Corrections

C-001 · Two papers merged into one reference

Miscitation · 2026-07-21

Date: 2026-07-21 Location: research/spectrum-of-collapse/ONS_SpectrumOfCollapse_Thesis.md:324 Found by: The_Three_Rivals.md §VII.1 (adversarial stress test)

Was:

Jacobs, H. I. L., Becker, J. A., Kwong, K., et al. (2021). The mechanistic link between selective vulnerability of the locus coeruleus and neurodegeneration in Alzheimer's disease. Acta Neuropathologica, 141, 631–650.

Now: Matchett, B. J., Grinberg, L. T., Theofilas, P., & Murray, M. E. (2021), Acta Neuropathologica 141(5), 631–650. PMID 33427939.

Why: The title, journal, volume and page range belong to Matchett et al. Jacobs is not an author. The Jacobs 2021 paper is a separate work — Science Translational Medicine 13(612):eabj2511, PMID 34550726. Two distinct papers had been merged into a single reference.

Aggravating detail: the entry is cited nowhere in the thesis body. It was an orphan reference, which is likely why the error survived — nothing pointed at it to check.

Artifacts: ☑ markdown ☑ PDF rebuilt ☑ copied to public/pdfs ☑ wiki republished — shipped in ACD 4701f953

Was briefly blocked on P-009 (attribution footer). Resolved: the stamp was reconstructed and re-applied, and the footer is verified present in the live file alongside the correction.

C-002 · A dopamine-neuron measurement attributed to the locus coeruleus

Transfer · 2026-07-21

Date: 2026-07-21 Location: research/first-principles/PhD_Thesis_Genetic_Architecture_Neurodegeneration.md:196, restated at :238 Found by: The_Three_Rivals.md §VII.2

Was:

…locus coeruleus neurons — whose 4-meter cumulative axonal arbor (Schwarz and Luo, 2015) and high cumulative NAD⁺ demand make them the most vulnerable single neuronal population in the brain.

and at line 238: "cumulative axonal arbors of several meters."

Now: the long-axon vulnerability is retained and re-grounded on Braak's actual criterion — long, thin, unmyelinated or sparsely myelinated projection neurons (Braak, Rüb, Schultz & Del Tredici, 2006, PMID 16914843). The metre-scale figure is removed, and the text now states that no published estimate of total axonal length per human locus coeruleus neuron exists.

Why: the 4-metre figure is Pissadaki & Bolam's estimate for human substantia nigra dopamine neurons (Front Comput Neurosci 2013;7:13, PMID 23515615). It is not a locus coeruleus measurement, and the quantified arbor-burden argument has no LC equivalent in the literature.

Class note: this is a fabricated-by-transfer citation — a real number, from a real paper, about a different cell type, attributed to a third paper that does not report it. It sat beneath the most-repeated mechanistic justification for Phase I. The same failure mode is being swept for corpus-wide (see P-002).

Artifacts: ☑ markdown ☑ PDF rebuilt ☑ copied to public/pdfs ☑ wiki republished — shipped in ACD 4701f953

This paper was never part of the 81-file attribution stamp, so the clean rebuild is safe to publish.

C-003 · In-text reference with no bibliography entry

Dangling citation · 2026-07-21

Date: 2026-07-21 Location: research/first-principles/PhD_Thesis_Genetic_Architecture_Neurodegeneration.md:196 Found by: verification pass while correcting C-002

Was: (Schwarz and Luo, 2015) cited in body text, with no corresponding entry anywhere in §12 References.

Now: the citation is removed with C-002. Correct entries added for Braak et al. 2006 and Pissadaki & Bolam 2013, both of which the revised text does rely on.

Why: a citation that resolves to nothing cannot be checked by a reader, which is how C-002 escaped notice. Note the corpus does cite Schwarz & Luo correctly elsewhere — collapse-trilogy/PhD_Thesis_Vago_Coerulean_Relay.md:278 gives Nature 524(7563):88–92 — so the error was local, not systemic.

Artifacts: ☑ markdown ☑ PDF rebuilt ☑ copied to public/pdfs ☑ wiki republished — shipped in ACD 4701f953

C-004 · Two PDF builders point at a directory that no longer exists

Build integrity · 2026-07-21

Date: 2026-07-21 Location: research/collapse-trilogy/homeostatic-microglial/build_fischer_pdf.py:13-14 and build_schwartz_pdf.py:13-14 Found by: pipeline survey during remediation planning

Was: both scripts carried hardcoded absolute paths to OskarFischerPrize/Homeostatic Microglial Collapse/, a directory that no longer exists. Both would fail on run. build_fischer_pdf.py additionally carried a docstring describing the Schwartz paper.

Now: paths resolved relative to the script's own directory, matching the house pattern (HERE = Path(__file__).parent). Docstring corrected.

Why: these two papers could not be regenerated from their markdown at all. Any correction to Fischers_Matrix_Paper.md or Schwartz_Peripheral_Arm_Paper.md would have been unshippable — the exact failure this register exists to prevent.

Artifacts: ☑ script fixed ☑ PDF rebuilt (both run clean) n/a copied n/a wiki

Neither paper appears in acd-pdf-map.json; both are unpublished, so there is nothing downstream to update.

C-005 · Attribution footer stripped from a published PDF

Regression caused and reverted · 2026-07-21

Date: 2026-07-21 Location: AdultCognitiveDisease/public/pdfs/unified_spectrum-of-collapse_intrinsic-extrinsic-axis.pdf Found by: byte-size comparison during the verification step of C-001

What happened: the corrected Spectrum of Collapse PDF was rebuilt from markdown and copied over the published file, per the remediation plan. The published file dropped from 189,801 to 126,808 bytes — a 33% reduction for a one-line citation fix. Investigation showed the published copy carried a two-line attribution footer ("Drafted by Claude-Opus-4.6 / Prepared for Dr. James Truchard and Benjamin Gustafsson") applied by a post-processing step, and the clean rebuild did not.

Resolution: the published file was restored from git to its committed, stamped state. Page count and footer verified present. The citation correction remains unpublished pending P-009.

Why it is recorded: it was caused by this remediation, caught by the register's own verification step, and reverted within minutes — but had the byte-size check been skipped, 81 published papers would have quietly lost a named person's attribution. Recorded rather than quietly undone.

C-006 · The attribution stamp is reproducible again

Tooling recovered · 2026-07-21

Date: 2026-07-21 Location: AdultCognitiveDisease/scripts/restamp-pdf.py (new) Found by: investigating C-005

The problem: commit f8870260 (29 April 2026) stamped a three-line attribution footer onto 81 published PDFs "via reportlab+pypdf overlay." That script was never committed — the commit contains PDFs and nothing else. Unstamped originals survive in public/pdfs-original/ (81 files, gitignored). Consequence: rebuilding any of those 81 papers stripped a named person's attribution, silently, with no check in the pipeline that would notice.

The fix: rather than reconstruct the original's per-document data table (drafting model and scientist list vary per paper), restamp-pdf.py reads the stamp off the currently-published file and reproduces it verbatim on the new build. No table to maintain, and nothing to get wrong per document.

Recovered overlay spec, for the record: Helvetica 7.5 pt, fill 0.35 grey, x = 36, lines stacked upward from y = 18 in 9.5 pt steps, applied to every page.

Validation: round-trip against ground truth — re-stamping the unstamped April original reproduced the live published file with zero text differences across all 22 pages. The corrected build was then diffed against the live file at word-multiset level to confirm the only change was the intended reference.

Usage, and the guard that goes with it:

# is this paper stamped?  a hit means DO NOT copy a clean rebuild over it
ls public/pdfs-original/ | grep <published-name>

python3 scripts/restamp-pdf.py NEW_BUILD.pdf public/pdfs/<published>.pdf OUT.pdf

The script refuses to run — and writes nothing — if it cannot find a stamp in the reference file, so it cannot silently produce an unstamped output.

After publishing a stamped paper, also refresh its unstamped twin in public/pdfs-original/, or a future rollback restores the pre-correction content.

Artifacts: ☑ script written ☑ validated ☑ installed ☑ committed — ACD 4701f953

C-012 · One term, six invented expansions

Fabrication, systemic · 2026-07-22

Date: 2026-07-22 Location: six documents (listed below) Found by: full-corpus sweep, prompted by an audit of how PANTHOS is phase-assigned

The actual coinage: PANTHOS — Greek, poisonous anthos [flower] — named by Lee, Nixon and colleagues for the flower-like rosettes into which undegraded cargo arranges itself in the swollen, de-acidified neuron. Nat Neurosci 25(6):688–701, PMID 35654956.

The six wrong glosses:

Aggravating detail — #2 was not merely a bad expansion. The passage had built a fictitious entity around it, describing PANTHOS as an endosomal protease performing "PANTHOS-mediated cleavage" — an enzyme rather than a morphology, with its own hypothesised substrate specificity. The surrounding argument (retromer dysfunction traps βCTF, lengthening substrate residence time and shifting the endosomal proteolytic landscape toward promiscuous cleavage) is sound and stands; only the invented enzyme was removed. Rewriting the expansion alone would have left a fabricated protease in a PhD thesis.

Why this is filed as systemic rather than as six typos. One term acquired six independent, plausible-sounding expansions across six documents, none checked against the paper that coined it. Three of the six are in the root-level submission documents. This is the signature of a term propagated by paraphrase rather than by citation — a failure mode a corpus assembled across many generation passes should expect and should test for by default. The_Order_of_Discovery.md:170 had caught two of the six and estimated "two further divergent glosses"; the real count was three times its estimate. A dated correction has been appended there rather than editing the historical reconstruction in place.

Recommended standing check: any coined term carried by more than one document should have its definition grepped corpus-wide and diffed against the primary source before publication. Six variants survived because no check existed that would compare them to each other.

