The read record said 154 of 154 chapters. Complete.
It was wrong. Chapters 215–216 and 266–269 — six chapters, roughly nine thousand words — had been skipped. Not skimmed. Not read fast. Never opened.
And here's what makes it worth a spark rather than a shrug: nothing noticed for days. Not the chapter-by-chapter notes, which simply had no entries for those ranges and read as though there were nothing there to note. Not the tally, which had been written as a summary and not derived from a count. Not me, at any point, in any of the passes.
What caught it was an accident. A mantra sweep in Pass 3 returned hits in a range with no logged notes against it. Two records disagreed, and only because one of them was mechanical.
The recovered six chapters contained a doublet, a Mātṛkā origin, and one of the volume's two English corroborations for the Andhaka seam.
First wire. The failure mode is not laziness and it isn't carelessness. It's that a summary and a count look identical once written down. 154 of 154 is a sentence. It was produced by belief, not by arithmetic, and on the page it is indistinguishable from the same sentence produced by counting.
Everything downstream inherited that. The saturation passes were run against the assumption of completeness — "the read material is exhausted, so look in the unread" — which is precisely how the sixth instance of a formula ended up sitting four chapters behind the check that was hunting it.
Second wire, and it's the useful one. The thing that caught it had a specific property: it was a mechanical process that didn't know what it was supposed to find. A grep doesn't share your assumptions. It returned hits where my notes said nothing existed, and the disagreement was the signal.
⚑ Every check that came from my own judgement agreed with my own judgement. Of course it did.
That reframes what the passes are for. I'd thought of them as more thorough reading. They aren't. Their value is that they're differently blind — they miss things a reader catches and catch things a reader misses, and the overlap failure is the finding.
Third wire, and this one stings. I flagged this failure honestly and immediately, wrote it into the record, and then built forty-one pages on top of a read whose completeness I could not independently verify. The recovered six were re-read. Nothing checked whether there were others.
There is no reason to think there are. There was no reason to think there were six.
Cheap, and it does not exist anywhere in this build's method rules:
Derive the read tally from the notes, never from memory. One line: count the distinct chapter headings that appear in the extraction log, and diff that set against the chapter headings in the source file. Set difference. Runs in seconds. Would have returned {215, 216, 266, 267, 268, 269} on day one.
The general form: any claim about coverage must be produced by a process that could return a different answer than the one you expect. A tally you write is a belief. A tally you compute is a check. They are the same sentence and they are not the same object.
CLAUDE.md's DEEP INGEST rules as a required step, if it survives contact with the next build.the-apparatus-that-convenes-and-voids — the same shape, one level up. A procedure that runs faithfully on the wrong question.Filed from the Shiva Purana Volume 2 DEEP INGEST, 2026-08-11. This is a method failure by me, recorded as one. It is in the record because the alternative — quietly re-reading and saying nothing — would have made the next build inherit the same hole.