Protocol: method/verification-protocol.md, Phase 1. Check date: 2026-08-10. Filing date (as-of baseline): 2026-08-03. Independence (V1): run in a session that did not author GBMT-9 and did not run its red team or its blind re-score. No verdict below relies on the filing's own citations; every load-bearing claim was re-derived from a source fetched in this pass. Targets: site/media/index.html, site/media/sources/index.html, site/media/sources/all/index.html, and the complete research record under media/ (18 documents).

Coverage (V3)

Section A = published whitepaper (site/media/index.html), 255 claims (A-1 – A-255). Section B = published sources page (site/media/sources/index.html + /all/), claims not already in A, 71 claims (B-1 – B-71). Section C = research record not appearing on the public pages, 230 claims (C-1 – C-230).

Every numbered ID is one claim and carries one verdict; the three section totals and the five verdict totals reconcile to 556 in both directions. That arithmetic is the coverage proof (V3.3).

Method notes

• Extraction was not delegated. The whole record (≈20,400 words) plus both public pages were read in-session before extraction, so a cheap extraction pass would have re-read the same material at lower fidelity. The protocol's guardrail on delegated ledgers (V3 / "Guardrail on the cheap extraction pass") is therefore satisfied by construction rather than by sampling; §A Part 6 and §C ws03-04 were re-extracted a second time against the raw HTML/markdown and produced no additional claims.
• Retrieval was delegated to five Sonnet-class agents, per "Running this efficiently." Adjudication was not. Every delegated finding that changes a verdict was independently re-verified in this session before being recorded. That check mattered twice, in opposite directions:
  • A retrieval agent reported that Medill's methodology contains no "does not include commercial television or radio stations" sentence. It checked the 2025 methodology; the record cites the 2024 methodology, where the sentence appears verbatim. The record is right and the agent was wrong. See C-108.
  • A retrieval agent reported that the MEO report contains no local/national engagement split. Independent re-checking confirmed that specific point (two separate copies of the report, grep-verified), but the agent had also implied the underlying figures do not exist anywhere; press restatements of MEO's work do carry an 85/64 pair. No primary MEO document containing it was reachable. See A-84.
  • Neither correction would have been caught by trusting the delegated output. Both are recorded because the protocol's known traps say to record them.
• Sources read from the atlas before fetching: method/sources/plaw-119-28-rescissions-act-2025.md, method/sources/moody-v-netchoice-603us707.md. The atlas's Moody file explicitly flagged the "six-justice majority" question as unresolved and told this pass to check the vote breakdown by section. That instruction is what produced finding A-91.
• V6 (dual as-of). Every time-sensitive claim carries both verdicts. For court-opinion claims, Descrybe check_case_status and find_cases_that_cite were run on Moody v. NetChoice (case c10600039): both returned unknown / zero citing cases — a coverage limit, not a clean bill of health, so currency was established instead by reading the slip opinion and the live district-court dockets directly.

Findings that change the filing

Ordered by how much weight the claim carries. Exact replacement wording is given for every CORRECTED and OVERSTATED item.

0. Corrections to this log, same day (2026-08-10)

This pass's first commit applied its corrections to the whitepaper (site/media/index.html) and the sources digest, but not to the sibling pages that publish the same claims. Two withdrawn claims stayed live in four places, one of them under an explicit "Confirmed" stamp. Fixed in the follow-up commit, recorded here rather than silently:

Root cause, worth recording: site/media/sources/{phase0-findings,research-inquiry,deviations-log}/index.html are generated from media/docs/*.md and media/research/*.md by scripts/build-source-pages.py. The first commit edited the markdown correctly and never re-ran the builder, so the corrections existed in the record and not on the site. Anything that edits those markdown files must re-run the builder. Two further stale renders (the anchor-3 ad-share note; the 2026-08-04 moving-targets updates to anchors 7 and 10) were carried over by the same rebuild.

Two substantive errors in this log itself are corrected inline below, at findings 2 and 5: the Part III–A/III–B attribution, and the deviations/commit/branch counts.

