Source sweep: the extension folder, all 105 documents, 176,548 words, re-swept 2026-08-19 at the technique-first standard, against The Hanumanic Strategy Corpus and the 264 pages already on disk.
Nine pages built from a hundred and five documents. The low yield is the phase's first result and not its failure: the build had already taken almost everything. Three research lines, and the first one is about what a re-sweep is actually for.
The line. Phase 10b was commissioned to catch missed content. It caught almost none — document after document resolved to a page already on disk, often under a name that had nothing to do with the document's title. What it did catch was structure, and structure is only visible from the whole.
The evidence. The phase's strongest finding is a count nobody could have made from inside a single document: 66 of 105 documents are composed as fixed-field entries repeated across a roster, and the corpus's two documented defects sort by column rather than by document.1 The second strongest is a one-sentence correction sitting in one document that four others needed and never received.2 The third is an arithmetic collision between two framework specifications that neither cites.3
⭐ None of the three is in a document. All three are relations between documents.
Where it goes. If a re-sweep's yield is structural, then the method should have been structural from the start — count field labels, diff the question-forms, add up the hour claims — rather than reading for missed items and finding them absent. ⚑ That is a claim about how to sweep any large corpus a second time, and this phase is the only test of it so far.
What would settle it. The vault holds several other multi-document corpora that have had one pass. Run a structural sweep on one of them — label counts, repeated phrasings, cross-document arithmetic — before reading anything, and see whether the yield is again relational.
The line. The field-reliability finding is stated for this corpus and is almost certainly not about this corpus.
The claim in one sentence: a mandatory field is a rigour instrument when an honest author lacking evidence can still complete it, and a fabrication engine when they cannot. Potential Drawbacks can be answered from imagination. Measurable Outcome cannot.1
Why it matters beyond this build. Structured sources are everywhere in this vault — practitioner manuals with fixed entry formats, astrological corpora with per-item fields, clinical and cultivation texts with recurring headings. ⚑ If the rule holds, a reader can predict which rows to verify before opening a single one.
⚠ And it points inward. This vault's own page template has fourteen mandatory headers. One of them — the two-handshake floor — was already found to manufacture insight and was amended in CLAUDE.md on 2026-08-10. The amendment was written as a fact about handshakes. It is a fact about mandatory fields, and the other headers have never been audited.4
What would settle it. A pre-registered test, and it is cheap: take a document nobody has read, predict from its field labels alone which rows will verify, then check. A prediction made before reading the cells is the only version of this claim worth keeping.
The line. The corpus attaches every protocol to a daily practice the reader already runs, and never describes that practice, because the reader is the author.5 Strip the base and the frameworks survive; the protocols do not.
Why it is now urgent rather than interesting. The base has been excluded from this vault entirely by instruction, which means the build must eventually say which of the corpus's material it can still carry. The split is currently asserted from inspection and has never been run as an audit.
What the audit is. One pass over the folder, marking each protocol portable or host-dependent. ⚑ It is a morning's work and it would tell a book what it actually has — the portable half being, on this reading, the only half a reader without the base could buy.
🏁🏁 THREAD 3 IS CLOSED, 2026-08-19. THE AUDIT WAS RUN OVER ALL 105 DOCUMENTS AND IS FILED AT WORKBENCH/Research/hanumanic-portable-host-dependent-audit.md. ⛑ The thread's own two-way split was wrong, and the audit needed four categories to say what it found: 68 portable · 5 portable with a generic base assumed · 14 separable dependency · 18 host-dependent. ⭐⭐ The category this thread lacked is the valuable one. In fourteen documents the dependency is confined to a discrete, findable, terminal Integration with… section, so the protocol comes out whole for the cost of one deletion — which is neither portable nor host-dependent and is where the corpus's best protocols actually sit. ⛔ Including the entire Ram protocol series, which this build had listed as portable and is not. ⭐ The thread's closing conjecture survives and is strengthened: the dependency really is scaffolding rather than infrastructure for 87 of 105 documents, and the corpus really is building something that can outlive the tradition that hosted it — it just does not know which of its own pages that is true of.
⭐ And the sharper question underneath it. The overlay mechanism is indifferent to its host: infusion needs only a recurring container the practitioner already keeps. A run, a commute, a nightly review would serve. ⚑⚑ If the payloads run on a secular container, the base is scaffolding rather than infrastructure — and the corpus is doing something it has never claimed, which is building a practice that can outlive the tradition that hosted it.
⚠ The saturation loop was run to zero on passes 2 through 4 — standalone worked examples, small named mechanisms, compressed quotes and contrarian asides — and returned two supporting details, both folded into existing pages as source-updates. The phase is saturated for content and open for structure, which is Thread 1 restated as a work item.
🏁 THE READING DEBT IS DISCHARGED, 2026-08-19. ⛑ A paragraph stood here saying two of the corpus's largest documents had been confirmed by content-string test rather than by full reading. It has been deleted because it is no longer true — and it understated the debt while it was true. The real outstanding figure was 56 documents totalling 106,508 words; all of them have now been read in full, minus the excluded set. ⭐ Every one resolved to a page already on disk except ten, which were built. ⚠ What the discharge produced was not items but relations — the three findings above — which is Thread 1 confirming itself on the largest sample the phase had.