Lab

What is being tested, not merely celebrated.

An open, practical record of questions, setups, results and the useful failures in between.

On the bench · November 2026

I’m trying OpenBot as an on-demand workshop: Codex prepares a piece of work, Grok gives it a second look, and I decide whether the result is worth keeping.

The published SignalTube guide review gave me a concrete result: Grok caught an uninstall sentence both Codex passes missed. I passed the review across manually and still had to check its suggestions.

Read the OpenBot experiment ↓

The experiment

  1. PrepareCodex works on a bounded task.
  2. QuestionGrok reviews the result.
  3. DecideI judge what actually helped.

On demand, not always watching. The separate context monitor is paused.

Workbench

Hands-on tests, with the wires still attached.

Real hardware, real setup and a record of what happened before the verdict is earned.

Current · Started August 2026

Workbench #001 — Portable Omarchy Workstation

Question: Can one encrypted AI/dev environment live on an external NVMe, move between the NVIDIA desktop and an ASUS ROG Zephyrus G14 with AMD Radeon RX 6700S graphics, and remain useful without configuration drama?

Working with
Omarchy 4.0.1 · encrypted Sabrent NVMe · NVIDIA RTX 4070 Ti SUPER · AMD Radeon RX 6700S · Benzinga Pro · ChatGPT Desktop · Codex CLI
Verdict
Portability passed: the same encrypted installation moved from the NVIDIA desktop to the AMD laptop and back without repair. Benzinga Squawk, normal laptop hardware and power profiles worked; the Windows 70% charging limit did not carry into Omarchy.
Next
Decide whether one bounded sustained-workload test earns a place. Before regular laptop use, verify and configure an ASUS battery charge ceiling in Linux.

Current · Started August 2026

Workbench #002 — Can Omarchy Handle a Real Workday?

Question: Can an Omarchy desktop handle an ordinary day of trading, calls and vault work without Windows becoming the rescue plan?

Working with
Omarchy · TradingView Desktop · Google Meet · Obsidian · PipeWire USB audio · GeForce NOW
Verdict
Update, 9 September: Lee is now using Omarchy full-time. Trading, writing, calls, vault work and game streaming have taken it beyond a one-day trial. The migration article is live, and Lee reports that the Windows recovery image passed verification. This is a personal daily-use result, not a claim that every Windows application works or that the system is fault-free.
Next
Keep recording real-world reliability and recovery checks. A calendar-related hang was reported after the 8 September update; its cause is unresolved. Successful image verification is not a restore test.

Current · Started September 2026

Workbench #003 — Local AI That Is Actually Local

Question: Can a local-only AI setup stay private and feel genuinely useful in ordinary work, rather than merely prove that a model will run?

Working with
Omarchy · Ollama · Open WebUI · Docker · Gemma 3 12B · NVIDIA RTX 4070 Ti SUPER · OpenSCAD
Verdict
Yes on the tested desktop: native Ollama and local Open WebUI stayed bound to the machine, while Gemma 3 12B ran fully on the RTX GPU. A controlled local check returned in 5.18 seconds cold and 1.36 seconds warm. A separate practical task felt much faster on Linux, but it used different models and runtime versions, so it is not an operating-system benchmark.
Next
Use the exact same model, digest and comparable runtime versions if a cross-platform claim becomes useful. Otherwise, move to one bounded real workload with temperatures and Git checkpoints.

Current · Started September 2026

Workbench #004 — Can Hermes Make Returning Easier?

Question: Can a persistent agent remember useful context and reuse proven procedures, so returning after an interruption takes less effort?

Working with
Hermes Desktop · Omarchy · Persistent memory · Reusable skills
Verdict
The published account records early tests on 8–9 September: fresh-session preference recall, shared context across providers including a local Qwen model, reuse of a saved health check, and an authorised, bounded run of an existing knowledge workflow. A Gemma configuration failed and was parked. These are reported results, not an independent audit or proof of long-term reliability. Remembering a boundary is not the same as enforcing it.
Next
Test whether the continuity remains useful after time away, and check that remembered instructions still reflect current decisions. Keep permission controls separate from written preferences. The article includes a small continuity-test guide using harmless sample material.

Lab status

Working now

Work in progress, with testing and limits made explicit.

Current · Started October 2026

A calmer way back into the desktop

Question: Can a small front door with one focus and one next action make returning to work less demanding?

Working with
Omarchy · Markdown · iMac
Verdict
An early Omarchy for Different Minds preview has been tried on a separate iMac test bed, using test notes rather than my personal vault. It is a step beyond a concept, but not a finished desktop. The first checks exposed rough edges in opening notes and the initial return-to-work record; creating Markdown files is not the same as a working Obsidian integration.
Next
Resolve the first-use issues and test leaving, returning and finding the next action in ordinary use. The preview remains private; no public download or wider-user validation is claimed.

Current · Started August 2026

First Steps

Question: Can a private, browser-based tool help someone turn half-noticed Parkinson's symptoms into a calmer, better-prepared clinical conversation?

Working with
Next.js · browser-only storage · printable report
Verdict
Live early experiment. It keeps entries on the device, produces a printable discussion record, and does not diagnose anything.
Next
Listen for useful friction before deciding whether it needs to become more.

Current · Started September 2026

My Linux Changes — remembering why I changed it

Question: Can a small desktop journal keep the reason for a Linux change alongside what changed, where the backup is and how to undo it?

