See what needs attention.
The dashboard and ORSON’s manual pilot are in use. The separate web-panel upgrade passed 145 checks and a no-write dry run; installation and device checks remain ahead.
Inside the ORSON buildPERSONAL TOOLS & EXPERIMENTS
Inside
the lab.
Small experiments, visible checks, and useful lessons for the next attempt.
I’m building ways to prepare better tasks, keep track of sources, and check AI-generated work. Some pieces are in use, others are bounded tests, and the connected workflow is still being developed.
RECORDED CHECKPOINTS
What works so far.
These summaries link to dated records.
They are not a live status display.
Keep the mistakes in the results.
A later Goose/Hermes comparison completed 32 local-model responses. Both produced caption mistakes; exact-answer checks caught failures that valid formatting would miss. One earlier synthetic-note read was separately approved and checked.
Read the tests and their limitsPrepare tasks. Keep the sources.
Prompt OS helps shape tasks and research briefs. Research OS brings sources and reviewed notes into the work. A recorded handoff stopped at a required status check; a later automated replay found seven filename mismatches.
Read the research incidentThe worker proposes code.
Separate checks decide whether a candidate can reach review. In the scripted September test, two successful attempts reached review and all seven intended negative attempts were kept from acceptance.
Read the scripted test
A SEPARATE, READ-ONLY CONNECTION
A small window for Muse.
Two approved context documents Muse. Persistent context was read back successfully. A later daily-update candidate remained unpublished, with the publisher prerequisite check still held in the latest consulted tracker.
This connection does not give Muse general file access or a way to write back into reviewed memory.
Read the Muse connection noteHOW THE PIECES FIT TOGETHER
A task, with room to say no.
This is the intended process. The complete connected flow is not installed. Manual steps, tested components, and planned work are labelled below.
- 01 / MY DECISION
I choose the task
Decide what is worth doing, what is in scope, and what a useful result should look like.
- 02 / PREPARE
Define the task and its checks
Prompt OS · Manual use; connected handoff in development
Prepare the instructions, relevant context, requirements, and evidence needed to judge the result.
- 03 / CHOOSE A PATH AS NEEDED
In testing Engineering
AEOS
AI workers propose a code candidate. Separate tooling checks the tested requirements.
Manual use · Integration in development Research
Research OS
Check reviewed memory first. Reuse suitable work, refresh it, or research the gaps, keeping the sources.
Planned Media
Media OS
Prepare storyboards, scene prompts, and references for appropriate generation tools.
- 04 / CHECK THE DRAFT
Check the result and its evidence
Use the task’s agreed checks. Code or format checks can be scripted; source quality, meaning, and creative judgement still need review. A passed check covers what that check actually tests.
↶ Failed or incomplete? Correct the draft or revisit the task. - 05 / MY DECISION
I review the result
Passing checks makes a candidate ready for review. I still decide whether it is useful and supported.
Approve → use or publishOnly the output I approve moves forward.
Revise ↶ prepare againChange the draft, evidence, or scope.
Reject → stopA result can be set aside.
How to read the labels
Manual use: I carry out or supervise that step. It does not imply an automatic connection.
In testing: the linked checks cover a bounded setup, not general reliability.
Planned: the component or connection is still a design. Media OS is planned even though I already make AI videos with individual tools.
Muse’s scoped connection is shown separately above. A local model is a possible worker inside an approved task; it does not choose its own permissions or approve its output.
THREADS I’M FOLLOWING
On the workbench.
NOTES, SOURCES & CORRECTIONS
Memory & learning.
Keep useful sources, decisions, and corrections together, so returning to a project doesn’t mean starting from scratch.
Explore the projectGIVE AN OLD PHONE WHEELS
A small rolling robot.
I’m planning a cheap wheeled build, potentially reusing my cyberdeck’s Moto G Stylus as its brain. Parts, assembly, and compatibility testing are still ahead. A walking robot is a longer-term goal.
Meet the phone behind the ideaAI VIDEO EXPERIMENTS
Generated Elsewhere.
Short AI-generated videos exploring impossible places and strange phenomena.
FIELD JACKET
One jacket, before a sample.
A visual concept for one personal jacket. Fit, construction, and a physical sample remain ahead.
Explore the conceptFollow the experiments.
The build notes keep the useful results, the failures, and the limits together.
Read the build notes