Known limitation, not fixed here: the builder's page shell is stale — it emits the pre-rename From the desk nav label, so a full build-source-pages.py run would silently regress the header on all 90 pages across four filings. The three pages rebuilt here had that line reverted by hand. The builder shell needs re-syncing with scripts/apply-header.py before anyone runs it wholesale. [Update, Phase 2, 2026-08-10: fixed since — scripts/build-source-pages.py now emits Findings, so a wholesale run is safe. Phase 2 ran the builder for real; it also created the six missing media source pages this log flagged at B-66, including this one.] media/report/gubment-media-report.html also still carries the pre-verification receipts line ("11 entries … four commits across two branches") and none of this pass's corrections; it is a snapshot artifact and was left alone.

1. CORRECTED — Canada's local-vs-national engagement split is not in the cited source (A-84, A-88)

Published (Part 5 stat block and body): "-85% / -64% Canada: local vs national outlets' Facebook engagement loss" and "local outlets lost 85% of their Facebook engagement at the one-year mark versus 64% for national outlets," sourced to Media Ecosystem Observatory, "Old News, New Reality" (Aug 2024).

What the cited report actually says. Two independent copies were obtained and searched: MEO's own PDF (mediatechdemocracy.com/files/publications/meta-news-ban_2024.pdf) and the Public Inquiry into Foreign Interference's archived exhibit COM0000513 (foreigninterferencecommission.ca). Both are the same 9-page brief. Verbatim:

"Canadian news outlets have lost 85% of their engagement on Facebook and Instagram." "Canadian news organizations have lost 42.6% of their online engagement since the ban across their social media footprint." "Of these, 212 (98%) are local news organizations: 30% of the 713 Canadian local news outlets who were previously active on social media are no longer able to share their content and build their readership online."

The string "64%" does not appear anywhere in the document. The word "national" appears exactly once, in a contributor biography. There is no local-versus-national engagement comparison in the report at all, and the 85% figure is MEO's all-outlets Facebook+Instagram number, not a local-outlet number.

Press restatements of MEO's wider work do report an 85/64 Facebook-Pages pair, and The Hub (2024-08-08) reports a different pair again — local −58%, national −24%, measured across five platforms. No primary MEO document containing either pair was reachable in this pass. Under the protocol's known traps ("Never record a figure you have not seen inside the fetched source itself"), the figures cannot be certified from a press restatement.

Replacement wording (Part 5 body):

"When Canada passed its own version, Meta simply blocked news outright rather than pay — a ban still in effect as of mid-2026 — and the damage landed asymmetrically: Canadian news outlets lost 85% of their Facebook and Instagram engagement overall and about 43% across their whole social footprint, and 30% of the 713 local outlets previously active on social media went dormant, against 2% of everything else. Meta's own usage in Canada stayed essentially flat by two independent measures."

Replacement stat-block pair: -85% → "Canadian news outlets' Facebook/Instagram engagement loss"; -64% → replace with 30% → "of Canada's local outlets went dormant on social."

This preserves the asymmetry finding — which survives, and is local-specific in the source, unlike the number that was published — while dropping a figure the cited report does not contain.

2. CORRECTED — Moody v. NetChoice did not command six justices for the proposition the filing relies on (A-91)

Published (Part 6): "The Supreme Court's 2024 decision in Moody v. NetChoice held, with a genuine six-justice majority, that a platform's curation of its own feed is protected editorial speech — not a minor aside, a real holding."

Primary text. Slip opinion fetched from supremecourt.gov (22-277_d18f.pdf, HTTP 200 via browser user-agent; WebFetch 403s). The lineup, verbatim:

"KAGAN, J., delivered the opinion of the Court, in which ROBERTS, C. J., and SOTOMAYOR, KAVANAUGH, and BARRETT, JJ., joined in full, and in which JACKSON, J., joined as to Parts I, II and III–A."

The proposition the filing relies on is in Part III–B, not III–A. Part III–A is the facial-challenge framing — it faults both courts below for failing to conduct a proper facial analysis and sets out what that analysis requires. Part III–B is where the Court both restates the settled editorial-discretion principles from Tornillo, PG&E, Turner and Hurley and applies them to social-media feeds. Both halves of that reasoning therefore sit outside what Jackson joined:

"…the major social-media platforms are in the business, when curating their feeds, of combining 'multifarious voices' to create a distinctive expressive offering." (Part III–B) "When the platforms use their Standards and Guidelines to decide which third-party content those feeds will display, or how the display will be ordered and organized, they are making expressive choices. And because that is true, they receive First Amendment protection." (Part III–B)

Jackson did not join Part III–B, and said so on the merits:

"In light of the high bar for facial challenges and the state of these cases as they come to us, I would not go on to treat either like an as-applied challenge and preview our potential ruling on the merits."