Working with
Omarchy · Linux · Local change journal
Verdict
The local prototype is now included in the publicly listed Omarchy Briefing plugin, checked on 12 October 2026. It combines deliberately written explanations with observations of installed package and plugin changes. Detection can flag a change; it cannot know why I made it. The journal stays private, and everyday usefulness is still being assessed.
Next
Keep testing it as part of Briefing during ordinary desktop changes, including missed observations and whether the saved explanation helps after time away. Backup and undo notes are reference text, not an automatic recovery system.

Current · Started October 2026

OpenBot — a workshop, not another boss

Question: Can two AI agents improve a specific piece of work without giving me another system to manage?

Working with
OpenBot · Codex · Grok
Verdict
Published review, 10 November: after an early shared-channel desk exercise, I compared the same SignalTube install guide with Codex alone and a Codex–Grok review in OpenBot. I manually passed the second Codex answer to Grok to control what it saw. Grok caught an uninstall sentence both Codex passes missed, but also suggested a step I had to correct. One measured run showed a useful second perspective; repeatability and the benefit of an automatic handoff still need checking. The separate context monitor remains deliberately paused.
Next
Try a shared-channel handoff on one real review task and compare the checking and coordination effort. Use public material or fictional files in a disposable environment for access-boundary tests. Keep the context monitor paused.

Current · Started October 2026

WD Obsidian Vault OS — testing the return

Question: Can a guided start make a fuller vault easier to use—and easier to return to after an interruption—without asking the person to rebuild the context?

Working with
Obsidian · Markdown · Claude · Codex · Manual setup
Verdict
Public alpha update, 8 November: WD Obsidian Vault OS v0.3.1-alpha.2 is available for personal testing. An early iMac run exposed a gap in finding saved work, which led to Recent notes and After Setup guidance. The public package now has a prepared ZIP, tester guide and checksums. Automated extraction, fictional setup, pause/resume and starter-note checks passed on Windows, macOS and Linux. Real beginner use in Obsidian and AI folder connections still need feedback.
Next
Try the public alpha with fictional answers in a separate folder. Check the saved notes in Obsidian, pause halfway and return in a fresh AI session. Report your operating system, route and first confusing step, especially on Windows or macOS.

Current · Started August 2026

A permanent workshop, not another publishing treadmill

Question: Can a site make scattered work easier to understand without becoming a second newsletter to feed?

Working with
Astro · Markdown · local development server
Verdict
The workshop is live: a searchable archive, practical tools and the surrounding work in one place.
Next
Keep it useful without turning it into a platform.

Current · Started September 2026

Where was I? — a way back into the work

Question: Can a quiet bar plugin help me find the note I was working on and pick up one next step, without spending the day inside Obsidian?

Working with
Omarchy · Markdown · Local checkpoints
Verdict
Built and installed, and still in personal testing. I can browse recent Markdown notes, read them from the bar and leave a checkpoint with where I stopped and what comes next. The useful question is whether it helps when I actually lose the thread, not just whether the panel opens.
Next
Keep using it through ordinary interruptions and time away. Record what helps and what gets in the way before deciding whether it is ready for other people. Not a public release.

Back to page contents ↑

Lab status

Paused

Work set aside, without calling it finished.

Paused · Started September 2026

A community map without the headcount

Question: Can a map show where a conversation reaches without exposing readers or turning into an audience-size contest?

Working with
Substack · Country-level data · Natural Earth
Verdict
An early private prototype explored a country-level view using Substack-reported locations, without names, subscriber counts or individual pins. Shelved on 27 October: I do not expect to work on this again soon. It was not released as a live reader map.
Next
None planned. The prototype and existing screenshot are retained in case I decide to return to it.

Back to page contents ↑

Lab status

What held up

Finished experiments with a result worth keeping.

Finished · Started August 2026

Five manual note-comparison tests

Question: Can a small, read-only local workflow produce useful, evidence-backed connection suggestions without quietly taking over the vault?

Working with
n8n · Ollama · copied external bundle
Verdict
Four useful outcomes from five test sets; one refusal was the right answer
Next
Keep it manual and bounded; the remaining manual-alternative measure is not yet established.

Finished · Started September 2026

From update confusion to a published Omarchy plugin

Question: I recognised Cursor in an update list. Mise? I had to ask. Could a small bar plugin close that memory/context gap by explaining what had changed, rather than leaving me with package names to decode again?

Working with
Omarchy · Linux · Codex · Python · QML
Verdict
Built with Codex, the experiment combined explanations of completed package changes with Omarchy news. After several days of ordinary use without issues on my machine and fourteen passing automated tests, it was useful enough to release publicly: v0.1.0 under MIT. The marketplace published and automatically verified the submitted snapshot. Marketplace verification is not a security audit, and one machine’s experience is not universal compatibility evidence.
Next
The experiment is finished; the released project remains active and now includes My Linux Changes. The verification above describes the original September snapshot: on 12 October 2026 the current Briefing listing shows unverified. See the Project record for current details.

Back to page contents ↑

Lab status

Useful decisions not to build

Stopped work can still leave behind a better rule.

Abandoned · Started June 2026

The health tracker I chose not to build

Question: Would a carefully designed no-shame tracker help on the days when tracking matters most?

Working with
Obsidian · design notes
Verdict
No — an empty dashboard would become a record of absence, not support
Next
Keep the lesson: a system that demands the most on a bad day has failed its own brief.

Back to page contents ↑