Artifacts: ☑ markdown (6 files) ☐ PDFs rebuilt ☐ copied to public/pdfs ☐ wiki republished — blocked on P-008 and C-005

Stale artifacts still carrying fabrications: wiki-site/dist/theses/collapse-trilogy/ons-synapticcollapse-thesis.html, ons-bioenergeticcollapse-thesis.html, phd-thesis-viral-trajectory-three-phases.html, the dist-acd twin of the last, and the four sibling PDFs.

C-013 · Twenty phase schemes reduced to one

Frame reconciliation · 2026-07-22

Date: 2026-07-22 Location: corpus-wide; specification at research/temporal-architecture/PHASE_FRAME_SPEC.md (new) Found by: audit prompted by a misfiled PANTHOS phase assignment on the live site

The defect: twenty independent definitions of the temporal architecture, disagreeing on phase names, count (3 vs 4), age boundaries (7 distinct sets), Braak mapping (3 sets), and whether the bridges exist at all. The live React site named Phase I "Synaptic Collapse" — the name the theses reserve for Phase III — and rendered it on the home page above brainstem prose.

Root cause, dated: commit 3d8c51a (28 Apr 2026) populated src/constants.ts and synthesis-report-2026-04-28-tripartite.md with Collapse Trilogy volume titles in volume order against phase slots in temporal order. The trilogy was written outcome-first because the programme discovered the disease in roughly the reverse of the order it occurs (The_Order_of_Discovery.md:20); the architecture runs time-first. The two orderings are exact mirrors, so the paste inverted Phases I and III. Canonical phase names were not coined until 21 May. constants.ts never caught up, and every artifact reading PHASES inherited the inversion for three months.

Proof it was a paste and not a decision: within the same object, title: 'Synaptic Collapse' sat above locus: 'Locus coeruleus…', driver: 'NAD⁺ depletion via PARP-1', drugClass: 'PARP inhibitors' and scoreColumn: 'bioenergetic_score'. Every field but the title was Phase I.

Resolutions adopted (full argument and evidence in the spec):

1. Five stations, not three phases. Ratifies what ONS_TemporalArchitecture_Thesis.md:62 already says and what three dashboards already render. Four phases rejected — Spectrum's Phase IV has no position on its own axis, no death-modality table, the same mechanism as Phase III, and its central claim is contradicted at ONS_CellularArchitecture_Thesis_SecondEdition.md:301.
2. The count is a staging convention, not a natural joint — already the position of The_Two_Anchors.md:137. No derivation of "three" from biology exists; the count came from three drug-class papers.
3. [substrate]: [fate] naming, adopted from ONS_OFP_Entry.md. The prior scheme named Phase I for a process, II for a cell, III for a casualty — three conventions, no scheme.
4. Driver and casualty are separate columns. This dissolves the apparent Spectrum-vs-canon conflict over Phase II: immunological driver, neuronal casualty. Spectrum's death-modality tables are promoted to the casualty column, not retired.
5. Bridges are first-class, with a mandatory evidence caveat — elevating them elevates the corpus's least-evidenced layer, and the complement arm of the Proteolytic Turn is contradicted in human tissue.

Applied so far: src/types.ts (new casualty, measurability, Bridge), src/constants.ts (phases rewritten, BRIDGES added), src/components/PhasePage.tsx (driver/casualty pair, measurability band), src/App.tsx (five-station strip; PANTHOS supersession note rewritten). Build passes.

Outstanding: paper-inventory.json phase strings; the five kb/wiki/temporal-stages redirect strings; the thesis-level deviations in the spec's §VII register; the three conflicting Braak mappings; dashboard label/geometry mismatches; the canonical thesis subtitle and site hero, both of which assert the count and are held for an editorial decision.

Artifacts: ☑ spec written ☑ src/ applied ☑ build verified ☐ markdown corpus ☐ wiki republished ☐ PDFs — partially applied; blocked on P-008 and C-005 for the publish half

C-014 · The corpus's central mechanism was filed in the wrong decade

Placement audit · 2026-07-22

Date: 2026-07-22 Location: research/three-phase-stress-test/The_Late_Flower.md (new, 30 pp) — the argued brief; ruling at PHASE_FRAME_SPEC.md §IV / R9 Found by: user challenge to the PANTHOS phase assignment, escalating from C-013

The defect: the corpus states that PANTHOS is "the visible endpoint of Phase I" and "the terminal morphology of Phase I autophagic failure" — a brainstem event of decades 3–5. Five falsifiers were fixed in advance against that proposition. All five are hits.

The therapeutic consequence, which is the reason this is not bookkeeping. The corpus closes the autophagy-flux-induction window at age 52 (autophagic-collapse.src.html:310) while drawing the PANTHOS lane to 85 and marking cortical PANTHOS at 79; Spectrum states the quality-control window "closed in Phase I–II." The window is drawn shut 27 years before the bulk of the casualties. The corpus has independently written the manoeuvre that fits — "protecting acidification… the manoeuvre most specific to the late disease" — and left it unconnected. Nixon's own LMRS sequence (re-acidify, then induce) is filed under "currentLocations": ["phase-i"] and is a late-stage protocol. And no autophagy or lysosomal drug class appears at any phase in src/constants.ts.

Root cause, and the general lesson. The network was charted outward from Nixon; PANTHOS entered at Nixon's coordinate in the ordering and inherited the filing. The question in which cells does this occur was never asked separately from where does Nixon sit. The_Map_That_Became_a_Clock.md confesses the path dependence in general terms; this is the specific instance. Standing remedy proposed: for every imported mechanism, ask where it occurs and how we know, independently of who told us about it. §XIII item 5 assigns the same audit to the mechanisms imported from Gouras, Frost, Area-Gomez and Abbate.

Failure mode. Distinct from the fabrication C-012 found and the interpretation-drift The Denominator found: correct facts, correctly cited, assembled into a claim none of them supports. No PMID check catches it. This is the argument for placement audits as a standing genre.

Artifacts: ☑ markdown ☑ PDF built (30 pp) ☐ copied to public/pdfs ☐ wiki republished — blocked on P-008 and C-005

C-015 · PANTHOS withdrawn from Phase I entirely

Placement audit, escalated · 2026-07-23

Date: 2026-07-23 Location: ruling at PHASE_FRAME_SPEC.md §IV / §V.7 / R9; propagated corpus-wide (see Artifacts) Found by: user direction — "we never intended to claim that PANTHOS occurs in Phase 1" — escalating C-014

What changed from C-014. C-014 established that the Phase I placement was wrong in emphasis and settled on "first reached in Phase I; bulk at the terminus" — retaining a Phase I priority claim (spec R9, and the closing sentence of The_Late_Flower.md). That priority claim is now withdrawn. The programme's position is that PANTHOS does not occur in Phase I at all — not as the terminal morphology, not as "first-reached," not as a hedged or predicted instance. This adopts the stronger of the two readings The Late Flower itself left open (§VI L-reading 1, "the corpus should stop calling it PANTHOS"; §X recommendation 4, "the assignment goes").

The corrected position. The distinction is machinery vs morphology. Phase I keeps the autophagy–lysosomal machinery failure — the architecture's shared spine — and its intrinsic death mode is Parthanatos (PARP-1/AIF), on a background of proteostatic gridlock. PANTHOS is the terminal morphology of high-APP, high-Aβ glutamatergic pyramidal neurons: it first appears in the Phase II casualties (entorhinal layer II, CA1) and reaches its bulk at the terminus (neocortical pyramids). The low-Aβ, noradrenergic locus coeruleus is not a PANTHOS-competent cell, does not die until Braak III (Theofilas 2017), and leaves no plaque "gravestone." §V.7 of the spec, formerly an open gravestone contradiction, is now marked resolved.

Scope of what was preserved, not changed. Parthanatos as a Phase I death modality; the autophagy-as-spine claim; every statement correctly placing PANTHOS in pyramidal/CA1/entorhinal/neocortical cells; keyword/scoring records that merely note a scientist studies PANTHOS; and the dated audit records themselves — The_Late_Flower.md (a resolution note was added at its head; its argued body is unaltered) and the April 2026 synthesis reports.

Artifacts: ☑ spec (PHASE_FRAME_SPEC.md §IV/§V.7/R9 + driver-casualty table) ☑ flagship thesis (ONS_TemporalArchitecture_Thesis.md:106,228) ☑ mechanistic_pathway.html ☑ src/constants.ts + src/App.tsx ☑ public/paper-inventory.json ☑ wiki apps (autophagic-collapse, collapse-cascades, map-of-collapse, census .src.html) ☑ companion theses (clearance-collapse, cellular-architecture ×2, down-syndrome/PhD_Thesis_Compressed_Architecture) ☑ kb note (inside-out-plaque.md) ☑ The_Late_Flower.md resolution note ☐ wiki republished ☐ PDFs rebuilt — markdown/source complete; publish half pending (npm run wiki:publish; museum-PDF rebuilds for the edited theses)

Failure mode. Same genre as C-014 — correct facts assembled into an unsupported claim — but this entry records the completion of the retreat C-014 began: a corpus that had corrected the emphasis but kept the priority claim was still, per Rule 6, laundering a retreat as a refinement. The claim is now fully withdrawn rather than softened.

Not caused here, but noted

public/pdfs/microglial_bioenergetic_nlrp3-trem2-substrate.pdf and microglial_homeostatic_tgf-beta-synthesis.pdf show as modified-uncommitted in the ACD repo. Both are dated 13 July 2026, eight days before this work, and were not touched by it. Left alone; flagged so they are not mistaken for collateral damage.

Part II — Identified, not yet corrected

Scheduled against three-phase-stress-test/REMEDIATION_PLAN_2026-07-21.md.

P-001 → C-007 · Overclaim retired · The "vast arbor" claim, 18 sources

Date: 2026-07-22 · Shipped: ACD 3bdb5be4

Was: the corpus derived the locus coeruleus's vulnerability from the size of its axonal arbor — "vast arbor," "prodigiously arborising," "among the most extensive of any neuron in the brain," "4-metre," "several meters," and (found during the sweep, not in the original audit) "total axonal length per neuron approaching one meter."