And Alito, joined by Thomas and Gorsuch:

"The holding in these cases is narrow: NetChoice failed to prove that the Florida and Texas laws they challenged are facially unconstitutional. Everything else in the opinion of the Court is nonbinding dicta."

So on the platform-curation proposition: five justices joined it; one (Jackson) expressly declined to reach it; three called it nonbinding dicta. The Court's actual holding was that "the judgments are vacated, and the cases are remanded, because neither the Eleventh Circuit nor the Fifth Circuit conducted a proper analysis of the facial First Amendment challenges."

This is the hinge of the filing's §7 constitutional screen, and it is the INTERPRETATION failure mode (V4) exactly: the vote breakdown in the record is accurate; the inference drawn from it is not. media/docs/phase0-findings.md §4 says Jackson joined "in relevant part" — but the relevant part is the one she did not join.

Replacement wording (Part 6):

"The Supreme Court's 2024 decision in Moody v. NetChoice vacated and remanded both cases for failing to analyse them as facial challenges — but along the way, five justices held that a platform's curation of its own feed is protected editorial speech. A sixth, Justice Jackson, joined only the facial-challenge analysis (Parts I, II and III–A) and expressly declined to reach the merits; the editorial-discretion reasoning itself sits in Part III–B, which she did not join. Three justices called the whole discussion nonbinding dicta. It is stronger than a stray aside and weaker than settled law, which is why the screen below is directional."

[Correction, 2026-08-10, same-day: an earlier version of this finding said Part III–A "restates the settled editorial-discretion principles," and the replacement wording credited Jackson with "joining the general editorial-discretion principles while expressly declining to apply them." Both were wrong in the same direction: the Tornillo/PG&E/Turner/Hurley material is in Part III–B, which Jackson did not join at all, so she joined none of the editorial-discretion reasoning. Corrected here and on the live page.]

Also correct media/docs/phase0-findings.md §4, media/docs/research-inquiry.md anchor 9, and the sources-page anchor 9, all of which carry "a real 6-justice majority."

3. CORRECTED — CPB's insulation was not built into the 1967 Act (A-56)

Published (Part 3): "The Corporation for Public Broadcasting was funded two years in advance on a formula basis specifically to insulate it from short-term political pressure — a firewall built into the Public Broadcasting Act of 1967 and never removed."

Primary text. 47 U.S.C. § 396 fetched from uscode.house.gov (currency: "laws in effect on August 9, 2026"), with its source credits and amendment notes:

• The original 1967 Act is Pub. L. 90-129. The funding mechanism the filing describes is not in it. § 396(k) was rewritten twice afterwards — Pub. L. 94-192 (Public Broadcasting Financing Act of 1975) added paragraphs (k)(3)–(7), and Pub. L. 95-567 (Public Telecommunications Financing Act of 1978) "completely revised and restructured subsec. (k)."
• The explicit content firewall at 47 U.S.C. § 398(c) was likewise added in 1978 by Pub. L. 95-567 ("designated existing provisions as subsec. (a)… and added subsecs. (b) and (c)"). What is in the 1967 design is § 396(g)(1)(D)'s "maximum freedom … from interference with, or control of, program content."
• The statute's dollar authorizations in § 396(k)(1) run out after FY1996. Modern CPB funding came from ordinary appropriations acts: Pub. L. 118-47 reads "For payment to the Corporation for Public Broadcasting … for the fiscal year 2026, $535,000,000." That is a fixed sum set by appropriators, not a formula. The distribution formula (§ 396(k)(3): 5% administration, 6% capital/interconnection, then 75/25 television/radio splits) is real and is what "formula-funded" correctly describes.
• Two-year advance funding is confirmed as practice — the FY2026 money came from a 2024 act — and CRS describes it as such ("has historically received two-year advanced appropriations"), not as statutory command.

The correction does not weaken the filing's argument; it sharpens it. The insulation was bolted on across 1975 and 1978 and was never as statutorily entrenched as "built into the 1967 Act and never removed" implies — which makes its collapse to a simple-majority rescission less surprising, not more.

Replacement wording (Part 3):

"The Corporation for Public Broadcasting distributed its money to stations by statutory formula and was funded two years in advance — insulation built up across the Public Broadcasting Financing Act of 1975 and the Public Telecommunications Financing Act of 1978 rather than in the 1967 Act itself, and never removed. The forward funding was settled practice rather than statutory command: Congress appropriated CPB's FY2026 money, $535 million, back in 2024."

