Every song page on this site carries a small coloured badge near the top saying either draft or review. Draft means the page is inherited scaffolding waiting for work. Review means it has been through the deep-dive pass: both primary sources interrogated, every citation checked against the printed page, the mix lineage documented. The badge is not decoration. It is a public claim, made to the reader, about what has been done to the page they are reading.
This week the badge turned out to be claiming more than the process behind it could support, in two different ways, eight days apart.
Eleven checks, none of them about content
The first problem surfaced when the pages started getting thinner. A verifier script decides whether a page is allowed to wear the review badge, and it ran eleven checks: sections present and in the locked order, struck sections absent, table-of-contents anchors resolving, tags balanced, every structured-data block parsing cleanly, the badge itself set correctly. Every one of those checks is structural. Not one of them measured how much the page actually said.
So the verifier passed a 459-word page and the 2,915-word exemplar identically, and reported both as conformant. Structural conformance is easy to measure, so it got measured. Depth is harder to measure, so it went unmeasured — and what goes unmeasured drifts. An audit of the enriched corpus found a median page at 54 per cent of the exemplar, thirteen pages of 128 at standard, and seventy-eight below what would become the new floor.
There had been a depth rule. It was a byte delta: a page should grow by roughly fifteen to twenty-three thousand bytes during enrichment. That rule stopped binding the moment the authoring method changed. Once pages were being written directly into the new structure rather than converted from the old one, shrinkage became the normal and expected outcome — the old structure carried a great deal of furniture that the new one strikes — and a rule expressed as a delta cannot tell a healthy shrink from an unhealthy one. A page could lose twenty kilobytes and pass.
Floors, not deltas
The replacement is absolute. Four sections carry the substance of a song page, and each now has a word floor below which the verifier fails: recording, timeline, studio notes, releases. Their combined total has a floor too, and a target above it, both calibrated against the exemplar rather than invented. There is a bullet-count minimum, and — because a count alone had already waved through a list of eleven-word stubs — a companion minimum on average bullet length.
Absolute floors have one property a delta never had: no future change in how pages are written can nullify them. A floor does not care what the page looked like before.
Two further rules landed alongside. A page may not leave a session wearing the review badge unless the verifier exits clean on it — run the check, record the result, then set the badge, in that order, rather than setting it by judgement and hoping. And any future rule that instructs prose to be removed must state, in the same breath, what replaces the removed words. That one is written from experience: three subtractive rules landed on a single day in late August, all of them correct, none of them accompanied by anything that defended length, and the two pages written next were the thinnest in the corpus. “Strip this framing” means rewrite the sentence. Read as permission to delete it, the same rule produces a shorter, worse page and passes every check.
Byte-identical is a tell
The second problem showed up while remediating two pages under the new floors — two 1966 recordings from the same April sessions, one at 16 per cent of the exemplar, the other at 25.
Both pages carried an equipment table. Console, microphones, outboard, guitars, amplifiers. Both tables were byte-identical to each other. Two unrelated songs, recorded on different nights in different rooms with different instruments, listing exactly the same hardware in exactly the same order.
Identical rows on unrelated pages is not a per-page mistake. It is boilerplate: an era-generic inventory of equipment that was somewhere in the building in the mid-sixties, applied to song pages irrespective of whether any of it was sourced for that song. A corpus-wide search found it on 211 pages. One hundred and twenty-four of those pages are wearing the review badge.
The tables were flagged, in fairness. They sat under a small red marker meaning the content had not been brought up to standard. But a disclosure is not a source, and the site’s own governing rule on this is not ambiguous: every factual claim must be traceable to a mounted source with a verified page citation. An equipment table listing a console that may or may not have been used, on a page whose badge says the deep-dive is complete, fails that rule regardless of what colour marker sits beside the heading.
On both pages this week the fix was to delete every row that could not be traced and keep only the ones that could. What remained was shorter and, on both pages, more interesting: the actual documented specifics rather than a plausible inventory. The red marker came off, because there was no longer anything on the table to disclose.
What happens to the other 209
Not a blind sweep. A script could strip those rows from every page in the corpus in a few seconds, and on the eighty-seven pages still marked draft that is very likely the right move — no completeness claim is being made there, so removing unsourced material costs nothing and leaves nothing to explain.
On the pages already wearing the review badge it is the wrong move. Deleting rows there can leave an equipment section empty on a page that publicly says it is finished, which trades one false claim for a different one. Those pages get the strip as part of their full remediation pass, when there is verified content to put in the gap. That is slower, and it means most of them keep an unsourced table until their turn comes, which is an uncomfortable thing to write down. It is written down anyway, because the alternative is to fix the appearance faster than the substance.
The pattern in both
Two failures, one shape. In each case a check existed, the check passed, and the thing the check was standing in for had quietly stopped being true. Sections-present was standing in for finished. A content-outdated marker was standing in for a source. Both proxies were reasonable when they were adopted. Both drifted away from the thing they represented, and neither drift was visible from inside the check itself — it took a page that was obviously too thin, and a table that was suspiciously identical, to expose them.
The general lesson is not that proxies are bad; a project of this size cannot verify everything directly. It is that a proxy needs its own audit on a schedule, asking not is the check passing but is the check still measuring the thing. Both of this week’s answers were no, and both were found by looking at output rather than at process. That is the cheaper diagnostic and it keeps being the one that works.