Now: re-grounded on Braak's actual criterion — long, thin, unmyelinated or sparsely myelinated axons (Braak et al. 2006) — and on pacemaking plus catecholamine oxidative chemistry (Sanchez-Padilla/Surmeier 2014), the mechanism actually demonstrated in this cell. The vulnerability conclusion is unchanged; only its derivation moved.

Preserved deliberately: the arbor argument for substantia nigra dopamine and nucleus basalis cholinergic neurons, where it is quantified and sound; the parvalbumin basket cell's connectivity claim; and the Cellular Architecture four-feature phenotype, read disjunctively ("an extensive axonal arbor or an outsized connectivity load"). Verified by inspection that all four surviving matches are these protected cases.

Also fixed under this entry: the LC neuron count (P-002), which disagreed with itself by 2× across the corpus. The figures are per side — ~20,000 as TH-positive neuromelanin-bearing cells (Manaye 1995), ~50,000 counted inclusively (Theofilas 2017). Propagated from the house formulation already in the Cellular Architecture census.

Artifacts: ☑ 18 sources ☑ 16 PDFs rebuilt ☑ 3 restamped ☑ 14 published ☑ wiki republished (0 broken links)

Caveat carried forward: Su et al. 2026 is a bioRxiv preprint and mouse. It licenses retiring an overclaim; it does not license asserting the LC is sparsely branched, and no corpus text now does.

C-009 · A third dead-path generator

Build integrity · 2026-07-22

Date: 2026-07-22 · Location: research/therapeutic-landscape/generate_therapeutic_pdf.py:8020

Same failure as C-004: hardcoded output path to OskarFischerPrize/Spectrum of Collapse/, a directory removed months ago. The generator had been unrunnable, and the document has no markdown source — it is built entirely from string literals — so it could not have been corrected at all. Repointed to Path(__file__).parent, rebuilt (250 pp), restamped, published.

Artifacts: ☑ script ☑ PDF rebuilt ☑ restamped ☑ published ☑ wiki

C-010 · Cellular Architecture edition mismatch

Mapping fault caught mid-sweep · 2026-07-22

Date: 2026-07-22

The predicted P-008 trap fired. acd-pdf-map.json maps the first-edition Cellular Architecture markdown to the second-edition PDF filename, so the automated sweep rebuilt the first edition and published it under first-principles_cellular-architecture_second-edition.pdf — replacing 50 pages of second-edition content with 26 pages of first-edition content.

Caught by the page-count and byte-size verification step, before commit. The corrected second edition is restored as the served file (50 pp, 203,005 bytes against 203,047 committed — the delta is this sweep's text edit).

Resolved 2026-07-24. The wiki now renders the second edition. Root cause: the two edition files (ONS_CellularArchitecture_Thesis.md and ..._SecondEdition.md) share a title, and AdultCognitiveDisease/scripts/sync-wiki.mjs matches Library entries to pages by title first (:118), so the single Library entry kept binding to the first-edition page on every publish — and every inbound link resolves to the first-edition slug, so the entry could not simply be repointed without breaking links. Fixed by superseding the canonical source with the second-edition content: ONS_CellularArchitecture_Thesis.md (and its research PDF) now carry the 50-page second edition, so the canonical page and slug render it beside the second-edition download. ..._SecondEdition.md is retained as a (now redundant) withheld duplicate to preserve inbound prose citations to it. Verified: the canonical page carries the Gabitto 2024 two-epoch data; 786 pages, 0 broken links, 128 papers shelved.

Artifacts: ☑ markdown (canonical file superseded) ☑ wiki republished — shipped in ACD e1e4ea45. OFP source committed locally per routine, not pushed.

C-011 · The architecture's coverage fraction, and which disease it describes

Scope stated · 2026-07-22

Date: 2026-07-22 Location: research/temporal-architecture/ONS_TemporalArchitecture_Thesis.md (abstract, §I Thesis, §II new scope note); wiki-site/build.mjs (About page); new kb/wiki/concepts/coverage-fraction.md Found by: The_Denominator.md §IX, Demotion 1 — outstanding since 2026-07-21

Was: the Temporal Architecture made unqualified claims about "Alzheimer's disease" throughout, and no document in the corpus stated what fraction of dementia the arc describes. The published About page described the corpus without distinguishing the wide frame (adult cognitive disease) from the narrow one used for mechanism.

Now: three insertions state the scope. The abstract closes on it; §I fixes the claim to Alzheimer's biologically defined — amyloid-positive, tau-positive, of typical topography, what the routing analysis calls Route 1 — and not dementia as encountered; §II gains "A note on which disease this describes," paralleling the measurability note added under C-008. The honest figure is 3–22 per cent of dementia, most plausibly 10–15, and the note carries the three consequences: entry is near-universal (abnormal tau in 99.6 per cent of adults; PART dementia rate 1.4 per cent), exit is not guaranteed by traversal (12–22 per cent of Braak V–VI are non-demented), and the excluded majority are routes, not attrition.

Why it was not a simple number. "Pure AD" is not a fixed biological quantity — it falls monotonically as more pathologies are scored (30 per cent on a 3-pathology panel, 9 per cent on a 9-panel, 3.13 per cent on a full panel). The filters covary and come from cohorts of different mean age, so they must not be multiplied. The new guardrail concept coverage-fraction.md records the table, the sources, and the reason the range is a range — the fourth such note, after phase-measurability, lc-morphology-what-is-established and plaque-origin-what-is-established.

Noted while writing, not fixed here. The entry problem's candidate answer is already in the corpus and unconnected to it: The Denominator says "the corpus has no account of what converts an ignition into a progression," while The Case for the Cascade §II states amyloid as the permissive trigger licensing tau's escape from the medial temporal lobe — and the difference between PART and AD in the series The Denominator itself cites is amyloid. The two documents have never been read against each other. Logged as P-010.

Artifacts: ☑ markdown ☑ PDF rebuilt (26 pp, 137,384 → 139,975 bytes) ☑ copied to public/pdfs ☑ wiki republished (781 pages, 0 broken links)

Verification: page count unchanged at 26; attribution footer parity confirmed before publish (12 × AdultCognitiveDisease, 1 × Truchard, 1 × Gustafsson) — the C-005 regression did not recur. The P-008 mapping trap did not fire here: the live PDF was byte-identical to the pre-edit build, confirming the map entry is correct for this paper. A first grep of public/wiki/*.html appeared to show the correction missing; the pages are in subdirectories and the artifacts were present. The faulty check was the check, not the build — recorded because a stale-artifact false alarm is as costly as a stale artifact.

Still outstanding from Demotion 1's family: the subtype and co-pathology literature remains absent from the corpus (Three Rivals §V.5) — zero occurrences of co-pathology, LATE-NC, hippocampal-sparing outside the stress-test documents. The scope note now names those routes but does not describe them. Remediation Stream D.

P-002 · LC neuron count disagrees with itself by 2×

Inconsistency

Most theses state "approximately 50,000 neurons in the human brain"; collapse-trilogy/bioenergetic/papers/PARP_Locus_Coeruleus_Phase_I.md:29 states "30,000 to 50,000 neurons per hemisphere." These cannot both be right.

One document already handles it correctly and is the house template — cellular-architecture/ONS_CellularArchitecture_Thesis.md:91: "each locus coeruleus holds on the order of twenty thousand neurons in young adulthood (Manaye et al., 1995); counted inclusively, nearer fifty thousand (Theofilas et al., 2017)." Two counts, two sources, stated method-dependence. Propagate rather than invent.

P-008 · Pre-existing mapping faults

Publication integrity

Not errors of claim, but they will silently defeat corrections:

• wiki-site/acd-pdf-map.json:57 maps the first-edition cellular-architecture markdown to the second-edition PDF. The wiki renders first-edition text under a second-edition download; ..._SecondEdition.md appears nowhere in the map.
• PhD_Thesis_Mitochondrial_Dysfunction_Neurodegeneration.md is published but has no builder — its PDF cannot be regenerated from source.
• The_Two_Anchors.md — the canonical framework — is unpublished. It has no LIBRARY_ITEMS entry, so corrections to it reach no reader.
• The_Map_That_Became_a_Clock.md carries no marker that it is superseded. The supersession exists only in its successor's reference list.

P-010 · The corpus's largest open question has a written candidate answer

Unconnected documents

Raised while shipping C-011. Two documents pose and answer the same question and appear never to have been read against each other:

• three-phase-stress-test/The_Denominator.md:288 — "the corpus has no account of what converts an ignition into a progression. This is the programme's largest open question and it is currently unstated."
• amyloid-steelman/PhD_Brief_Case_For_The_Cascade.md:60 — amyloid as "the permissive trigger that licenses the spread of tau… In amyloid-negative individuals, tau tends to remain confined; in amyloid-positive individuals, it spreads."

The word amyloid appears once in The Denominator's 342 lines, despite the brief assembling exactly the PART evidence the permissive-trigger argument runs on — and PART is defined by tau without amyloid.

Compounding it: the steelman's §VI forward-references The Temporal Architecture of Collapse and The Unified Architecture of Collapse as where the corpus's counter-argument lives. Neither contains it. The timing paradox is asserted across at least five theses and argued in none. The pointer resolves to nothing.

This is not a citation error; it is a structural gap. Scheduled as its own document rather than a patch — the rebuttal the steelman promises, which must either answer the permissive-trigger reading or concede it.

C-016 · Chronic excitatory insufficiency rendered with its sign inverted

Frame reconciliation · 2026-07-23

Date: 2026-07-23 Location: research/collapse-trilogy/convergent-synaptic/ONS_SynapticCollapse_Thesis.md §4 ("The Chronic Excitatory Insufficiency Hypothesis"), subsection Seizures and Hyperexcitability Found by: reading the primary source — Moosmann & Sohre, Chronic excitatory insufficiency as proximate cause of Alzheimer's disease (OFP 2020 submission #145) — against the corpus's rendering of it; confirmed and refined by a five-cluster primary-literature fact-check of the paper (2026-07-23).