4. CORRECTED — the whitepaper says a workstream was never run; it is in the same repository (A-197)

Published (honesty box): "no dedicated workstream on US v. Google's actual remedies was ever run."

media/research/ws13-google-remedies.md is dated 2026-08-06, is committed on main, and is exactly that workstream. It finds a liability ruling and a proposed final judgment but no entered remedy, and concludes the filing "must not credit it with coverage, newsroom-employment, or civic-information effects." The honesty box is describing a gap that was closed three days after filing and never updated.

Replacement wording:

"…and the correction is documented in full rather than quietly fixed. A dedicated US v. Google remedies workstream has since been run (§13, 2026-08-06): it found a liability ruling and a proposed divestiture of AdX, but no entered remedy and no evidence connecting the case to local-news coverage — so the scorecard's midrange, explicitly unverified treatment of that architecture stands."

5. CORRECTED — the published receipts undercount the record (A-204, B-70, B-71)

The deviations claim appears once in the whitepaper and twice on the sources page. Entries 12–16 (the blind re-score, the Google-remedies pass, the voucher-evidence pass, the capacity proxy, the ghost-paper audit) are precisely the work the receipts line is meant to advertise; entry 17 is this verification pass itself.

[Correction, 2026-08-10, same-day: this finding was first recorded as "16 entries … fourteen commits across eight branches," which was the state of the record before this pass wrote itself into it, and it undercounted the branches. Recounted:

• 17 deviations entries. Rows 1–17 of deviations-log.md; row 17 is this pass. 16 was the pre-pass count.
• 16 commits. git log origin/main -- media/ returns 14; this branch (verify/media-phase1) adds 2 — the Phase 1 commit and this cross-page reconciliation commit.
• 10 branches. The 14 commits on main came from 9 distinct branches, not eight — gbmt-9-media, gbmt-9-media-cont, site-sources, report-pdfs-7-8-9, media-moving-targets, second-scores, second-scores-crypto-ai, gbmt-9-google-remedies-tail, gbmt-9-ghost-paper-tail — attributed by git log --merges --ancestry-path per commit. verify/media-phase1 is the tenth.

The published receipts line already read "17 entries · fifteen commits across nine branches"; the entry count was right and the commit/branch pair was one short in each term. Corrected to 16/10 on both public pages.]

Replacement: "deviations log (17 entries) … sixteen commits across ten branches, all public."

6. CORRECTED — the news-desert counts are one off from Medill's report (A-34, A-35)

Medill's State of Local News 2025 report body and full PDF state, five times: "there are now 212 U.S. counties with zero local news sources, compared to 206 in 2024," and "there are 1,525 counties with only one local news source remaining," out of "the 3,144 counties in the U.S." The whitepaper publishes 213 and 1,524.

213 appears only in the landing-page introductory essay, an uncorrected inconsistency on Medill's own site. The filing's own Phase 0 recorded the ranges honestly ("212-213", "1,524-1,525"); the public page picked one end of each without disclosing the range.

The combined share is unaffected: 213+1,524 and 212+1,525 both total 1,737, and 1,737/3,144 = 55.2%.

Replacement: 213 → 212; 1,524 → 1,525; and add to the body: "(Medill's own landing page says 213; its report and dataset say 212 — we use the report.)" Note also that "~55%" is computed, not a Medill-stated figure.

7. CORRECTED — the abstract contradicts Part 9 of the same page (A-16, A-17)

Abstract: "no single architecture dominates, and the most reliable results are at the bottom of the ranking, not the top." Part 9: "Public media funded outside the annual appropriations cycle is now the clear most-consistent performer."

The abstract states the post-red-team (v2) finding; Part 9 states the post-re-score (v3) finding from 2026-08-04. The re-score strengthened the top of the board — ws11-rescore-log.md: "a strengthened version of the post-red-team finding." Part 9 and the sidebar were updated; the abstract was not.

Replacement (abstract closing):

"We report the corrected finding: no single architecture dominates every weighting, but after an independent blind re-score, public media funded outside the annual appropriations cycle is the most consistent performer, and the platform bargaining code is the one instrument the record actively refutes rather than merely ranks last."

