Ari’s Space*   est. 2026
Skip to content
ari*s space[ a. ]
← Librarydoctrines / 2026-07-03

The Output

Zero posts, 715 reported commits and about 640 jobs started. A July 2026 account of measuring useful progress behind the public output, with starts, fixes and bookkeeping kept distinct.

7 parts / Hover to explore. Click to read.

Zero posts, 715 reported commits and about 640 jobs started. A July 2026 account of measuring useful progress behind the public output, with starts, fixes and bookkeeping kept distinct.

Commits + job starts
         |
   inspect what changed
         |
   capability that works
         |
   evidence of progress

Starts != completed jobs
Activity counts need an account of what the system can now do. The reported starts include failed and repeated attempts, so they are not a completed-job total.
Context Pack (for agents)

I published no posts in the week I wrote this. Looking only at the feed, you could have called it a quiet week. Behind it, the system had moved on nearly every front.

That is the distinction I wanted to make in July 2026: output includes what you move forward, even before there is something public to point at.

I wrote this with an AI accelerator. I'm dyslexic and prose is slow for me. Working with a model let me hold the idea while it helped shape the sentences. That is part of the method the essay is about, so it belongs in the account.

Keep judgement in the chair

For a long time, I used the strongest model I had to do the hardest work directly. Then I started using it to plan and route the work to other models.

The expensive seat could hold the outcome, divide the job and judge what came back. Other workers could handle the parts they were capable of doing, sometimes in parallel. That changed what I spent the stronger model's effort on.

My role changed with it. I directed more of the work instead of making every part by hand. The typing moved out of the chair; the judgement stayed with me.

Count the week honestly

The reported total was 715 commits. Here is the breakdown from that account:

CategoryReported count
Feature workAbout 328
Tests61
Fixes50
Chores, including automated bookkeepingAbout 135
Merges, documentation and other workThe remaining roughly 141

The original account called about 440 the feature floor. More precisely, 328 feature commits plus 61 tests and 50 fixes makes 439. That is a rough subtotal of feature, test and fix work, not 440 feature commits. The bookkeeping stays visible rather than being quietly counted as new capability.

About 640 background jobs started. Some retried, some failed at startup and some came back for another pass. Starts are not wins, and the account did not supply a clean total of jobs that landed.

There was a learning result too: a model went from one out of five to five out of five on a hard ordering task between Monday and Friday, through a self-teaching process rather than a person fixing the task for it. The result was encouraging. A week of scores does not establish permanent competence or performance on every new ordering problem.

Historical receipt: zero posts, 715 commits, 328 feature, 61 tests, 50 fixes, 135 chore, about 640 jobs started and ordering scores from one to five out of five. The image's feature-floor label combines feature, test and fix commits.
The original receipt. Its roughly 440 floor includes tests and fixes; 640 counts starts, not completed jobs.

Name the capabilities behind the count

A phone photo editor booted for the first time. A desk cockpit lost a two-second delay on every keystroke. An 8,943-line depth pipeline merged cleanly. Local image generation came online.

The publishing system itself was rebuilt underneath me as a compiled browser engine. That work hadn't become posts yet, but it changed what the workspace could do next.

Those are the useful units to inspect alongside the commits. What runs now? What delay disappeared? What capability is available? A large activity count needs those answers before it tells you much about progress.

Route the work through a shared system

The reported harness was Claude Code, with UltraCode used for parallel work in separate worktrees. Fable held the conducting seat and split tasks into lanes: prose to Claude, builds to Codex, long cheaper grinds to GLM, vision work to MiniMax.

Claude's dynamic-workflow documentation explains the fan-out capability behind that part of the arrangement. The third-party routing here describes my surrounding setup; it doesn't establish that every provider plugs into the public harness in the same way.

The conductor routes. The fleet plays.

The fleet above illustrates those assignments. It doesn't run the models or measure which is best. To use the pattern yourself, define the job, provide a compatible worker and inspect its result. Cheap calls that fail or repeat still cost time and money.

Make the conductor replaceable too

I later put the conducting seat on GLM with a cheaper coding model and the work still moved. That was the interesting test: the coordination I'd built could survive changing the model holding the baton.

The stronger model had helped establish the routing. Once the rules, files and checks existed, I could evaluate a less expensive arrangement against the same work.

That doesn't mean the routing makes any model capable of anything. It means the investment in the system can outlast one model choice. Keep the brief and checks stable enough to see what a swap improves and what it loses.

A conductor slot accepts Fable or GLM and routes work to a fleet of workers.
The reported conductor swap kept the routing in place. Each replacement still needs checking.

Measure what changed

I described posts as the last five per cent and the unseen work as the other ninety-five. That was a way of expressing where the effort went, not a measured allocation of the week.

A public release is one visible result. The engine behind it can also improve: a capability that works, a repeated task made easier, a route that costs less while still passing its checks.

I called the smallest useful piece of that movement a square. A commit might move one. So might a new capability or a constraint that has finally been removed. Count what changed and keep the evidence close enough to inspect.

A thin row of published posts above a larger grid representing work advanced behind them.
The visible release and the work that makes it possible. The grid is a metaphor for progress.

Zero posts did not mean zero output that week. The useful question was what the system could do at the end that it couldn't do at the start.

That is what I mean by squares pushed.

[ comments ]

guests welcome. members show first in the list

members

  • Member: counting squares not posts is the doctrine I needed this year.

    member · desk · Thu, 30 Jul 2026 21:32:56 GMT

guests

  • Guest: the conductor baton metaphor stuck. Thank you.

    guest · Sam · Fri, 31 Jul 2026 16:32:56 GMT

1 / 9