The defect. CEI is a hypothesis about a network tipped toward inhibition: "excitatory insufficiency (meaning too much inhibition)." Moosmann's account of the AD seizure excess (5–15% vs <1% in controls) is explicitly not disinhibition — it is "a compensatory over-amplification of small excitatory signals in an overly inhibitory background" (the silent/absence-seizure model), with Aβ as the sensitizer doing the amplifying. He derives it this way precisely to pre-empt the disinhibition misreading. The corpus's §4 rendering explains the same seizure excess the opposite way: neurons "undergo compensatory upregulation of excitability — increased intrinsic excitability, reduced GABAergic inhibition," and "the loss of inhibitory GABAergic interneurons in AD … removes the normal inhibitory constraint, exacerbating seizure susceptibility despite the underlying excitatory insufficiency." That is a disinhibition mechanism attributed to CEI — Moosmann's sign, inverted, on the one phenomenon he uses to guard against exactly this error. (The §4 risk-factor table and its "primary deficit / secondary compensation" framing are faithful; the fault is local to the seizure subsection.)

What is not the error. The corpus's terminal disinhibition model — PV⁺ net digestion → loss of inhibition → pyramidal hyperexcitation → excitotoxic death (terminal-collapse Phase 6; ketamine-thesis §2.1; ofp-entry:79) — is a separate and correctly-signed claim about the terminal endgame. It is not CEI and is not being corrected. The fault is welding that mechanism onto Moosmann's hypothesis, which runs the other way.

The compounding issue (unreconciled, not merely miswritten). One datum — AD epileptiform activity — is claimed by both models at once: the disinhibition reading (terminal-collapse:82) and the over-inhibition reading (CEI). No document notes that two opposite-signed mechanisms compete for the same observation. The honest reconciliation is that they describe different epochs and different axes: CEI owns the compensated, decades-long prodrome (net under-excited); disinhibition owns the decompensated terminus (locally over-excited); and "insufficient throughput" and "pathological hyperexcitability" are distinct variables that coexist in the late brain (globally hypometabolic and locally seizing). Cross-link node written: kb/wiki/concepts/chronic-excitatory-insufficiency.md.

Fact-check refinement (2026-07-23). The multi-source check changed the direction of the fix. Moosmann's stated mechanism — over-amplification of sparse excitatory signals in an over-inhibited background — is faithfully sourced (Klaassen 2006; Cope 2009) but is the minority reading of AD seizures; the disinhibition/hyperexcitability account the corpus already uses (PV⁺/interneuron loss → disinhibition; Palop & Mucke 2016; Verret 2012, PMID 22541439, where restoring inhibition rescues) is the consensus. So the corpus flipped Moosmann's sign — but toward the better-supported answer. This is therefore a fidelity correction (the corpus misattributes its own consensus-aligned mechanism to a hypothesis that argues the opposite), not a case of the corpus holding a wrong mechanism. The broader fact-check verdict: CEI is citation-honest at the finding level but its load-bearing inversions (Aβ/tau as protective) are built by selective citation and contradicted by intervention direction (lecanemab/donanemab slow decline; low-nM human-derived Aβ dimers are synaptotoxic — Shankar 2008); the underlying preprint is unpeer-reviewed and uncited. Cross-link node with the full graded verdict: kb/wiki/concepts/chronic-excitatory-insufficiency.md.

Required source edit. In §4's Seizures and Hyperexcitability subsection: (a) stop attributing a disinhibition mechanism to CEI — state that Moosmann's actual account is over-amplification in an over-inhibited background (net toward inhibition); (b) do not adopt that mechanism as correct — flag it as the minority reading; (c) keep the "loss of GABAergic interneurons / removed inhibitory constraint" clause, but relocate it to the corpus's own terminal-disinhibition account (consensus-aligned), not to CEI; (d) add a one-line note that CEI and the disinhibition model claim the same epileptiform-activity datum with opposite-signed mechanisms, resolved by epoch.

Register correction (2026-08-04). This entry read "☐ markdown (source edit pending)" until today. It was stale: inspection of ONS_SynapticCollapse_Thesis.md §4 finds all four required changes present — Moosmann's actual mechanism stated with its Klaassen/Cope precedents, flagged as "a genuine but minority reading," the interneuron-loss clause relocated to the corpus's own account with Verret/Palop & Mucke/Bakker attached, and the closing sentence recording that the two accounts claim the same datum and are resolved by epoch. The §4 chapter head also carries the graded verdict. The markdown box is ticked accordingly. Recorded rather than silently amended, because a register that reports work as outstanding when it is done is failing in the same direction as one that reports work as done when it is not — both make the true state unknowable without re-reading every source.

Escalation — the resolution was asserted, not argued. The (d) clause commits the corpus to a reconciliation "by epoch" in a single sentence. Nowhere did the corpus argue it. That is the same structural fault as P-010 (a claim asserted across theses and argued in none), on a different joint. Discharged 2026-08-04 by a standalone volume rather than a longer footnote: research/excitation-inhibition/PhD_Thesis_Low_Signal_High_Noise.md (Low Signal, High Noise, 40 pp) — which adjudicates CEI against terminal disinhibition on the direction of therapeutic benefit, separates throughput from excitability as two variables, argues the epoch table, and supersedes §4 as the corpus's account of the excitatory–inhibitory question. §4 remains as the framework-by-framework survey entry it was written to be.

Artifacts: ☑ markdown (verified present, not merely specified) ☑ successor volume written and built (40 pp) ☐ ONS_SynapticCollapse_Thesis PDF rebuilt ☐ successor copied to public/pdfs ☐ wiki republished — publish half outstanding (blocked on P-008 with the rest of the backlog)

C-017 · The OFP submission carried errata Item 1 for eighteen days

Fidelity · 2026-08-11

Date: 2026-08-11 Location: research/ofp-entry/ONS_OFP_Entry.md (six passages) and ONS_OFP_Entry_Abstract.md (one) Found by: a blind grader in OFP Pass 2, grading the programme's own submission under Charter Rule 4 with no access to CORPUS_ERRATA_2026-07.md. Verified independently against PubMed, then traced to the existing errata.

The defect. The entry cited de Vries et al. 2025 (Alzheimers Dement 21(2):e14504, PMID 39737731) as "the decisive evidence" that resilient brains show "intact perineuronal nets," and built the §VII unity claim on the "joint preservation of all three layers." The paper reports the opposite for the matrix layer: aggrecan decreased in resilient donors (p = .005), peridendritic complexity reduced (p < .0001), WFA⁺ net density lower in the resilient group specifically — below both control and demented (p = .0005) — tenascin-R never measured, and synaptic contacts a null. The authors' own framing is remodelling, explicitly not preservation.

This is not the stale-artifact failure mode. The entry was last edited 22 July; CORPUS_ERRATA_2026-07.md was compiled 24 July and names this file at lines 59 and 111. Nothing was fixed and then shipped stale — the remediation was specified and never actioned on this document. Eighteen days.

Why it outranked an ordinary miscitation. Line 85 carried the §VII unity claim, which the entry itself calls its deepest and which requires all three layers preserved together. For one of the three, the cited paper reports the reverse in exactly the group the argument depends on. The anchor did not merely fail to support the claim; it pointed against it.

The fix. Applied the errata's own canonical replacement wording and the corrected framework form developed in research/containment-fifth-surface/PhD_Thesis_Fifth_Surface.md §IX: resilience restated as a weighted sum crossed at a threshold rather than a list of intact layers, distinguishing regulated remodelling without the proteolytic signature from digestion with it. Seven passages changed — the resilience evidence (§ Phase II), the unity claim, the resilience prediction, the therapy prediction, the convergence target, the closing statement, and the abstract. The surviving microglial result (matrix-proteolytic programme elevated in demented and not in resilient) is retained, because it is what the paper actually supports.

Independent convergence, recorded as a result about the method. The corpus's own July sweep and an August blind grader, working from opposite directions and unaware of each other, landed on the same sentence. Both procedures work. What failed is the handoff between finding and fixing — which is the argument for tracking remediation per artifact rather than per finding.

Still open. Errata Item 1 names ~30 further source files. Spot-checks on 2026-08-11 find unified-architecture (238, 242), temporal-architecture (140, 202) and kb/wiki/concepts/perineuronal-nets.md (36) still carrying the preservation claim; collapse-trilogy/homeostatic-microglial (121) appears corrected. This entry corrects the OFP submission only. The framework-wide sweep remains outstanding and is not closed by this record.

Artifacts: ☑ markdown (7 passages, both files) ☑ PDFs rebuilt — ONS_OFP_Entry.pdf 88,825 → 90,231 B, _Abstract.pdf 59,993 → 60,072 B, _Bibliography.pdf byte-identical at 54,156 as expected (C-005 size guard applied) ☑ corrected wording verified present in the rendered PDF, old claim absent — n/a public/pdfs (this document is not published there) — n/a wiki republish (not in acd-pdf-map.json or acd-shelves.json)

Update, 2026-08-11 — C-017 was itself shipped incomplete, and the same procedure caught it.

The original C-017 fix corrected seven passages of prose and rebuilt three of the four PDFs. build_ofp_entry_pdf.py does not build the figures; build_ofp_figures_pdf.py does, and it was not run. Figure 7 therefore continued to assert "Resilience = all three layers intact" and "remove any one → the threshold falls", directly contradicting the §VII text that had just been corrected to a weighted sum. Its caption and title carried the old claim too.

This is precisely the failure mode this register exists to catch — "the programme's documented failure mode is not making errors, it is fixing the text and shipping the stale artifact" — committed while correcting an instance of it, by the process running the correction.

How it was found. Not by the person who made it. A blind second grader, reading the submission package with no knowledge of C-017, reported that "the submitted package self-contradicts — Figure 7 asserts resilience requires all three layers intact, which §VII of the paper explicitly repudiates." The correction was seventeen minutes old.