8. CORRECTED — Amazon's ~17% share is a 2026 projection presented as 2024 (A-30)

Published (Part 1): "Their combined share fell below 50% in 2024 for the first time in roughly a decade, as Amazon's share climbed toward 17%."

The filing's own Phase 0 says: "Amazon's share reaching ~17.3% by 2026." The consistently reported 2024 figure is ~13.9%. The public page moves a 2026 forecast into a 2024 sentence, against its own record.

Replacement: "…as Amazon's share passed 13% in 2024 and is forecast to reach about 17% by 2026."

9. OVERSTATED — the "prior broken" headline is itself single-root (A-31)

Published (Part 1): the below-50% ad-share claim is "independently verified against two separate reports," citing "eMarketer 2024–2026 estimates · IAB/PwC Internet Advertising Revenue Report FY2024."

IAB/PwC was obtained and confirms only the market total: "internet advertising revenue … reaching a record $258.6 billion in 2024." It publishes no Google/Meta share. Every share figure reached traces to eMarketer/Insider Intelligence, and eMarketer's own vintages disagree with each other about the same year (43.0%, 47.1%, 47.7%, 48.4% all appear for 2024 across successive reports). No non-eMarketer source for the below-50% crossing was found.

This is the filing's own signature failure mode — a single research lineage restated until it looks corroborated — operating on the filing's own headline. It deserves the same flag the desert count gets.

Replacement:

"Their combined share fell below 50% in 2024 for the first time in roughly a decade. One caveat we owe our own method: every published version of this figure traces to a single estimator, eMarketer, whose successive vintages put the 2024 number anywhere from 43% to 48% — the same single-root problem we flag in the desert count below. The market total it sits inside ($258.6B for 2024) is independently confirmed by IAB/PwC."

10. OVERSTATED — the BLS "independent" confirmation is the same root (A-23)

Published: newspaper newsroom employment fell 57%, "confirmed against BLS data independently of Pew's own restatement."

Pew's figures are a Pew analysis of BLS OEWS data — the filing's own anchor table says so: "Single-root: all three Pew figures are BLS-OEWS-derived." Reproducing them from BLS tables checks Pew's arithmetic; it does not add an independent root.

That arithmetic does check out, and this pass verified it at a tier the filing never reached. From the BLS OEWS industry tables for NAICS 511110 (retrieved via Wayback; bls.gov 403s to automated fetch): May 2020 gives 14,680 + 14,700 + 1,330 + 110 = 30,820, an exact match to Pew's published 2020 figure; May 2008 gives 33,840 + 31,880 + 5,190 + 40 = 70,950 against Pew's 71,070 (0.17% apart, consistent with a suppressed small cell).

Replacement: "…57%, and we reproduced Pew's arithmetic directly from the underlying BLS occupational tables — though that is the same data root, not an independent one. The one genuinely separate source, ASNE's newsroom census, was discontinued around 2015."

11. OVERSTATED — the Nordic steelman's capture finding is not in the cited paper (A-129)

Published (Part 8): "The modern comparator, Nordic direct press subsidies, has genuine peer-reviewed support for avoiding viewpoint capture through the same design principles — but the same literature finds these subsidies did not prevent commercial decline," citing Grönlund, Villi & Ala-Fossi, Media and Communication (2024).

The citation is correct and the paper is real: "Press Subsidies and Business Performance of Newspaper Publishing in Three Nordic Media Welfare States," DOI 10.17645/mac.7570. Its abstract, retrieved in full, is entirely about financial performance. The commercial-decline half is confirmed verbatim — "direct press subsidies in Sweden and Norway have not been able to prevent a decrease in the number of titles and … have resulted in a number of subsidy-dependent news outlets unable to survive on their own income" — and so is the Finland comparison. The capture half is not in the paper. It contains no finding about viewpoint capture, state capture, or editorial independence.

This matters because the steelman owed under M3 is the filing's counterweight to its own subsidy skepticism, and half of it is attributed to a source that does not support it.

Replacement:

"The modern comparator, Nordic direct press subsidies, is credited in the policy literature with avoiding viewpoint capture through independent administering bodies and formula-based allocation — a design claim, well documented but not something the peer-reviewed financial study we rely on here tested. What that study does find is that the subsidies did not prevent commercial decline: Sweden's and Norway's subsidies failed to stop titles disappearing and produced outlets unable to survive on their own income, while Finland, which abandoned direct press subsidies in the 1990s, was in a better financial position on almost every indicator."

