Kept rather than deleted, per the standing rule: a broken finding and its obituary are worth as much as a finding that held, because deleting it removes the evidence that the method was ever tested.
Phase 1 of this ingest wrote a generative-tail forecast — twelve collisions and fourteen sparks, predicted from extraction notes before a single concept page existed. Phase 2 then built two hundred and thirty-odd pages against the same source.
Two of those predictions turn out to be wrong, and they are wrong in two different ways.
[FALSIFIED] — T4: "deflation asymmetry along a school boundary, not a magnitude boundary."
The prediction was that the corpus demotes Hindu and Advaitic attainment reports while leaving Buddhist ones alone — a sectarian tell dressed as a criterion.
The pages killed it. On one page of the fourth-dhyāna chapter, Ramakrishna's five samādhis are demoted to chi-channel phenomena on a stated rule — vivid bodily sensation disqualifies a state from being an absorption — and then Buddhism's own five meditative blisses get exactly the same treatment on the same page, with the same criterion and almost the same words. Swami Rama's stopped heart goes the same way elsewhere.
So the knife does not check the passport. The boundary is phenomenological, not sectarian: anything reported in vivid bodily or affective terms gets demoted, whoever reported it.
What survives the obituary, and it is the better finding. A criterion applied evenly across schools is exactly what gives it authority — and it raises a question the sectarian version could never have reached: if vivid sensation disqualifies a state from being an absorption, how much of the world's mystical literature survives the rule, and does anything survive it except the author's own scheme? One genuine outlier remains and it is not about schools at all: an illness read backwards as a verdict on a person.
[FALSIFIED] — E63: "the five elements are the five wisdoms obscured — an identity Bodri prints twice and never links."
This one is worse, because it failed at the level of the referent. The corpus has no "five wisdoms." It has four — Great Mirror, Universal Equality, Subtle Discerning, Accomplishing Action — at a single page, arriving late and once. A search of the full source for the phrase returns nothing. There is no identity printed twice, because one of the two things being identified does not occur.
The prediction was assembled from extraction notes, and somewhere between the note and the forecast a four became a five and an adjacency became an identity.
The two failures are not the same species, and the difference is the useful part.
T4 failed by being too specific in the wrong dimension. The observation underneath it — the corpus demotes reports — was true. The forecast added a boundary the evidence never supported, and the added boundary was the one that made the finding sound like an accusation. Sharpening a claim toward the more damning reading is a reliable way to break it.
E63 failed by never having a referent. No amount of reading would have confirmed it, because the thing it described is not in the book. It is the failure mode the standing rules name directly: a claim about a source, generated from notes rather than from the source, and carried forward because it sounded like the kind of thing the source would do.
The general shape: a forecast made from notes is a claim about your notes. It gets treated as a claim about the source by everyone downstream, including the person who wrote it, and the misfiling is invisible until somebody goes back to the pages.
Against the value of Phase-1 forecasting itself. Twenty-six predictions were written before any page existed. Of those checked so far, most were absorbed into concept pages during Phase 2 — which means the forecast was right about where the tensions were, and its main function turned out to be a to-do list rather than a set of findings. Two were wrong, and both wrong ones had reached the PRD in the confident register.
Promotion criterion: count the whole forecast. If the hit rate is high, the practice is sound and the register is the only problem — forecasts should be written as questions rather than as claims. If it is low, Phase-1 tail prediction is costing more than it returns and should be replaced by a tension log with no forecasting layer at all.
Next action: audit the remaining §10 predictions against the built pages and record each as absorbed, live, or [FALSIFIED]. Roughly half are already done; the register question above cannot be answered without the count.