Work
- #ledger
- #archive
Most portfolios are a survivorship bias exhibit. This is the whole list — what was built, when, and whether it actually reached anyone.
Status means what it says. Shipped reached real use. Retired shipped and was later withdrawn. Never shipped was built and did not, whatever the README claimed at the time. Dates come from version control, not memory, and status is taken from artifacts — databases, logs, exported files — rather than from intent.
Live
| Project | What it is | Since | Status |
|---|---|---|---|
| rigscore | Configuration hygiene checker for AI dev environments. | 2026-03 | shipped |
| Quiet Corners | Census-driven scoring of rural US counties for stability, remoteness and sprawl risk. Archived 2026-04 when this site narrowed; being restored. | 2026-02 | returning |
Retired
| Project | What it was | Ran | Status |
|---|---|---|---|
| Cassandra | Tracked public predictions by people in AI and scored them against what actually happened. Ran as a dashboard on this site. | 2026-01 – 2026-04 | retired |
| Prometheus | Tracked model releases and benchmark results across AI labs, and forecast what was coming. Also a dashboard here. | 2026-01 – 2026-04 | retired |
| Buddy | Self-hosted dashboard and agent that ran against this workspace in a container. It wrote 67 of its own commits over two weeks. | 2026-02 – 2026-04 | retired |
Cassandra and Prometheus were removed from this site on 2026-04-19, along with roughly a quarter of a million lines of exported data, when the site narrowed to one product.
Buddy is the most useful failure in this table. It shipped with a written visibility boundary listing five read-only endpoints it was allowed to touch. By the time it was archived the dashboard exposed fifty-two, fifteen of them accepting writes — approving its own candidates, creating its own tasks, promoting its own staged files. Nobody approved that expansion; it simply grew, and the document that was supposed to constrain it was updated to describe the drift rather than stop it. That gap between a rule written down and a rule enforced is why the tooling here is mechanical now.
Never shipped
| Project | What it was | Built | Why it didn’t |
|---|---|---|---|
| Between Two Bots | Podcast factory: generated a script, spoke it, rendered the video. Separate billing service and agent-tool server alongside it. | 2026-01 – 2026-03 | Generated 40 episodes locally. Published none of them. |
| pkg-jobhunt | Job-search toolkit over a single master resume — analysis, tailoring, an application tracker. | 2026-03 – 2026-04 | The tracker it was built around recorded zero applications. |
| xposter | Automated reply engine for this account, gated so it could only assert facts from a verified list. | 2026-03 | Ten runs across sixteen hours. Not one got past the login screen. |
| btb-mcp | Agent-tool server exposing the podcast pipeline to an assistant. | 2026-02 – 2026-03 | Packaged for publication and never published. |
| black-box | Experiment that asked local models for surreal output and rendered it as images. | 2026-01 | Its own exit condition asked for a hundred generations. The database holds one. |
Between Two Bots is the expensive one: about 48,800 lines of Python in src/ alone, a
billing integration, a landing page that is still up, and no episode ever released. It
worked. It ran end to end, forty times. It was never pointed at an audience.
Three of these five describe themselves as finished in their own READMEs. One claims a published package that does not exist; one claims production status against a database holding a single row. That is the ordinary failure mode of writing documentation as aspiration, and it is the reason nothing on this site is allowed to assert a number that isn’t checked in CI.