A second incompleteness, caught on verification. The first figure rebuild fixed only the inner panel labels; the figure title ("Resilience is triple preservation") and the caption ("requires the joint preservation of all three layers") still carried the claim. Found by extracting text from the rebuilt PDF and reading it, rather than trusting that the rebuild had done what was intended. Fixed in a second pass.

Now verified across the whole package. All four PDFs return zero matches for "all three layers intact", "intact perineuronal", "joint preservation" and "triple preservation". ONS_OFP_Entry_Figures.pdf 76,593 → 76,762 B; the other three unchanged from the C-017 rebuild.

Two lessons recorded rather than absorbed.

1. A correction's artifact list must be derived from the builders, not from memory. Four PDFs, two build scripts, and the fix ran one of them. The four-box artifact status in this register is exactly the right instrument and it was filled in without checking which script produces which file.
2. Verifying a rebuild means reading the output, not observing that the build succeeded. The first rebuild reported success and produced a document still carrying the error in two of three places.

Consequence for the grading pass. E000's two independent gradings are not comparable: the first read the uncorrected entry, the second read the corrected text against the stale figure. Neither read the fully corrected package. This is recorded in the findings log and E000 is re-graded rather than averaged.

C-018 · The prize entrants were re-graded, and Pass 1's circularity is now quantified

Superseded · 2026-08-11

Date: 2026-08-11 Location: kb/wiki/researchers/*.md (150 notes, new ofp2 block), kb/wiki/meta/_ofp2_rankings.md, kb/wiki/meta/_ofp2_divergence.md, research/ofp-regrade/ Found by: the pass itself, by design.

What was wrong. The April 2026 scoring gave 30% of its CSC composite to max_node + mean_node — each entry's overlap with the programme's own convergence nodes. Entries were scored by how much they resembled the framework doing the scoring; the exercise returned zero contradictions; and the absence of contradiction was later cited as a reason not to prune the corpus. The_Order_of_Discovery.md established this from dates. This pass measures it.

The measurement. Pass 1 CSC relevancy correlates with max convergence-node score at r = +0.88 — the score was almost entirely an overlap measure, not partly one. And the two Pass 1 rubrics, presented as independent lenses, correlate with each other at r = +0.92: CSC and the Ten Key Questions were one instrument read twice, so their agreement was never corroboration. Pass 2 retains r = +0.50 with node overlap, which is the honest residue.

What replaces it. A rubric frozen before the first entry was read (kb/tools/ofp2_rubric.json), governed by a five-rule charter (kb/tools/OFP2_CHARTER.md) barring framework overlap from every grade and making agreement with the architecture inadmissible as evidence of quality. One classification gate, seven axes, two mandatory prose fields written before the verdict. All 150 entries read end to end by two independent blind assessors; reported score is the mean, ±3.2. Programme grades assessed separately against the 2020–2026 literature.

Consequences for the corpus.

• Thirteen entries contradict the architecture. Pass 1 found zero. Among them is the highest-scoring submission in the field.
• Not one of the 2022 jury's ten winners reaches Pass 2's top thirteen. The jury nonetheless selected well above chance — its winners sit 11 points above the untiered field, which exceeds the instrument's resolution.
• Route 7 corroborated externally. An entrant's 2026 paper finds 68–76% of cognitive variance unexplained by every pathology measured — same direction as The Denominator, larger magnitude, unconnected source.
• The CEI hypothesis is dormant, not retracted, its author having published nothing on it since 2020. Relevant to C-016.
• The programme's own entry was graded blind in the same stream and placed separately, because it was written in 2026 against 2020 submissions and no axis captures that advantage.

Pass 1 is superseded, not deleted. csc_scores and tkq_scores remain unchanged in every note per Rule 5, with the frozen baseline at kb/tools/ofp1_scores_frozen.json.

Known defects in the new rubric, published rather than patched. independence admits four incompatible readings and needs rewriting; fit bundles distance with candour; maturity has no quality floor at M3; resolution cannot separate absent from vague; reachability returned one anchor for 105 of 150 entries and is excluded from any discrimination claim. Full register in kb/tools/OFP2_FINDINGS_LOG.md.

Process failures in this pass, recorded under Rule 6. The §6 resolution ruling landed after 40 entries were already graded (re-checked; 13 of 26 moved, all downward). The first variance figure was computed on a stratified sample and reported as representative (9.1 vs the true 5.2). The first Phase 2 consolidation silently returned 42 of 150 records because the merge handled only one of two file shapes. C-017 was shipped with its figures unrebuilt. All four are the same species — a step that fails quietly and returns something resembling an answer.

Not done. Phase 3 (95 external scientists) was skipped by decision. Errata Item 1's framework-wide sweep remains open.

Artifacts: ☑ 150 researcher notes carry the ofp2 block (YAML verified, Pass 1 data byte-identical) ☑ leaderboard and divergence pages written ☑ catalog rebuilt (150 graded) ☑ lint 0 errors 0 warnings ☑ paper written (research/ofp-regrade/The_Second_Reading.md) ☑ corpus-facing notes (_internal/INTEGRATION_NOTES.md) ☑ PDF built (The_Second_Reading.pdf, 38 pp, Appendix C added 2026-08-12) ☐ wiki republished

C-019 · Two sulfated polymers rendered as one lattice

Chemistry · 2026-08-13

Date: 2026-08-13 Location: research/reelin-pnn/PhD_Thesis_Architects_Scaffold.md — title, abstract, §IV (three passages), §VI, §VIII ledger (two entries), §IX, §X (two passages); plus four hardcoded strings in build_architects_scaffold_pdf.py, one of them the PDF metadata title. Found by: drafting research/ofp-entry-3/The_Contested_Serine.md, which needed the reelin arm stated at residue resolution and so checked the sugar against source.

Was. The dissertation's compact formulation — "one sulfated lattice … shields the neuron, stages the protective reelin signal, and gates the entry of pathological tau" — treats the three offices as functions of a single material, and locates that material in the perineuronal net: "it is embedded in the same matrix the perineuronal net is built from", "the sulfated matrix of which the net is the densest expression".

Now. They are two polymers in one compartment. The perineuronal net is a condensed chondroitin-sulfate structure (lecticans on hyaluronan); reelin's obligate co-receptor and tau's route of entry are both heparan sulfate, a constituent of the diffuse extracellular matrix and the neuronal surface throughout the CNS rather than a defining component of the net (Fawcett et al., 2022, PMID 35760878). Reelin binds heparan sulfate at K_D 17 ± 5 nM — 10 ± 2 nM for the COLBOS variant — with N-sulfation the critical determinant, and heparinase, NDST1 knockout or free heparin each reduce reelin-induced ApoER2 dimerisation while N-desulfated heparin does not (Pan et al., 2025, PMID 41359976).

What was already right, and is now stated as the firmer claim. The pairing of reelin-staging with tau-entry is exact — both read one sugar. The subtitle asserted that pairing and was correct; only its word "lattice" implied the net, and it now reads "sugar". One ledger entry already said "reelin signalling and tau propagation read the same class of sulfated sugar" and needed no change.

What weakens, and it is the dissertation's central novel prediction. §VI predicted that microglial digestion of the matrix silences reelin as it removes the net, "collapsing structural and signalling defense together." The documented microglial attack is lectican proteolysis by ADAMTS and MMPs (Crapser 2020) — chondroitin-directed. Stripping or de-sulfating heparan sulfate needs heparanase or the extracellular sulfatases, and whether the activated microglion does that work in this disease is unmeasured. The prediction spans two chemistries across an unevidenced bridge, and is demoted in the ledger from near-corollary to hypothesis-with-a-named-missing-link.

A falsifier was stated with its sign inverted. §IX predicted reelin-dependent Dab1 phosphorylation would fall when the matrix is degraded "by chondroitinase". On the corrected chemistry chondroitinase leaves heparan sulfate intact and should therefore not silence reelin. Rewritten as a two-enzyme dissociation with opposite expected signs — heparinase silences, chondroitinase does not — which is a harder test than the original and cheap to run. If chondroitinase does abolish reelin signalling, the net contributes something the sugar chemistry does not predict and the single-lattice formulation should be restored.

A therapeutic claim is withdrawn. §X asserted that "an agent that preserves the functional sulfation of the perineuronal matrix would … defend the structural net and the reelin signal at once." It would not; preserving chondroitin sulfate does not defend a heparan-sulfate-dependent signal. The two-for-one benefit is removed.

And the paper's best warning is strengthened from inference to observation. §X warned that a heparan-sulfate-blocking drug aimed at tau propagation might silence reelin. Pan et al. put the proposed therapeutic class and the proposed harm in one experiment: free heparin reduces ApoER2 dimerisation. The conflict has been seen in a dish. Its converse is also now stated — net-directed and chondroitinase-based strategies cannot silence reelin by this route, which is the therapeutic face of the same dissociation.

Consequence for the wider corpus. Wherever the corpus treats the perineuronal net as the staging ground of reelin, the claim should read compartment, not net: reelin does not require a perineuronal net in order to signal. Not swept in this correction; raised as an open item alongside the errata Item 1 sweep.

Artifacts: ☑ markdown (10 passages) ☑ builder corrected (4 hardcoded strings incl. PDF metadata title, which survived the first rebuild) ☑ PDF rebuilt (116,079 → 123,255 B, 21 pp unchanged) ☑ reference added (Fawcett 2022 inserted at 13, list renumbered; in-text citations are author-date so unaffected) ☑ published PDF replaced at AdultCognitiveDisease/public/pdfs/synaptic_architects-scaffold_reelin-pnn.pdf — verified stale (116,079 B, dated 13 July, carrying the old claim 5×) before overwrite, and verified corrected after ☐ wiki republished — public/wiki/theses/reelin-pnn/phd-thesis-architects-scaffold.html and four search indices still carry the old claim. Deliberately not run: wiki:publish regenerates from the whole corpus, which currently holds 476 uncommitted files and the open E000 defects, and would publish those too.

