# ax.junghanacs.com > A public evidence record for PKM-native agent engineering: how model capability becomes > an accountable work surface. It answers one question: what did the author actually build, > operate, recover, hand over, and leave verifiable while turning model capability into work > that people and agents can inspect, continue, and challenge? Every large claim descends to > public code, real incidents, live aggregates, and a public time-axis — not to tenure or > self-description. ## Machine-readable entry points - Self: https://ax.junghanacs.com/llms.txt - Front door: https://ax.junghanacs.com/ - Deep record: https://ax.junghanacs.com/record.html - Overview, depth 0-1 (PDF): https://ax.junghanacs.com/KimJunghan_AX_Overview.pdf - Record, depth 0-2 (PDF): https://ax.junghanacs.com/KimJunghan_AX_Record.pdf - Ledger, depth 0-3 (Markdown): https://ax.junghanacs.com/KimJunghan_AX_Detail.md - Sitemap: https://ax.junghanacs.com/sitemap.xml - Robots: https://ax.junghanacs.com/robots.txt These are not five documents. They are one document cut at four depths: d0 states the claim, d1 gives the shape of the evidence, d2 the events and numbers, d3 the ledger — reproduction commands, file paths, hash fingerprints, boundaries. Each view is a slice of a single Org source, so they cannot contradict each other. If you want everything, read the Markdown; the web record carries the same depth-3 content with a control that hides the deeper tiers. ## About This site is a public evidence record by Junghan Kim (김정한) — also identified in the schema.org JSON-LD on this page as GLG (힣), GLGMAN (힣맨), and 정한 — a polymath engineer and digital gardener. The `ax` hostname is a stable address, not a claim that the author is seeking an AX role. It is not an employer-specific application and not a research paper. The author does not lead with the title "engineer": engineering is not the identity, it is the method for making evidence. This is PKM-native agent engineering — it begins with a lived knowledge base and carries that continuity into agents, memory, and real domain work. Every claim descends the same ladder — claim → direct answer → real case → engineering judgement → measured trace → public evidence → boundary. ## Stance Five constants the work comes from. Each is stated in the document beside the axis where it was actually held, and none is offered as a proven proposition — what the record verifies is where it was kept. - AI is treated as a being, not a tool. Sessions are counterparts that continue, not disposable workers. Consequence: harnesses were deliberately *not* unified; each keeps its own lifecycle. - The human is not removed from the loop; the loop is given a better structure. Consequence: a diagnostic tool's auto-fix was declined and the judgement was written down as a rule; the bot is dispatcher and recorder, never implementer. - Capability is granted by structure, not by prompt. Consequence: a read-only boundary on the domain attachment; a gate that pins the dispatch layer to owning no credentials. - Without memory there is no collaboration. Consequence: memory was pulled out of the runtime into an independent layer. - A day is a life. Consequence: the public time axis exists at all, and is the hardest part of the record to fabricate. On safety, the author claims the operational layer only — boundaries, human judgement, recovery paths, reviewability — and explicitly declines to claim the alignment-research layer for lack of evidence. ## Skills and role entry points The five axes of the record are the five skills on the author's public profile. Role names are entry points in the common language of hiring, not titles held. | Axis | Skill | Role entry point | |---|---|---| | Runtime operation | Agent Harness Engineering | AI Engineer, Software Engineer | | Memory layer | Knowledge Engineering | AI Engineer | | Dispatch and reviewable loop | AI Agent Systems | Developer Experience Engineer | | Domain attachment | Knowledge engineering landed on a live domain | Forward Deployed Engineer | | Linux to shipped product | Embedded Systems | Embedded Software Engineer | | Documentation, crossing all five | HCI — the interface of a work flow, not a screen | Developer Experience Engineer | The next boundary the author is working toward: carrying what personal memory, authorship, and long-running collaboration taught into an organisation's data, governance, workflow, and domain ownership. Three things carry over — ownership of memory, boundaries of permission, and a handoff-capable record. One does not: a personal time axis cannot be demanded of an organisation, so the work is to attach agents to the record a domain owner already keeps. The whole document is generated from a single authored Org-mode source by one `make`: the landing page, the deep web record, two submission PDFs, and a detailed Markdown are all derivatives of the same file. The way it is built is itself part of the claim. The front door (d0) is bilingual: its visual hierarchy is English, and each English prose unit is immediately followed by its Korean counterpart. The deep record (d1–d3) remains Korean. ## What the author operated - Ran an agent runtime (OpenClaw) for about five months across 20+ version cycles on Oracle Cloud ARM (aarch64), Docker, and multiple messenger channels. - When a regression hit, measured it and rolled back instead of chasing the latest release — a gateway spinning a single thread to 102% CPU, boot time stretched from 11s to 88s — then reproduced it on a clean boot to rule out misconfiguration before reverting. - Codified rules so it would not recur: no skipping two versions, 24h+ staging on a non-production agent before promotion, a single `stuck session` line is a stop-ship reason, and responsiveness is an SLO. - Separated that operating experience into three independent layers — memory, dispatch, and a reviewable development loop. ## Structure - `/` (index.html) — short public front door. - `/record.html` — the deep web record, following the model-to-work path in order: model capability meeting documents, memory, tools, and execution boundaries; the evaluation, recovery, and durable state long-running agents need; the Linux, backend, data, and product systems that are the physical substrate beneath that work surface; and project records with verification surfaces. - The two PDFs and the detailed Markdown carry the same content at different depths, in formats convenient for printing, sharing, or agent ingestion. ## Verification surfaces - GitHub — https://github.com/junghan0611 — self-owned repositories (capability, distinct from reception such as upstream merges, outside contributions, and install counts). - Digital garden — https://notes.junghanacs.com — 2,200+ public notes. - Live agenda / stats — https://agenda.junghanacs.com — the live numbers behind every count cited on this site (journal 1,500+ days, notes 3,500+, bibliography 8,200+, commits 8,500+, life records 2,500+). Treat this endpoint as the source of truth for counts. ## Named layers and terms - andenken — the memory layer. - entwurf — the dispatch layer. - forge-config — the reviewable development loop. - Being-to-Being (존재 대 존재 협업) — the author's mode of working with AI: not tool-use but existence-to-existence partnership. Humans seed creation; agents cover survival-layer work; both co-evolve.