mujun.dev — I build systems that contradict themselves on purpose, then listen for what survives the argument.

[ CALLSIGN :: K. ARASHI ] [ ROLE :: SYSTEMS ENGINEER / ARCHIVIST ] [ STATUS :: ACTIVE ] [ NODE :: 37.7621N / 122.4543W ]

SCROLL ↓ VERTICAL :: HORIZONTAL DRIFT LOCKED CONTRADICTION FACTOR :: 0.71

— DOCTRINE 02.A

two engines, one room, opposing directions.

I do not believe systems should agree with themselves. The clean architecture diagram is a lie a team tells the room before lunch. Real software is two engines bolted to opposite walls, each spinning the floor in a different direction; the floor stays still only because their contradiction cancels.

What I build, then, is not a tower of consistent abstractions but a held tension. A type system arguing with a runtime. A schema arguing with the data it has to swallow. A latency budget arguing with a feature flag. The job is not to resolve these arguments — the job is to instrument them so the contradiction becomes a measurable quantity.

// AXIOM 矛盾

“A system that never contradicts itself has stopped describing the world. Listen for the friction. That is where the real schema lives.”

Most of my career has been spent in the seam between things that were never supposed to touch — a payments rail meeting a real-time bidder, a cold archive meeting a live editor, a Cold War mainframe protocol meeting a 2026 GraphQL gateway. The seam is the only honest part of any system. Everything else is documentation.

— end of doctrine 02.A // continued in 04.00 //

SECTION 03.00 — WORK / DECLASSIFIED DOSSIER

Four entries selected from a longer file. Names redacted where contracts demanded it. Years are honest. Outcomes are measured in the friction they generated, not the features they shipped.

CLEARED // REV-A // 2024
FILE NO. 03.A · 2024 · ARCH CLIENT [ ██████████ / FINTECH ] SCALE 12 services · 04 regions · 41M req/d STATUS SHIPPED · STILL RUNNING
Project: HADAL — a settlement engine that disagrees with the ledger on purpose.

Designed and shipped a dual-write reconciliation core where the canonical ledger and the streaming view are deliberately permitted to disagree by up to 30 seconds. Drift becomes the alarm rather than the bug. Reconciliation drift dropped 92% because the contradiction was finally measurable.
- rust
- kafka
- postgres·logical
- nats
- opentelemetry
REDACTED // REV-B // 2023
FILE NO. 03.B · 2023 · GOV CLIENT [ ████ ████ / PUBLIC SECTOR ] SCALE 1 mainframe · 1 modern stack · 41 yrs of pain STATUS DELIVERED · UNDER NDA
Project: TRENCH-LINE — a translator between a 1983 protocol and a 2023 graph.

Wrote a long-running adapter that speaks EBCDIC on one side and protobuf-over-grpc on the other. Pages of edge cases. The contradiction: every modernization plan asked us to kill the mainframe; the mainframe was, measurably, the most reliable component in the building. We kept it. We shipped the seam.
- cobol-bridge
- go
- grpc
- ebcdic
- kerberos
CLEARED // REV-C // 2022
FILE NO. 03.C · 2022 · OSS CLIENT [ SELF / OPEN SOURCE ] SCALE ~8.4k stars · 211 contributors STATUS v2.0 · MAINTAINED
Project: ANGLER — an observability lib that emits its own self-doubt.

Open source telemetry library that, alongside metrics, emits a continuous stream of “contradiction events” — places where the trace, the metric, and the log disagree about the same operation. Operators see the disagreement before they see the outage. Adopted by 04 mid-stage startups; the failure mode it made visible is now a vendor category.
- typescript
- otel
- open-source
- v2.0
CLEARED // REV-D // 2021
FILE NO. 03.D · 2021 · MEDIA CLIENT [ ████████ / NEWSROOM ] SCALE 04 desks · 1 archive · 41 years of copy STATUS RETIRED · ARCHIVED CLEAN
Project: PALIMPSEST — a CMS that refuses to forget its previous selves.

Editorial CMS where every story carries the full lineage of its earlier drafts as a first-class object — not a diff log, but a parallel document. Editors stopped fighting auto-save and started using contradiction between drafts as the basis for headline selection. Fact-checking time fell 30%.
- elixir
- phoenix
- postgres
- cms
- archival

SECTION 04.00 — PROCESS / METHODOLOGY

Five steps. Each step contains the seed of its own contradiction. The friction between adjacent steps is where the work actually happens.

04.01
listen for the disagreement.

Before reading the spec, I read the meeting notes — especially the parts that were not minuted. The contradiction in any system is announced in the room before it is announced in code. I write down the two sentences that should not both be true. Those become the schema.
// anti-step: ignore the explicit ask
04.02
build the seam first.

I never start with the “core domain.” I start with the boundary that two teams refuse to own. That seam — the contract that nobody wants to write — is the highest-information surface in the system. Everything else is downstream of how honest that one interface is.
// anti-step: skip the unowned middle
04.03
instrument the contradiction.

I add a metric whose only job is to be > 0 when the two sides of the seam disagree. Then I make that metric a first-class deliverable. Dashboards, runbooks, paging — the disagreement becomes operationally visible before the feature does.
// anti-step: hide drift behind a retry
04.04
refuse premature resolution.

When a stakeholder asks why drift is non-zero, the temptation is to fix it. I argue, every time, that drift at 0 is the actual bug — it means the measurement is broken. A healthy seam has steady, bounded contradiction. We tune the ceiling, not the floor.
// anti-step: ship a clamp() that hides the truth
04.05
archive the friction, not the win.

The post-mortem I write is for the contradiction, not the launch. Every shipped system gets a one-page document — paper, not Confluence — describing the two opposed forces it holds in tension. That page survives the team. The launch does not.
// anti-step: write a press release

SECTION 05.00 — RESOLUTION / CONTACT

CONTACT :: 37°45'43.6"N // 122°27'15.5"W / k.arashi@mujun.dev

// one channel. plain text. monospaced replies only. expect 04.21 days latency, on average.

[ DOSSIER · 03.00 ]
[ PROCESS · 04.00 ]
[ ENTRY · 01.00 ]

mujun.dev — I build systems that contradict themselves on purpose, then listen for what survives the argument.

Project: HADAL — a settlement engine that disagrees with the ledger on purpose.

Project: TRENCH-LINE — a translator between a 1983 protocol and a 2023 graph.

Project: ANGLER — an observability lib that emits its own self-doubt.

Project: PALIMPSEST — a CMS that refuses to forget its previous selves.

listen for the disagreement.

build the seam first.

instrument the contradiction.

refuse premature resolution.

archive the friction, not the win.