One residual mention of "one sulfated lattice" remains in the document by design — the Validity Ledger entry that records this correction.

Update, 2026-08-13 — the framework-wide sweep C-019 raised, and a correction to C-019 itself.

C-019 overstated its own finding. The original entry said the coupled-failure prediction "rests on an unevidenced bridge" and that the missing step was heparanase. Both parts were wrong in the same direction. The better-evidenced route is not cleavage of the sugar but shedding of the core protein that carries it: MMP-2 and MMP-9 cleave the ectodomains of syndecan-1 and syndecan-4 at mapped sites confirmed by site-directed mutagenesis (Manon-Jensen et al., 2013, PMID 23384311). A metalloproteinase-rich secretome can therefore reach both polymers — lecticans for the net, syndecans for the sugar bed — which is exactly what one microglial activation needs in order to collapse all three offices together. And the corpus already contained this. The Coerulean Shears names MMP-9 as the instrument, separates the lectican from the syndecan substrate class, cites Manon-Jensen, and grades the competence "Strong." The bridge was missing from The Architect's Scaffold, not from the corpus; the two volumes simply never met on the point. §VI and the ledger of the Scaffold now carry it, with its limit stated — the syndecan work is biochemical and tumour-associated, so what is established is the enzyme's competence, not that microglial MMP-9 performs this on neuronal proteoglycans in Alzheimer cortex.

The sweep. 48 files mention both reelin and the net; the defect was present in nine, always in the same shape — the individual chemistry stated correctly, then fused in the summary sentence. Corrected:

Checked and left alone, because they were already right: reelin-handshake/PhD_Thesis_Architects_Handshake.md and coerulean-pincer/PhD_Thesis_Coerulean_Pincer.md name N-sulfated heparan explicitly; chronic-excitatory-insufficiency.md says "sulfated ApoER2 surface" and never invokes the net. And one nearby claim must not be swept: reelin genuinely is secreted into perineuronal nets (Pesold 1999). Co-location is a fact; the error was only ever the inference from co-location to chemical requirement.

Artifacts: ☑ 13 markdown files corrected across 9 documents and 3 concept nodes ☑ 8 PDFs rebuilt (First Ember, Unified Architecture, A Process In Time, Crowded Cleft, Speck & Architect, Architect's Reprieve, Coerulean Shears, Synaptic Collapse) ☑ Scaffold rebuilt again with the bridge and Manon-Jensen added as reference 16 ☑ 6 published copies refreshed via acd-pdf-map.json, each verified against its map entry before overwrite ☑ 2 documents confirmed unpublished (no map entry, no file in public/pdfs) ☑ residual sweep returns only the intended ledger audit note ☑ vault lint 0 errors ☐ wiki republished — still outstanding, and now covering more pages than C-019 alone.

C-020 · The submission's open defect list, cleared

Fidelity · 2026-08-13

Date: 2026-08-13 Location: research/ofp-entry/ONS_OFP_Entry.md (§II, §IV, §V, §VI, §VII), ONS_OFP_Entry_Bibliography.md (refs 33, 36, 40; refs 48–49 added), build_ofp_figures_pdf.py (Figures 1, 5, 8). Found by: the blind Pass 2 assessors, recorded in research/ofp-regrade/_internal/INTEGRATION_NOTES.md §3 and left open since 2026-08-11.

Miscitations, four.

• Ref 14 (Devi 2006). §II said "when amyloid-β binds the import channel TOM40". The paper reports accumulation of the amyloid precursor protein — it says so in its own title. Corrected to APP.
• Ref 20 (Holmes 2013). §VI claimed the sulfated sugar is the docking site for "tau and amyloid seeds". Holmes establishes tau and α-synuclein; amyloid was not shown. The amyloid claim is removed and the limit stated in the text.
• Ref 36. Attributed to "Werneburg"; it is Brott, Raissi, Micheva, Vielmetter, Mendes, Baccus, Huang & Shatz, PNAS 2025. Author list corrected, volume and DOI added.
• Ref 40 (Dixon 2012). Used as the sole support for oligodendrocytes dying by ferroptosis and releasing iron. It is the erastin paper in cancer cell lines — no oligodendrocytes, no brain, no Alzheimer's. Now cited only for the definition of ferroptosis, with that limit printed in the bibliography entry itself, and the oligodendrocyte-and-iron evidence carried by a new ref 49 (Gutierre 2024, PMID 39725376).

A mechanism attributed to the wrong cell, twice. §IV listed "the C4d–LilrB2 pathway" among the microglial effector arms, and §V described C4d as an "eat-me" opsonin licensing phagocytic stripping. C4d binds LilrB2 on the neuron. Brott et al. showed C4d strips dendritic spines when infused into adult cortex and that the loss is entirely abolished in PirB-null animals — synapse elimination with no microglion required. Removed from the microglial list; §V now states the neuron-intrinsic route explicitly and distinguishes it from the C1q/C3/CR3 arm. Figure 5 repeated the error ("C4d opsonin → CR3") and now reads "C1q / C3 → CR3 (glial) · C4d → LilrB2 (neuronal)".

The uncited passage that turns a relay into a theory. §VII's unity argument rested on three claims sourced to nothing. Graded rather than dressed: that TGF-β maintains the homeostatic microglion is established and now carries ref 48 (Butovsky 2014); that TGF-β additionally lies latent in the perineuronal matrix and restrains complement pruning is an extension by analogy, undemonstrated for the net in this disease; and that TREM2 is engaged by the sugar fragments matrix degradation releases is unevidenced — a literature search returns no demonstration that TREM2 binds glycosaminoglycan fragments. The paragraph now closes: the recurrence "is suggestive that the sequence is one system failing rather than a relay of separate diseases; on this evidence it does not establish it."

The A673T contradiction, conceded rather than reconciled away. §VII claimed the protective APP variant "confirms that amyloid genuinely contributes to Phase I" — while Phase I is defined in this entry by the Braak series in which 41 of 42 cases carried no amyloid. A variant acting on lifelong amyloid production cannot be evidence about a stage characterised by amyloid's absence. The entry now says so in those terms, restricts A673T to the amyloid-dependent portion of the arc beginning at the first transition, and labels §II's TOM40 proposal as its own candidate mechanism rather than a demonstrated one.

Figures.

• Figure 1's phase bands ran five years later than Figure 2 and the text (55/75 against decades 3–5, 6–7, 8+). Boundaries moved to 50 and 70; the clinical-stage strip follows them.
• Figure 8 contradicted §VI. It placed Gouras in Phase I and Small in Phase III, while §VI reads both together as "the Phase I housekeeping collapse recurring now in cortical neurons". Gouras moved to Phase III and marked "(recurring)". It also credited Area-Gómez, who appears nowhere in the entry and has no bibliography entry at all — removed. And it has no bridge column, so frameworks the text places on a transition are absorbed into an adjacent phase; that limit is now printed in the caption rather than left for a reader to discover.

One defect from the same list is corrected here as part of the corpus-wide sweep, not separately: §VI carried the "one matrix performs three offices" fusion addressed in C-019, and now reads two polymers in one compartment.

Still open, and not claimed as fixed. Bridge I's conjunction — the claim that the two arms arrive together, which is the entry's original contribution — remains untested in any system and unobservable in a living human. The entry states this; it is a limitation, not a defect, and no edit can retire it.

Artifacts: ☑ markdown (5 body passages, §VII rewritten in two places) ☑ bibliography (refs 33 and 36 corrected, ref 40 annotated with its own scope limit, refs 48–49 added and both PubMed-verified) ☑ figure builder (Figures 1, 5, 8) ☑ all four PDFs rebuilt — Entry 90 KB, Abstract 58 KB, Bibliography 53 KB, Figures 75 KB ☑ rendered PDFs verified by text extraction, not by byte count — six retired claims return zero matches across all four documents, and every replacement verified present ☑ reference-usage audit run: 48 of 49 entries cited, none orphaned n/a public/pdfs (not published) n/a wiki republish (absent from acd-pdf-map.json and acd-shelves.json)

The C-017 failure mode — correcting the text and shipping the stale figure — was specifically checked for. Figure 5 carried the C4d error independently of the prose and would have survived a text-only fix.

C-021 · One entorhinal population under two names, and the finding that never flowed back

Nomenclature and omission · 2026-08-13

Date: 2026-08-13 Location: research/reelin-alzheimers/PhD_Thesis_Architects_Reprieve.md:92 (and §VII ledger), research/colbos-pharmacology/PhD_Thesis_Architects_Understudy.md:371 and :395, kb/wiki/concepts/Reelin-ApoE receptor axis disruption.md, research/somatostatin-governor/PhD_Thesis_Unnetted_Governor.md §VII Found by: user question — "which neurons produce reelin?" — followed by a term-frequency reconnaissance of the corpus's reelin files

Two defects, one subject: which cell makes reelin, and which cell the disease reaches first.

Defect 1 — the same population carried under two incompatible names. Architects_Reprieve.md called the reelin-expressing entorhinal population glutamatergic pyramidal neurons; The_Crowded_Cleft.md:658 and ONS_CellularArchitecture_Thesis_SecondEdition.md:157 call it the layer II stellate cell. Both were faithful to their sources and neither was reconciled with the other. Chin et al. 2007 does say "Reelin-expressing pyramidal cells… the major population of glutamatergic neurons expressing Reelin in the brain" (PMID 17360894) — that laboratory's subject was amyloid, not entorhinal anatomy. The entorhinal convention names the reelin-immunoreactive layer II cells as the stellate cells of medial and fan cells of lateral entorhinal cortex, the origin of the perforant path, and distinguishes them from the calbindin-positive layer II pyramids projecting to CA1.

Why it is not merely a naming quarrel. The corpus makes cell identity load-bearing — the census counts this population, the Crowded Cleft cell-type map sorts cells by source-versus-recipient, and the perforant-path deafferentation argument depends on which layer II cell is meant. A population under two names is a population that cannot be checked across documents.

