AI-assisted solo engineer studying how unreliable code generation can enter reliable software systems without quietly acquiring authority.
Generating code, obtaining a green check, and establishing an engineering fact are different events. My projects began with five dependency-free HTML pages, grew through complete business systems and distributed-system exercises, reached real single-machine stop lines, and eventually separated into three laboratories with distinct authority boundaries.
“Student” describes my current identity and age, not a project maturity level. Project maturity is stated through evidence, release state, and explicit limitations.
Start here: Project journey · Three laboratories · Selected work · Engineering method
Cognitive branch: AI increases the rate of cognitive variation; it does not guarantee cognitive progress. Selection quality, retained counterexamples, and explicit ownership of goals, evidence, vetoes, and revisions still determine what survives. The current long-form edition is available in the Chinese essay library.
The philosophers of 1965 competed for forks. Today's agents also interpret goals, choose actions, report results, retry, delegate, and consume one another's claims.
Read -> a report is not a fact
Delete -> ability is not authorization
Update -> a local observation is not the complete fact
Create -> a correct end state does not make the execution path lawful
Traditional operating systems are good at deciding whether a concrete process may perform a concrete operation. I am interested in the additional boundary created by an executor that interprets natural-language goals: why its selected action remains inside the user's authorization, when old execution rights expire, and what qualifies its report to become system fact.
For now, I test these boundaries above existing operating systems. A new kernel primitive becomes worth discussing only when retained counterexamples show that a property cannot be established correctly in user space. Read The KPI Philosophers.
Exact milestones, frozen baselines, and active tasks belong to each project repository. This Profile records stable project roles and routes; it does not duplicate high-frequency project state.
This is not a product matrix invented after the fact, and it is not a single integrated runtime chain. Some questions continued forward; others split into independent systems with their own fact ownership.
- InkNarratives preserves five dependency-free HTML works used to learn narrative structure, typography, interaction, and public presentation.
- DarkRoomLibrary turned pages into the first complete business system and introduced role, data, collaboration, and delivery boundaries.
- Qixu grew from campus-space needs exposed by DarkRoomLibrary, but became an independent system. DarkRoomLibrary owns accounts and library facts; Qixu owns spaces, eligibility, allocation results, reservations, usage rights, and governance facts. Their explicit identity adapter is not SSO, does not inherit upstream roles, and does not merge databases.
- PlainJournal became a training ground for distributed behavior, reliability, degradation, multiple instances, and real acceptance. It also made a 16 GiB single-machine boundary—and the refusal to claim what was not proven—impossible to ignore.
- PlainJournalPro preserves the future multi-merchant, platform-ledger, and cross-machine problem. It is a planned architecture, not an implemented product.
- Those stop lines exposed a separate question: AI can help produce code, but it cannot use its own output to prove source, test, environment, and release facts. That question became VeriTrail.
- Asking who owns computation and who owns system capabilities then separated the problem into JPyxis and FlowKernel.
Question: What did this run actually prove, and is the retained evidence sufficient for the sealed claim?
Boundary: VeriTrail provides a local-first Core, Workbench, Entry surface, and GitHub Evidence path. It does not own source-system facts, world truth, or the human's final disposition.
Question: Who defines computation, who authorizes an invocation, who executes it, and who explains lifecycle and failure?
Boundary: JPyxis has a reproducible single-node heterogeneous-compute baseline. It does not thereby acquire host resource authority or business truth.
Question: How can an unreliable agent, model, or rule act inside revocable, attributable, and observable capability and resource boundaries?
Boundary: FlowKernel is a planned operating-system-level trust and execution substrate. Implementation has not started. A C-first target and a Linux reference lab remain proposed experimental vehicles, not existing cross-platform capability.
Each laboratory can close its own problem independently. When versioned contracts eventually connect them, their combination may produce capabilities none of them owns alone. Composition does not merge authority.
- PlainJournal — distributed business and reliability laboratory; a continuing source of capacity, evidence, and runtime questions.
- VeriTrail — local system for controlled variables, immutable evidence, browser observation, and deterministic bounded verdicts.
- JPyxis — contract-driven heterogeneous-compute framework separating Control, definition frontends, and runtimes.
- FlowKernel — planned OS-level trust and execution substrate for bounded agentic authority, resources, revocation, and recovery.
- InkNarratives — five zero-dependency HTML works and a unified visual gallery.
- DarkRoomLibrary — the first complete Spring Boot and Vue business baseline.
- Qixu — campus-space reservation and usage-right management, with explicit ownership of eligibility, allocation, waiting lists, short reservations, venues, and governance.
- MiniSpringBoot — reconstruction of IoC, AOP, Web/MVC, configuration, JDBC, transactions, and startup mechanics, challenged through a real React and MySQL path.
- MiniLinux — a bootable C kernel laboratory for memory, CPU, user boundaries, files, init, shell, live LAB, and replay; not a Linux-compatible implementation.
All project cards still route to their laboratory repositories. Language separation does not authorize a distribution route. A card will move to Engineering Gallery only after that project is released, has a usable Chinese edition, is cataloged, and becomes PROFILE_ROUTABLE.
The journey diagram shows historical causality and feedback. It does not turn the repositories into a mandatory deployment chain.
FlowKernel may eventually constrain operating-system capabilities and resources. JPyxis owns heterogeneous-compute contracts and execution. Evidence adapters observe bounded source facts. VeriTrail judges existing evidence against a sealed Plan. Humans own premises, sealing, and final disposition. Reality owns truth.
GitHub owns and exposes facts inside its platform trust domain; it is not an oracle for the external world. Review Attention proposes where a person should inspect first; it is not a verdict engine.
Read the full Repository System Map.
The current project repositories are laboratories: source, construction history, retained failures, evidence, and milestone coordinates remain together.
Engineering Gallery is a separate English-first distribution surface for clean, bounded release projections. It currently contains a reference showroom but no cataloged project exhibits.
Laboratory repository
engineering truth + history + evidence
|
| exact qualified projection
v
Engineering Gallery
clean release + Quick Start + limitations
|
| independently revisioned language view
v
NoctilumeDev-ZH
Chinese explanation bound to an explicit source revision
Lab optimizes for engineering truth. Gallery optimizes for public consumption. The Chinese edition optimizes for a separate language reader. Ownership is decoupled; provenance remains connected.
My current working loop is intentionally smaller than any one project's gate system:
bind the current fact coordinate and the smallest plan
-> actively search for counterexamples to the premise
-> execute the smallest authorized change
-> keep these states separate:
plan defined
!= execution complete
!= qualification established
!= state effective
!= next step authorized
-> let new evidence rewrite the next plan
Gate strength follows project risk and contract. VeriTrail's heavier qualification chain is not mechanically copied into every repository.
The full Chinese method library remains available in the Chinese edition. English articles are published selectively rather than through forced directory symmetry.
Available in English:
- Engineering Judgment
- Decision Epistemology
- The KPI Philosophers: When Executors Begin to Interpret Goals
- Protecting Zero: From Generated Answers to Qualified Facts
- Adversarial Engineering Validation: Making “Done” Survive Its Author
- AI’s Ceiling Is Not in the Answer
- Repository System Map
The complete current essay and engineering-method library is available in NoctilumeDev-ZH. The English collection is a curated publication surface, not a complete translation mirror. Existing English-repository article URLs remain as thin forwarding pages where link continuity matters.
NoctilumeDev
= canonical English Profile
NoctilumeDev-ZH
= provenance-bound Chinese edition
EngineeringGallery
= canonical distributable project surface
project laboratories
= exact engineering state, evidence, and history
Language migration is not project-routing authorization.