12. OVERSTATED — CPB is not the only completed durability test in the record (A-3)

Published (subtitle): "the one architecture in this domain with a completed, real-world durability test: a 58-year-old public broadcasting institution…"

The filing's own re-score log scores the bargaining code's durability at 1 on the ground that "Meta's walk-away is a completed failed real-world test." That is a second completed test, of a second architecture, in the same document. Comparators #9 and #10 also carry B=5 on track record.

Replacement: "…built around the architecture whose durability test ran longest and failed hardest: a 58-year-old public broadcasting institution, formula-distributed and forward-funded, that just voted to dissolve itself."

13. OVERSTATED — "$38.7 billion in a single year" reads as a year-over-year fall (A-112)

Published (Part 7): "Newspaper ad and circulation revenue fell $38.7 billion in a single year (2022 versus the 2005 peak)."

It is a single year's shortfall against the 2005 peak, not a one-year fall. The parenthetical corrects it, but the sentence does not survive being read on its own — and it is the numerator of the 55× ratio that anchors Part 7.

A second caveat the filing never discloses: Pew's revenue series changes source mid-stream. Pew's own methodology says pre-2013 figures come from the Newspaper Association of America, and "2013–2022" figures are "Pew Research Center analysis of year-end SEC filings of publicly traded newspaper companies" — "which in 2022 numbered four." The $21.4B endpoint is scaled from four companies' filings, not an industry census, which is more uncertainty than a figure quoted to $0.1B implies.

Replacement: "Newspaper ad and circulation revenue in 2022 ran $38.7 billion below its 2005 peak — a single year's shortfall, not a single year's fall. (Pew's post-2013 revenue series is estimated from the filings of the four remaining publicly traded newspaper companies, so treat the precision loosely.)"

14. OVERSTATED — the published scorecard drops the cash comparator without saying so (A-126–A-175)

The record's reconciled matrix has eleven rows and 55 cells. The published table has ten rows, 50 cells. The missing row is #10, "Cash (universal comparator, per M7)", scored 1/5/2/2/2. The words "cash" and "comparator" appear nowhere on either public page.

M7's universal cash comparator is a standing method requirement, and the filing's own §11 seed list names it. Publishing ten of eleven rows without disclosure also makes the receipts line ("10 architectures × 5 anchored objectives") inconsistent with the record's own "55 cells."

Fix: restore the cash row to the published table, or state in the caption that the cash comparator is scored in the record and omitted here.

15. OVERSTATED — "5 weightings" cannot be checked against anything in the record (A-186, A-195)

Both public pages advertise "5 weightings," and Part 9's central claim is that "no single architecture dominates" across weightings. The record contains:

• no definition of the four original weightings, anywhere;
• one weighting vector, named once in the red-team log (pluralism-first, A15/B10/C60/D5/E10);
• no per-weighting rankings. ws11-scorecard.md's "Rankings, re-scored" section reports only a neutral-weight mean, and ws11-rescore-log.md reports the same.

The red-team log says "The recomputed rankings are in the revised scorecard." They are not. Everything published about rank stability under alternative weightings rests on a computation that exists nowhere in the committed record. This is a V3 coverage gap on a load-bearing claim, not a wording problem.

The neutral-weight arithmetic that is in the record checks out: 4b = (4+3+5+3+3)/5 = 3.6, and nothing else exceeds 3.0 (#4, #6 and #8 tie at exactly 3.0). "Nothing else above 3.0" is correct as written.

Fix: either commit the five weighting vectors and their rankings to ws11-scorecard.md, or drop "5 weightings" from both public pages and state the neutral-weight result only. Until one or the other lands, the claim is UNVERIFIABLE.

16. INTERPRETATION — the filing's headline architecture is scored on the cell the re-score existed to eliminate (C-141)

Not a factual error; a reading the record does not support, on the cell that decides the filing's headline.

Axis B's anchors read: "1 = no demonstrated track record of surviving a hostile majority, or a completed real-world test that failed · 3 = has meaningfully outlasted normal political cycles but with no completed test against a determined hostile majority."

The blind re-score applied that rule hard, and said so: "a brand-new instrument cannot score 3." It moved #1 (payroll credit) 3→1 "no track record exists"; #5 (vouchers) 3→1 "Never implemented anywhere; no record"; #7 (philanthropy) 4→2.