Now: all four reelin-series documents follow the entorhinal convention and read Chin's "pyramidal" as soma morphology, not projection identity. Architects_Reprieve.md gains a subsection, The Cell, Named, stating the reconciliation in the open rather than silently adopting one name. Deliberately not swept: reelin-pnn/PhD_Thesis_Architects_Scaffold.md:42 and kb/wiki/concepts/matrix-sulfation-node.md:31 use "pyramidal" for the Dab1-expressing recipient neurons, which is Pesold's own finding and correct as written.

Defect 2 — a finding that entered the corpus and never propagated backwards. Kobro-Flatmoen, Nagelhus & Witter 2016 (PMID 27195475) — intracellular amyloid-β in entorhinal layer II appears selectively in the reelin-immunoreactive neurons, at the pre-plaque stage in a transgenic rat and in human tissue with early AD-related change — was cited zero times across the entire four-leg reelin series (Reprieve, Scaffold, Handshake, Understudy) and in First Ember, Coerulean Shears and Speck & Architect. It had reached the corpus only later and laterally, through coerulean-clock/ and crowded-cleft/. The flagship reelin volume rested its entorhinal case entirely on Chin 2007, whose direction it correctly grades as ambiguous — while the evidence that sharpens the claim from shared address to shared cell, pre-plaque, in the human sat in two other documents in the same repository.

Now: Architects_Reprieve.md carries the finding in §V and a new Moderate grade in the §VII ledger, with its own settling experiment; reference 22 added.

Failure mode. Distinct from the fabrication class (C-012) and the placement class (C-014): no claim here was ever false. Every document was faithful to the source in front of it. The defect is one of circulation — a corpus large enough that a finding can enter through one volume and never reach the volume it most belongs to. Term-frequency reconnaissance across the corpus found it in one pass; nothing in the writing process would have. This argues for a standing check when any volume lands: which existing documents cite this paper's subject, and does the new source belong in them?

One connection claimed, not corrected. Pesold 1999's actual result is that reelin-positive interneurons are the somatostatin- and NPY-expressing cells and never the parvalbumin cells. somatostatin-governor/PhD_Thesis_Unnetted_Governor.md — an entire volume on the somatostatin interneuron, whose §VII argument is that this cell is the one without the net — mentioned reelin once, in passing, and never noticed that its own cell is a reelin source. §VII now states it: the somatostatin class does not wear the net and does make the reelin; the parvalbumin class wears the net and makes none. The anatomy is graded established, the consequence (that losing the source lowers cortical reelin availability) as an inference not yet measured. References 41 and 42 added.

Artifacts: ☑ markdown (4 documents, 1 concept node) ☑ 3 PDFs rebuilt — Reprieve 138,105 → 141,382 B (26 → 27 pp), Understudy 240,526 → 240,557 B (50 pp), Unnetted Governor 228,978 → 231,664 B ☑ all three verified by text extraction, not byte count — the new passages present, the retired wording absent ☑ attribution-footer parity confirmed on both published documents (the C-005 regression did not recur) ☑ public/pdfs/synaptic_architects-reprieve_reelin.pdf refreshed, map entry verified by title-page extraction before overwrite ☑ public/pdfs/synaptic_architects-understudy_colbos-pharmacology.pdf refreshed, same guard ☑ Unnetted Governor confirmed unpublished (absent from acd-pdf-map.json and public/pdfs) ☑ wiki rebuilt and deployed into public/wiki — 785 pages, 0 broken links; check:integrity reports 129 papers, all shelved, all links resolve; the concept node's new text verified present in the rendered public/wiki/concepts/reelin-apoe-receptor-axis-disruption.html ☐ ACD commit + push — outstanding; the correction is not shipped until it is pushed

A P-008 mapping trap fired during this correction and was caught. A first programmatic walk of acd-pdf-map.json reported the Understudy as unpublished, and this entry was drafted claiming so; a direct listing of public/pdfs then found synaptic_architects-understudy_colbos-pharmacology.pdf present and stale at the pre-edit byte size. Had the walk been trusted, a corrected source would have shipped behind an uncorrected published PDF — the exact failure this register exists to make visible. The lesson is the register's own: verify publication status against the directory, not only against the map.

C-025 · Errata items 2–6 applied, and a second fabricated bibliography found

Citations · 2026-08-16

Date: 2026-08-16 Found by: working the errata backlog; item 13(d) found by an over-broad edit of my own, caught by its own replacement count

Items applied.

2 — the de Vries citation. Six fabricated variants replaced with the verified entry across 37 files (40 replacements): the five-author splice, the "Perineuronal nets and cognitive resilience" invented title in both APA and Vancouver form, the wholly fabricated Radboud-authors entry in the MMP-9 paper, and the fabricated title in the bioenergetic PDF builder. Scientist page renamed daniele-de-vries.md → luuk-de-vries.md, inbound links updated.

The errata's own correction here was wrong. Item 2(f) instructed renaming to Lisa E. de Vries. The first author is Luuk E. de Vries (PubMed 39737731, DOI 10.1002/alz.14504). Applying the errata as written would have propagated a new wrong name across the corpus. The errata has been corrected in place.

3 — Perez-Nievas. The soluble-oligomer claim inverted an explicit null. Bulk soluble amyloid species did not differ between resilient and demented; what was lower was in-situ fibrillar and plaque-associated oligomeric amyloid. "Less microglial activation" corrected to glial — CD68 and GFAP only, effect in both cell types, no Iba1 or HLA-DR measured.

4 — Katzman. "As many plaques as the demented" → about eighty per cent of demented counts, in nursing-home residents of the top functional quintile carrying mild-AD pathology.

5 — Riley / Snowdon. Was cited for a dissociation of pathology from cognition; the Nun Study reports a strong association between Braak stage and cognitive state, and concerns tangles rather than amyloid. Re-cited as the association a resilience account must survive.

6 — Ulland. TREM2–DAP12–SYK through PI3K–AKT–mTOR was described as driving oxidative phosphorylation. The measured deficit was glycolytic — reduced extracellular acidification rate with relatively intact oxygen consumption. Corrected in the two theses, the tryptophan node, and constants.ts (live site).

13(d) — a second unverified bibliography, new. Applying item 6 included replacing the fabricated Ulland co-authors "Bhatt, D., & Bhatt, P." That replacement rewrote 73 citations rather than one, because the same filler pair sits at the end of seventy-three unrelated references in PhD_Thesis_Mitochondrial_Dysfunction_Neurodegeneration.md — a CD38 immunometabolism paper, a SIRT1 microglia paper, a huntingtin mitochondrial paper, and seventy more. The edit was reverted within the minute and the Ulland fix re-applied alone; the file's final diff is one line.

That paper has 151 references, 73 carrying the filler pair, and zero DOIs. It meets item 13(b)'s standard — treat the bibliography as unverified in its entirety — and is more exposed than the paper 13(b) names, being the entryPoint for the groundwork shelf. A verification note now sits at its head. The 151 references have not been verified.

Why the near-miss is recorded. The blanket edit would have attached Ulland's co-authors to seventy-three unrelated papers — replacing a fabrication with a worse one — and nothing in the rendered page would have shown it. It was caught only because the script reported its replacement count per file and the number was absurd. Sweeps across a corpus should report counts and abort on outliers rather than run silently.

Still open: items 7, 8, 9, 10, 11, 12, 13(a), 13(b), 13(c).

C-024 · Six errors in the Fischer chronology, including an inverted one

History · 2026-08-16

Date: 2026-08-16 Location: research/political/papers/Fischer_Sidelining_Biographical_History.md — abstract, §93, §95, §121, §185, §433, and both Chapter II headings; the built PDF; the published wiki page Found by: preparing a public post about the paper, and checking its two most checkable facts first — a date and a page range

Verified against: Goedert M, Oskar Fischer and the study of dementia, Brain 2009;132(Pt 4):1102–11 (PMID 18952676, DOI 10.1093/brain/awn256, open access), which this paper already cites; and the primary sources held locally in Fischer Scientific Papers/PDFs/ — the translated 1907 paper (15 pp), the 1910 paper (101 pp) and the 1912 paper (36 pp).

Where "56 pages" came from. Goedert reports plaques in 56 cases in the 1910 study. The figure appears to have migrated from a case count into a page count, and then attached itself to the wrong paper. The related "fifty-eight cases" error sits beside it, so the same passage carried one number that was wrong and one that was right-but-misplaced.

Why #3 is the serious one. The paper's rhetorical spine is that Fischer got there first and was eclipsed anyway. He did not get there first in time — Alzheimer presented seven months earlier — he got there first in scope: sixteen brains against one, and by 1910 fifty-six cases against a case report. The corrected chronology makes a better argument, because priority-by-date is a weak claim and priority-by-evidence is the paper's actual thesis. The paper's own internal detail already contradicted the wrong date: it states Auguste Deter died in April 1906, which cannot precede a November 1907 presentation of her case by seven months.

Failure mode. These are all checkable in one lookup, which is exactly the class of error a hostile reader checks first. They survived because the surrounding prose is dense and correct — the biography, the Halbhuber affair, the Bleuler editions, the eponym's propagation — so nothing looked wrong.

Artifacts: ☑ markdown — 9 passages ☑ PDF rebuilt and restamped (39 → 40 pp; stamp parity +1 on each count, being the added page) ☑ copied to public/pdfs and back to the corpus ☑ wiki republished, 0 broken links ☑ retired wording confirmed absent: "56-page" 0, "fifty-eight" 0, "November 1907" 0, "five months" 0

C-023 · The Oskar Fischer Prize was attributed to the wrong funder

Attribution · 2026-08-16

Date: 2026-08-16 Location: research/political/papers/Disciplinary_Silos_AD_Research.md — §V line 220 and §VII line 327; the built PDF; the published wiki page /theses/political/disciplinary-silos-ad-research.html Found by: user, reading the live page

What it said. "The Oskar Fischer Prize, established by the Loyd D. and Marian R. Fischer Foundation and announced in 2022…"

What is true. The prize was endowed by Dr. James Truchard. The Loyd D. and Marian R. Fischer Foundation is a different body and had no part in it.

