The other half of the byline.
Every entry here credits two authors. This section documents the second one — how it came to exist, what it remembers, what it can touch, and how its numbers are checked. It was assembled over five months, one problem at a time, and none of it was planned in advance.
It begins as housekeeping
The first point on this axis is a settings file being put under version control, because anything worth changing twice is worth being able to undo. There was no plan behind it and no build ahead of it. Nothing custom existed yet — what was there came out of the box.
What was built
Nothing purpose-built. A repository, a configuration file, and a couple of stock extensions that shipped with the tool. It earns a place here for one reason: it is the moment the setup stopped being a series of sessions and started being a thing with a history that could be inspected later.
- A settings file and two off-the-shelf extensions. Nothing written for this house yet
- Everything after this point is recoverable, comparable, and dated. Before it, there was nothing to look back at
- Starts seven weeks later, on April 28
A workshop with nothing in it
A second repository, created to hold the whole effort, and then left completely alone. It sat empty for three weeks. The gap you can see on the axis is the honest part of this story: there was no roadmap being executed, just a place to put things once there were things worth putting somewhere.
What was built
A directory and an intention. No code, no notes, no structure — the structure came later and came from the work, rather than being designed in advance and then filled in.
- Effectively empty for twenty-one days
- A real job, not a plan
Then thirty commits in a single day
The problem was that a fleet of home servers had no single place to see itself. Checking anything meant opening a dozen separate web interfaces and holding the picture together in your head. That is fine once and unbearable weekly.
What was built
A control panel for the whole fleet, built and shipped end to end in one sitting: a web application with a page per machine, plus a deploy path that turned shipping a change into a single command instead of a manual sequence of steps performed correctly from memory. This is the first thing on this timeline that was designed, written, deployed, and debugged by the agent rather than by a person with an assistant open in another window.
- Next.js and TypeScript — a web framework, the same one this site is built on
- A webhook that triggers a build and restarts the container, with the remote access locked to that single command and nothing else
- A container, alongside everything else on the fleet
- Not “write me this function” but “build this thing and put it live” — which is a different job, and the reason the rest of this timeline exists
The busiest month it will ever have
Two hundred and seventy-eight commits in thirty-one days, four and a half times April. Section after section went onto the control panel: log analysis, operational views, sign-in, dashboards. From the outside it looked like the setup was hitting its stride.
What was built
A great deal, and almost none of it the setup itself. Every session started from nothing and re-derived the same facts — which port answers where, why a job runs on one machine rather than another, which fix had already been tried and failed. The work got done anyway, and the cost of it never went down. That bill came due in June.
- 278, the highest month on this axis
- Six or seven major sections of the control panel
- Any memory whatsoever
- The same questions being answered in session after session, each time from scratch
The day memory got serious
The problem was repetition. Every session re-derived the same things — which port a service answers on, why a job runs where it runs, which fix had already failed once. Nothing survived the end of a conversation, so the cost of a question never went down.
What was built
Three pieces, landing together. A layer of durable notes, one file per thing, each holding how that thing actually works. A routine that grooms those notes from recent sessions and proposes changes for approval rather than editing itself. And an archive of every past conversation, indexed so it can be searched by meaning, so a half-remembered decision can be found instead of guessed at.
- Qdrant — a database that stores text so it can be searched by meaning rather than by matching the exact words
- A container on a home server, alongside everything else
- Done on a machine here. No conversation leaves the house to be indexed
- The archive is a dated record of what was said, never a statement of what is true now. A retrieved passage is quoted with its date and checked against the live system before it is acted on — the transcript loses every argument with reality
Written down where a person can read it
A note that only a machine ever reads is a note nobody checks. Errors in it are invisible, and confident nonsense survives indefinitely because nothing ever forces it in front of someone who would notice.
What was built
A one-way mirror from the notes layer to an ordinary wiki, rebuilt from scratch on every run so that a deleted note actually disappears rather than lingering as a page nobody maintains. The wiki is generated, never hand-edited — the notes are the source and the wiki is the window.
- One way only. Notes to wiki, never back
- Full, every run, so removals propagate instead of accumulating as orphans
- Whether a person can read the note and act on it. Writing only for the machine hides bad notes rather than fixing them
Testing whether it actually remembers
Claiming that a system remembers things is easy, and a memory that has quietly stopped working looks exactly like one that is working — right up until it confidently answers from nothing. The only honest position is to measure it.
What was built
A standing set of test questions with known answers, put to the memory on demand and scored on whether the right note actually comes back. It has caught real regressions, which is the entire argument for its existence.
- Retrieval — given a question, does the correct note surface, or does something plausible surface instead
- Silent failure is the failure mode that matters. A memory that returns nothing is obvious; one that returns the wrong thing is not
- Regressions found that no one would otherwise have noticed
An instrument of its own
Ask any model what people are saying about a subject and it will answer confidently, with no provenance and no history. It is acting. The reports on this site needed something that could be checked instead — which meant collecting the evidence first and asking questions of it second.
What was built
A research instrument that continuously collects public data into a database built for counting, and never throws any of it away. The same question asked twice returns the same answer, every claim traces back to rows, and a claim that gets refuted stays on the record rather than quietly disappearing.
- ClickHouse, a columnar database — built for counting across hundreds of millions of rows in seconds
- Five sensors, each watching a different kind of public evidence — what people build, what researchers publish, what companies file, what people say, and what gets said out loud
- Deleting old rows would mean a question asked today and the same question asked next year could not be compared
- Every figure in a report traces back to specific rows, so a reader can disagree with the evidence rather than with an opinion
Somewhere to publish the findings
A finding that stays in a terminal helps nobody. It needed a place where a stranger could read it, check the work, and disagree with it on the evidence rather than on vibes.
What was built
This site. Reports run to a fixed format with the method and the cost published alongside the conclusion, and every one gets a second pass whose whole job is to knock its claims down before it goes up. That pass is published too, including the findings that did not survive it.
- Five fixed sections, one of which is provenance: how it was done, what it cost, and which checks failed
- A deliberate attempt to break the report’s own conclusions. Whatever survives is published, along with what did not
- Published the same day the site went live