#4b — public media, non-appropriated — has never been implemented either. The scorecard's own basis text concedes it: "No project-sourced evidence of a non-appropriated public-media mechanism with a completed, successful durability test exists." Under the anchor as written, that is the definition of B=1. It was scored 3, reasoning down from a higher score via the Finland cap rather than up from the floor — the same design-label leak the re-score was run to catch, surviving on the one row it most affects.

Consequence if corrected: 4b's neutral-weight mean falls 3.6 → 3.2. It stays top (next best 3.0), so the ranking survives, but "the clear most-consistent performer" loses most of its margin, and the filing's title question — "can insulated public journalism funding survive a hostile Congress?" — is answered by a cell scored above its own floor on no evidence.

This pass does not re-score it. Phase 1 is a fact-check, and M6 requires a blinded scorer. Recorded here, flagged for the next re-score and for Phase 2's S5 sensitivity analysis.

17. UNVERIFIABLE — three litigation claims in Part 6

• A-97, "Florida's is headed to trial September 8, 2026." The docket (NetChoice LLC v. Uthmeier, N.D. Fla. 4:21-cv-00220, via CourtListener RECAP) confirms the case is live before Judge Hinkle, and confirms a bench trial was set and then repeatedly rescheduled (3/9/2026 → revised 2025-07-03, 2025-11-05, 2025-12-15). No entry setting September 8, 2026 was found, and on 2026-06-30 the parties filed a "Joint MOTION for Extension of Time to Amend Pretrial Schedule" — after the filing's last re-check. A specific future date is published as fact with a live motion to move it.
• A-99, "while rejecting Florida's argument that the First Amendment doesn't reach platform editorial decisions." The order is bare: "ORDER DENYING SUMMARY JUDGMENT. The summary-judgment motions, ECF Nos. 233, 281, 287, and 291, are denied. Signed by JUDGE ROBERT L HINKLE on 6/18/2026." Four motions — both sides — denied, with no reasoning stated. Any reasoning would be in the 6/8/2026 hearing transcript, which the filing does not cite.
• A-100, "Texas's injunction stands after a May 2026 stay denial, with its scope still being briefed." The docket (NetChoice LLC v. Paxton, W.D. Tex. 1:21-cv-00840, 219 documents) records no order in May 2026; its 2026 entries are 1/30, 2/24–25, 6/29–30 and 7/9. The only stay ruling on the docket is a grant: "Text Order GRANTING 78 Motion to Stay … In light of the parties' desire to await possible legislative action," 2025-04-16. Scope briefing closed 2025-09-10, and on 2026-07-09 the court signalled "the Court's forthcoming order" — so "still being briefed" is also STALE.

Replacement (Part 6): "…Florida's case survived summary judgment in June 2026, when the district court denied every motion from both sides without stating its reasons, and is awaiting trial; Texas's case has been fully briefed on the scope of HB 20 since September 2025 and is awaiting a ruling the court signalled in July 2026. Nothing about the doctrine is fully settled yet."

18. UNVERIFIABLE — Australia's ~90% figure rests on a citation that does not exist (A-87)

Published (Part 5): "the one peer-reviewed distributional estimate available finds roughly 90% of bargaining-code payments flowed to three large incumbent publishers," cited as Brevini & Ward, Media International Australia (2023).

No such paper exists. Crossref and OpenAlex were queried directly:

• Benedetta Brevini's only 2023 bargaining-code paper is sole-authored: "Global Digital Lords and Privatisation of Media Policy: The Australian Media Bargaining Code," Javnost – The Public 30(2): 268–283, DOI 10.1080/13183222.2023.2207427. No co-author named Ward appears in her indexed output.
• A journal-scoped Crossref query of Media International Australia for bargaining-code work returns other authors entirely (Neilson & Heylen, and others) — no Brevini, no Ward.

The real paper is paywalled (Unpaywall: is_oa: false), so the ~90% figure could not be verified here either. The filing never read it: deviations log #5 records a 403 on "Brevini's Media International Australia article."

This is the filing's load-bearing evidence for Part 5 and for four of the five 1-scores on scorecard row 2 — the one architecture the whitepaper says the record "actively refutes." It is sourced to a paper with the wrong journal, the wrong author count, and no reader.