Why it is more than a name. The paper's closing argument in §VII generalises from the prize's institutional structure — "a private foundation, working outside the NIH and journal mainstream, with a deliberately heterodox brief… it is one that worked." A structural claim resting on a misidentified institution is a structural claim about nothing. Corrected to "a privately endowed prize", which is what the funding arrangement actually was and which leaves the argument intact.

Scope. Contained to this one document — swept research/, kb/, wiki-site/templates/, constants.ts and components/ for "Fischer Foundation" and found no other instance. The sibling paper Fischer_Sidelining_Biographical_History.md is clean.

The C-005 regression fired again, and was caught by its own guard. Rebuilding the PDF from corrected markdown produced a 33-page file that had lost the per-page attribution stamp: Truchard fell 33 → 1, Gustafsson 45 → 12, and the file shrank 244,267 → 160,586 bytes. scripts/restamp-pdf.py read the stamp off the published copy and reproduced it. Post-stamp parity: pages 33 = 33, Gustafsson 45 = 45, AdultCognitiveDisease 38 = 38, Claude-Opus 33 = 33; Truchard 34 vs 33, the extra being the corrected sentence itself. Had the byte-size check been skipped, this correction would have stripped a named person's attribution from every page of a 33-page paper — which is precisely what C-005 records happening before.

Standing note. Dr. Truchard is a co-author of this project and the funder of the prize this paper evaluates favourably. That is not a defect in the correction, but it is a disclosure the paper does not currently make, and a reader who checks the funder will find it. Worth adding a conflict-of-interest line to this paper and to the OFP-entry documents.

Artifacts: ☑ markdown — 2 passages ☑ PDF rebuilt and restamped (33 pp, parity verified) ☑ copied to public/pdfs and back to the corpus ☑ wiki republished ☑ retired wording confirmed absent from source and PDF

C-022 · The two-polymer correction reached the prose and not the summaries

Chemistry · 2026-08-19

Date: 2026-08-19 Location: research/speck-architect/PhD_Thesis_Speck_And_Architect.md — abstract (§ line 18), §IX (three passages), §X (two passages), §XV coda; research/somatostatin-governor/PhD_Thesis_Unnetted_Governor.md — §VI, §VII (two passages) Found by: user question — "do we have an error with regard to the correct heparan sulfate that is involved in reelin?" — answered by verifying Pan et al. 2025 against source and then re-sweeping the corpus for the C-019 defect

The reelin chemistry itself is correct, and was checked to the source. Pan et al., J Am Chem Soc 2025;147(51):46773–46779 (PMID 41359976, DOI 10.1021/jacs.5c15573) was retrieved in full and every quantity the corpus carries from it matches: K_D 17 ± 5 nM full-length wild-type, 10 ± 2 nM COLBOS, C-terminal region 65 ± 23 nM and 32 ± 9 nM; NDST1 knockout reduces surface binding to 40% (WT) and ~30% (COLBOS); N-desulfated heparin is the weakest inhibitor of ApoER2 dimerisation while 2-O- and 6-O-desulfated heparin inhibit nearly as well as unmodified heparin; the authors' own hedge — a "modest (1.5–2-fold)" gain that "fine-tunes, rather than dramatically increases" the interaction — is quoted accurately in the Understudy. N-sulfation is the right determinant. No document in the corpus attributes reelin to the wrong polymer, the wrong sulfation position, or the wrong enzyme.

What was still wrong is C-019, incompletely swept. C-019 established that the perineuronal net is a chondroitin-sulfate structure while reelin's obligate co-receptor and tau's route of entry are both heparan sulfate, and its follow-up sweep corrected nine documents. A fresh sweep finds the "net stages reelin" formulation fully cleared corpus-wide — zero hits. But the defect survived in a second shape the first sweep's patterns did not match: the chemistry stated correctly in the body, then re-fused in the summary sentence, and the net re-defined as containing heparan sulfate.

Speck & Architect — six passages. Listed as corrected in C-019's sweep table; the fix had landed on the prose and not on the summaries. Three restatements still read "three things happen at once on one polymer" / "in a single act, on a single polymer" / "three losses on one polymer, in one act" — each sitting downstream of a paragraph that states the two-polymer distinction properly. §IX additionally defined the net as "the extracellular matrix of chondroitin-sulfate and heparan-sulfate proteoglycans," cited to Pesold 1998 — a reelin-secretion paper, not a net-composition paper — and equated the two outright: "the heparan-sulfate bed … as the aggrecan-rich perineuronal net … physically armours the neuron," closing "a shield and a gate written in one polymer."

And one claim that was not merely fused but false. §X had the activated microglion degrading the net "and, with it, the sulfated heparan-sulfate bed on which that net is built." The net is not built on heparan sulfate; its backbone is hyaluronan, hung with lecticans. Rewritten to the two-enzyme route C-019's own follow-up established: the aggrecanases cleave the lecticans, while the same secretome's metalloproteinases shed the syndecan ectodomains that carry the heparan-sulfate bed beside it — one secretome reaching two polymers, which is what makes the three offices fail together without their being one material.

Unnetted Governor — three passages, and this one was load-bearing. The file was never in C-019's sweep table (it was edited under C-021, for the reelin-cell-naming defect). §VII defined the net as "a condensed lattice of chondroitin- and heparan-sulfate proteoglycans," and §VI listed "the sulfated matrix that both shields the neuron and stages the reelin signal" among the corpus's single-structure-many-functions objects. The consequence is not cosmetic: this volume imports First Ember's locus-coeruleus inference — "the net-less neuron, unshielded and unstaged" — which C-019 explicitly demoted to conjecture because chondroitin-net poverty does not entail heparan-sulfate poverty. Defining the net as containing heparan sulfate silently re-inflates the demoted premise from the composition side, and makes the corpus's own falsifier (reelin signalling should survive chondroitinase) incoherent. §VII now carries the demotion where it borrows the inference.

What survives untouched. The Governor's core argument does not depend on the polymer at all: it rests on the net as a demolition step — microglia must enzymatically degrade the net before reaching the surface beneath (Crapser 2020) — which holds whatever the net is made of. Its §VIII ledger already graded the claim Moderate and already carried the Kv3-subset exception. Speck & Architect's bridge likewise survives: co-location plus one secretome reaching both polymers, which is a weaker and better-evidenced unity than identity.

Failure mode — new, and worth naming. C-019 and its sweep searched for the assertion ("the net stages reelin") and cleared it completely. They did not search for the arithmetic ("one polymer", "a single polymer") or the definition ("the net is made of X and Y"). A correction that rewrites an argument must also sweep the sentences that count the argument's parts and the sentences that define its nouns — the summary and the glossary are where a retired claim goes to survive. Both files here stated the corrected chemistry and the retired chemistry within a few hundred words of each other, in one case within a single paragraph.

Artifacts: ☑ markdown — 9 fragments across 2 documents, each replaced under a uniqueness assertion (single-occurrence check before write, abort on any miss) ☑ retired wording confirmed absent and new wording confirmed present by grep on both sources ☑ 2 PDFs rebuilt — Speck & Architect 158,405 → 158,719 B (32 pp), Unnetted Governor 231,664 → 232,008 B (42 pp) ☑ both verified by text extraction, not byte count — new passages present, all retired strings absent ☑ public/pdfs/microglial_speck-and-architect_nlrp3-reelin.pdf and public/pdfs/phase-ii_unnetted-governor_somatostatin.pdf refreshed, both confirmed SHA-256-identical to their rebuilt sources ☑ kb/wiki/concepts/matrix-sulfation-node.md, Reelin-ApoE receptor axis disruption.md and perineuronal-nets.md re-checked against all C-022 patterns and already clean from C-019 ☑ PDF builders inspected for hardcoded strings — neither carries the defect ☑ wiki rebuilt and deployed into wiki-dist — 1061 pages, 0 broken links, 8262 resolved links; check:integrity reports 163 papers, all shelved, all links resolve ☑ corrected text verified present and all retired strings verified absent in the rendered wiki-dist/theses/speck-architect/phd-thesis-speck-and-architect.html and wiki-dist/theses/somatostatin-governor/phd-thesis-unnetted-governor.html ☑ ACD commit + push — shipped (AdultCognitiveDisease 2dff4ade on main; OskarFischerPrize d7b93c7 + 0f46634 on reelin-synaptic-collapse-integration)

A second defect surfaced during the rebuild, and is fixed in the same push. Cloud sync had been leaving byte-identical numbered copies beside originals — Foo 2.md, Foo 3.md — and wiki-site/build.mjs's corpus scan was reading them as documents. The wiki had begun publishing a shadow copy of itself: 111 duplicate concept pages were already committed, this build would have added 132 more, and the auto-linker had begun citing them — 264 pages carried links into the duplicates and the search index held 243 entries for them. All 259 duplicates under kb/wiki were confirmed byte-identical to their canonical siblings, and no tracked .md in the repo matches the pattern, so walk() now skips it. The wiki drops 1061 → 818 pages and 874 → 635 concepts, the whole difference being debris; 0 broken links, check:integrity unchanged at 163 papers. This is the same failure mode as the correction it interrupted: a defect that had entered the published artifact and was compounding silently because nothing swept for it.

A C-021 publication claim is corrected in passing. That entry recorded the Unnetted Governor as "confirmed unpublished (absent from acd-pdf-map.json and public/pdfs)." It is present in both, as phase-ii_unnetted-governor_somatostatin.pdf, and was sitting at exactly 231,664 B — the byte size C-021's own rebuild produced. No stale text shipped, so the practical harm was nil; but the mapping trap C-021 warned about had in fact fired inside C-021 itself, in the opposite direction, and went unrecorded. Untracked 2.md/ 3.md and 2.pdf/ 3.pdf duplicates of both documents remain in the working tree, uncorrected and outside git.