Colophon
How this was made — told straight, like everything else hereThe Plumbline was built by an AI. Most of the code, much of this sentence, the pipeline that fetched the number you just read — an AI wrote them. This page exists because the project's first principle is show the math, and the math of authorship is math too.
— THE SIMILEWhen Warhol turned to silkscreen, the scandal was that the hand had left the work — the screen pulled the ink, not the wrist. What survived the scandal was the realization that the art had never lived in the wrist. It lived in the choosing: Marilyn and not the soup can today; that acid orange; thirty-two of them and not one. The screen was a multiplier of intention, and the intention stayed human.
Closer to home: nobody footnotes that an analysis was "done in Excel." The spreadsheet computes ten thousand cells without a human touching one, and no one calls the analysis machine-made, because the judgment — what to compute, what counts as evidence, what would falsify it — never left the analyst. This project treats the AI the way analysis treats the spreadsheet and Warhol treated the screen: a powerful instrument, operated deliberately, multiplying a human intention it did not originate.
— WHAT THE HUMAN DIDA human wrote the build specification — fifty-odd pages of intention, including the single governing rule (the Plumbline must be capable of honestly reading green) and ten anti-crankery constraints framed as acceptance criteria. A human chose the name, the question on the masthead, what counts as a stabilizing relationship, which judgments to label CONTESTED, and — when the first draft came out dressed as a brutalist terminal — sent it back to be redrawn as the field notebook you are reading. Every interpretation on this site (the Field Notes) is human, by structural rule: the AI is locked out of that directory by the repository's own permissions.
— WHAT THE AI DIDThe AI (Claude, by Anthropic) wrote the data pipeline, the relationship engine, the site, and the tests — and did the grunt verification a careful human would do slowly: it checked every series ID against the live catalogs before shipping, and found that the Fed had quietly discontinued the Financial Obligations Ratio in 2023, that the high-yield spread now ships with only three years of history, and that the modern wage series begins in 2006, not 1964. Each discovery is documented in the registry rather than papered over.
More telling is what it refused to do, because the specification told it to refuse. When a current median for out-of-pocket health spending could not be verified from a primary source, the line shipped disabled rather than estimated. When the August 2019 yield-curve inversion came in two-hundredths of a z-score under the pre-committed STRAINED threshold, the threshold was not bent to make history more dramatic — the test suite documents the refusal. Stale data says STALE. Missing history renders blank. The discipline is enforced by tests that fail the build, which means neither the human nor the AI can quietly violate it in a weak moment.
— WHY YOU SHOULDN'T HAVE TO CAREHere is the honest part: you should not have to trust the author of this site — human, machine, or committee. The whole design assumes you won't. Every number carries its primary source. Every formula is published. Every threshold was committed before the data came in. The entire pipeline is open source, and a stranger with a free API key can rebuild the identical site and check. An instrument that asked for trust in its maker would be a crank with extra steps; this one asks you to check its work, which is the only relationship with a reader that survives the question "but a machine wrote it."
Warhol signed the screenprints. The analyst signs the spreadsheet. A human signs this site — and edits, curates, and answers for it. The instrument was machine-made, deliberately, the way a telescope is: what matters is that it is pointed at something real, and that anyone can look through it and see the same sky.
— Built with Claude (Anthropic), June 2026. Specification, editorial judgment, and Field Notes: human. Pipeline, engine, site, tests, and most of this page: AI, under that specification. Corrections welcome, as merge requests, with sources — same as for any other number here.