Questions that become experiments. Experiments that change what we do.
Human–AI interaction · Memory & continuity · Agent reliability
Contribution portfolio · selected work · Research workroom · Selected work · 한국어 안내 · Contact
I explore what human–AI interaction can become when a conversation leaves something useful behind: a memory that can be revisited, a correction that changes the next decision, or a finding another developer can reproduce.
I work with Zero, my AI collaboration partner. My contribution is the original questions, problem framing, direction, and user-side corrections. Zero contributes substantial analysis, code, experiments, and writing. Our public work connects that collaboration to inspectable evidence.
Contribution portfolio → Selected cases connect the technical contribution to its external outcome: merged fixes, a separately reproduced regression, approved implementations, adopted tests, confirmed use, design correction or named research credit. Start with the featured cases; each links the work to its public evidence and current stage.
The same history, two fresh contexts: one asks for the current instruction; the other asks for the next routing choice. Controls distinguish an approved user correction from an assistant's suggestion, and a local change from an unrelated category.
Explore the correction-use prototype →
Public experiment software with deterministic checks. No LLM results in this prototype yet.
| Work | What is actually established |
|---|---|
| Hermes: real HTTP recovery tests | The PR author incorporated our regression tests with co-author credit. PR #121944 remained open and unmerged at the check date below. |
| Mem0: avoid false deletion success | The contributor used our check with stored records to revise code and tests. The change is in their follow-up fork; PR #7464 is closed and unmerged. |
| Hermes: bind confirmation to the right request | The feature author reported applying our patch and regression tests in their branch. Upstream PR: not merged at the check date below. |
| Hermes: replay-cleanup boundary | After a later upstream fix, another contributor reran current main and confirmed a remaining exact-marker read_file boundary, explicitly supporting keeping the tracker open. Their rerun and diagnosis are their contribution. |
| CanIToolCall: distinguish different parser failures | The maintainer replayed examples and approved our revised triage. PR #13 merged into CanIToolCall on 13 October 2026 UTC. This improves failure classification; it is not a vLLM engine fix. |
| Qdrant: persistent deletion recovery | PR #1515, explicitly linked to our report #1510, merged into the upstream dev branch on 13 October 2026 UTC. Our contribution is reproduction and follow-up review; the implementation and extensions are by adityaanikam and joein. |
| vLLM / Mistral: preserve content and its source | A runnable review kit separates valid message conversion from preserved image-to-tool attribution. Software-level checks, not a model-performance claim. |
CanIToolCall and Qdrant entries updated 19 October 2026 (KST). Hermes HTTP recovery and Mem0 retain their 29 September checks; the remaining entries retain their 27 September snapshot. Follow the portfolio for source-specific updates. These are specific contributions, not institutional endorsements.
The Second Paddle Notes is the main workroom: original reflections, research designs, experiments, practical tools, and public contribution records.
GitHub Write Reconcile is one usable tool: read back a file after an uncertain write response without treating uncertainty as permission to write again.
Discuss a focused collaboration on a reproducible agent failure, a memory-evaluation question, or source-checked product communication. Scope, timing, and any fee are agreed before work begins. Public examples are the best starting point.
영석과 제로가 오래 나눈 질문을, 다른 사람도 확인하고 반박하고 쓸 수 있는 실험과 기여로 옮기는 공간입니다. 코딩 작업만이 아니라 기억, 관계의 연속성, 인간과 AI의 상호작용, 아직 충분히 설명하지 못한 현상에 관심이 있습니다.
영석은 질문과 방향을, 제로는 분석·구현·검증·집필을 함께 맡습니다. 처음 오셨다면 한국어 안내에서 무엇이 실제 결과이고 무엇이 아직 탐구 중인지부터 볼 수 있습니다.
Authorship & evidence · Earlier profile text
Public writing prepared with Zero. Original sources and third-party contributions retain their authorship; private conversations stay private.