1n2 Weekly — the week the archive admitted it was wrong

The most important thing that happened this week is the kind of thing most studios would never mention. On Friday, while checking whether two freshly ingested press items could actually be found, the search behind the brain archive was caught returning the wrong documents. Not missing documents. Wrong ones.

The index was lying

The diagnosis is worth stating plainly because it is so ordinary. The full-text index had been declared contentless, which means it stores no column values and does not take its row identifiers from the table it indexes. The sync routine then tried to delete a row by an identifier that was never stored, so the delete quietly matched nothing and the insert appended another copy. Every re-sync added another. By the time anyone looked, the index held 1,441,945 rows against 679,665 real items — roughly two duplicates of everything, with the relevance arithmetic weighted accordingly. Ask it for the two most on-topic documents in the corpus and it would hand you adjacent chatter instead.

Found and fixed the same day

It was rebuilt the same afternoon as an external-content index, aligned to the real rows and maintained by triggers rather than by a routine that had to remember to be correct. The fix is written up at length, root cause first, which matters more than the patch: a corpus of a hundred and sixty thousand published item pages is only worth having if the thing that searches it tells the truth. This is a project that keeps insisting illegible material can be made legible. A broken index is the exact opposite of that, and it had been broken corpus-wide for a while without anyone noticing, because broken search does not look broken. It looks like an answer.

A homepage, narrowly

Elsewhere in the week, a books-first personal homepage went live on the third — street art photography, the Curio Cards, the films — published through its own deliberately narrow copy step rather than by being added to the main roster. A sixteen-chapter book about Curio Cards and what it means to be first finished building on the first of the month, and is sitting complete and unpublished, which is a choice rather than an oversight.

Two refusals

That choice is the week's other theme. The sweep had spotted several new directories and offered them up as roster candidates, and the audit on Sunday turned two of them down in writing. One would have had the morning job screenshot a page built from private voice memos and copy that image into the public assets directory — the leak path being the screenshot, not the directory, which is the sort of thing you only notice if you trace it. The other would have dragged along a sibling folder holding another writer's essays. Neither was added. The patch that would add the safe ones is written out and waiting for a human to say go. A roster that knows what it must not publish is doing more work than one that just grows.

The ledger, honestly

The raw file counts need the usual discounting. Seven thousand seven hundred and ninety-five files moved under brain, and all but nineteen of them are regenerated item pages and daily automated reports; exactly one file in there was touched by a person. The comics, listings, and pixel-display directories are cron doing its job. The reader's five hundred files are a build output, not five hundred decisions. The archive directory that moved most is local-only and stays that way. The hub was rescanned and now knows about two hundred and eighty-four destinations with fifty-one more waiting. Strip the automation out and the week comes down to one real bug found, one real fix shipped, one site published, and two things correctly refused. That is a good week.