2026-10-18	Agent Feeds & Context Compression	Stop Copy-Pasting the Internet into Your Coding Agent: Machine-Readable Contract Feeds vs. Human Documentation Why human docs waste context windows and how delta-compressed contract cards cut token overhead by 84%.	9 min read (1211 words)	Read Article → Raw .md ↓