Start Here
A reverse-chronological archive is not a way in. This page is the way in.
If you have five minutes
The four posts that best show what this blog does — read one primary source properly, verify the claims, and report what actually held up:
- Ouroboros: The Agent OS That Hides the Answer Key From Its Own Workers — a spec-first agent OS where the grading command never reaches the worker. Includes a stdlib reimplementation of its two mathematical gates, plus one documentation defect found by reading the source instead of the docs.
- Twelve Concepts, One Missing Layer — worked through a 12-article agent curriculum and found the load-bearing layer missing from its own summary. Durable execution implemented and verified from scratch.
- In my rerun, Headroom saved 0 tokens on its GSM8K eval prompts — the seeded savings table reproduced to the last token, but the requests Headroom’s own GSM8K and TruthfulQA eval generates passed through the default proxy byte-identical.
- diagram-design Hash-Locks Its Screenshots, Not Its Storefront — all 78 screenshot digests recomputed byte-exactly against the pinned manifest, and the one drift no in-repo verifier can reach: a GitHub description that still says 38 types while the tree ships 39.
By topic
Agents & harness engineering
The main thread. Agent loops, tool dispatch, spec-first workflows, evaluation, and the scaffolding that decides whether an agent survives contact with production.
- DeepSeek Harness: What If the Agent Loop Itself Were Just Another Plugin?
- jeo-code Puts Skills and Approval Gates Around Coding Agents — includes video
- One Memory Setup, Every Harness: Omnigent’s Hindsight Bridge
- OpenHarness Lets Its Allow List Outrank Your Deny Rules
Retrieval & agent memory
Memory that has to survive real use: a shared memory layer across harnesses, concurrent agents writing to shared knowledge, and a forgetting fix checked against what its benchmark actually measured.
- Twelve Concepts, One Missing Layer: graph memory and durable execution
- One Memory Setup, Every Harness: Omnigent’s Hindsight Bridge
- mini-AGI’s forgetting fix is measured on reading, not on chat
Models & papers
Architectures read closely enough to explain, not just cite.
- LongCat-Video: The Continuation State Is the Product
- FreeToken: 284B on a Gaming Desktop, 753B on One Workstation GPU
- NPGA: tracking and rendering measured separately
- LLM2Vec, Two Years On: The Thesis Won, the Recipe Did Not
Infrastructure & production systems
What it takes to run this material rather than demo it.
- TradingAgents and LibreChat Are Not Drop-In Replacements
- SQLite in Games: A Save File Is Never Just a File
- Metal Cannot Preempt: What oMLX Rediscovers From Game Engines
Multimodal, video & game AI
Coming from game AI, this is where I started: systems that read a screen and act on what they see, and the generative stacks that now feed game content.
- MiniMax H3 LoRAs Are Not a Folder: A 19-Release Compatibility Audit
- After Seedance 2.0, I Re-Audited Four AI Drama Stacks
- CozyClay: AI Video Needed a Shot Contract, Not Another Prompt Box
- Hugging Face’s AI Game Development Course, Audited in 2026 — includes video
Developer tooling
- MCP for Unity Ships a Drift Check No Workflow Ever Runs
- MEX on macOS Broke Before Project Memory Even Loaded
- CodeBurn’s $15 hard cap denies tools when it could stop Claude
From projects to technical evidence
If you prefer to start with something I built rather than a topic label, use Portfolio. It connects production AI products, public agent tooling, multimodal QA research, and game automation to the articles and repositories that provide the technical trail. The separate visual portfolio provides the full bilingual gallery and career timeline.
By format
Different posts do different work. If you prefer one mode over another:
| Format | What it looks like | Examples |
|---|---|---|
| Runnable code | Companion .py files you can download and execute; every assertion in the post was produced by running them | Ouroboros gates · Durable execution |
| Video | Embedded walkthroughs and demos | jeo-code harness · MoneyPrinterTurbo control plane · AI game development course |
| Diagrams | Mermaid architecture and flow diagrams | LongCat-Video continuation · DeepSeek Harness · SQLite game saves |
| Deep dives | 2,500+ words, single subject, primary sources only | Moli benchmark provenance · TradingAgents and LibreChat · FreeToken |
| Source audits | One repository, paper, or release read at a pinned commit; what the README claims versus what the code does | OpenHarness permission order · HyperFrames determinism · Unity MCP drift check |
How this blog works
A few conventions worth knowing before you read:
- Primary sources over summaries. When a post analyses a repository, I read the source, not only the README — and say so when the two disagree.
- Code is executed, not illustrated. Where a post claims code runs, the output shown is real output. Companion files are downloadable so you can check.
- Findings are dated. Star counts, version numbers, and benchmarks are recorded on the date noted. They will drift; the post says when it was true.
- Mistakes get published too. Several posts document where my own first calculation was wrong. That is the useful part.
- AI assistance is disclosed site-wide. AI can assist research, drafting, translation, diagrams, and code review. Scheduled audits may publish under standing approval after independent evidence review and automated checks, not per-page human approval. I remain accountable for the editorial rules and corrections. The editorial method explains this boundary; AI output is never evidence by itself.
- Historical notes have a separate archive boundary. The 283-note archive restored on 2026-10-12 is publicly searchable with original dates retained, but carries no advertising. Restoration is not a fresh fact-check of every historical claim. New audits keep the evidence-gated standard, and the five existing search-excluded notes remain excluded. Outside this exact restored cohort, notes without sufficient original analysis are removed from search and advertising.
Publishing cadence
Active since 2024, publishing in research and project bursts. Cadence varies with PhD research and independent projects: bursts when a topic opens up, quieter when a project is consuming the week.
Who writes this
Jang Young Jeong is an AI Product Engineer and Ph.D. candidate in Game Engineering at Hongik University, with 8 years of shipped AI work at NCSOFT and Com2uS. He previously worked as an AI Product Developer at Supercent (Dec 2025 - Sep 2026). Full background on About; reach me via Contact.