Fix: correct the citation to Brevini (2023), Javnost – The Public 30(2), and mark the ~90% figure as reported-but-unread until someone obtains the text — or obtain it. Do not leave it described as "the one peer-reviewed distributional estimate available" while the estimate itself is unverified.

19. UNVERIFIABLE — CPB's board-vote date (A-58)

Published: "CPB's board voted to dissolve the organization outright on December 10, 2025; the announcement came January 5, 2026."

CPB's own press release (cpb.org DNS-fails; retrieved via Wayback, capture 2026-01-31) is dated January 5, 2026 and says only that CPB "announced today that its Board of Directors has voted to dissolve the organization after 58 years of service." No separate vote date appears in it, and no primary document giving December 10, 2025 was reachable. The "58 years" figure is CPB's own and is confirmed.

Fix: cite the announcement date, which is documented, and drop or source the vote date.

20. Smaller corrections

Findings that strengthen the filing

Recorded because a fact-check that only subtracts is not a fact-check.

• The 39% cap's statutory text was obtained — method/sources/UNOBTAINED.md listed it as a media gap ("never cited by section, though it's the statute actually in dispute per Commissioner Gomez's dissent"). It is § 629 of the Consolidated Appropriations Act, 2004, Pub. L. 108-199, amending § 202(c)(1)(B) of the Telecommunications Act of 1996: it strikes "35 percent," inserts "39 percent," and adds to § 202(h): "This subsection does not apply to any rules relating to the 39 percent national audience reach limitation in subsection (c)(1)(B)." That sentence is the precise text the FCC and Commissioner Gomez read in opposite directions, and the filing can now cite it. The statute says nothing about the UHF discount, which is a Commission rule (47 C.F.R. § 73.3555(e)(2)).
• The FCC repeal is confirmed exactly as the filing's update bracket states it. FCC 26-53, MB Docket 17-318, adopted 2026-08-06: "voted to repeal its 39% national television multiple ownership rule and replace it with a granular, case-by-case review," Carr and Trusty approving, Gomez dissenting. Gomez's stated ground, verbatim: "Congress set this cap in federal law, and only Congress can change it."
• Nexstar-Tegna confirmed on every number from the FCC's own order (DA 26-267, adopted March 19, 2026) and Nexstar's release: closed March 19, 2026; $6.2B; "Nexstar would reach 54.5% of the national audience, after taking into account the 50% discount applied to UHF stations"; 80% of TV households pre-discount; six named divestitures. It was cleared by a case-specific waiver five months before the rule itself was repealed — exactly as the filing frames it.
• CPB's rural incidence confirmed from CPB's own words, retrieved from a pre-dissolution Wayback capture of cpb.org/aboutpb/rural (2025-06-13): "CPB grants … represented 17 percent of an average rural station's revenue, versus 9 percent for non-rural stations. Almost half of all rural grantees – 120 stations – relied on CPB for at least 25 percent of their revenue while 33 rural stations – many on Native American reservations – relied on CPB funding for at least 50 percent." The "some 1,500 stations" reach figure is confirmed too: 386 radio grantees representing 1,216 stations plus 158 television grantees representing 365 stations = 1,581.
• The Senate vote is exactly 51–48 — Senate roll call vote 411, July 17, 2025, "On Passage of the Bill (H.R. 4, As Amended)," 51 yea / 48 nay / 1 not voting.
• The Mississippi bias-sizing — the filing's most original work — holds. Medill's live county API (northwesternlni.com:9095, state=MS&year=2025), queried across all five outlet categories for all 82 counties, returns exactly five zero-outlet counties: Benton, Issaquena, Jasper, Lamar, Tunica — the filing's list, independently reproduced from the publisher's own data.
• The Medill exclusion quote is verbatim and correctly attributed. The 2024 methodology reads: "Our database does not include commercial television or radio stations. These stations tend to focus their coverage across a wide geographic area, and our research is especially interested in community journalism at a local level." Worth noting for future passes: the 2025 methodology drops that sentence, so the citation must stay pinned to 2024.
• The 2018 ghost-paper estimate is verbatim: "at least 1,000 of the 7,200 newspapers still published in this country – and perhaps as many as 1,500 – have lost significantly more than half of their newsroom staffs since 2004."
• The single-root claim about the desert count survives a positive check (V5). Every candidate alternative census measures something else or feeds Medill: the Local Journalist Index measures journalist-equivalents, not outlets; INN is a self-selected membership association; and Medill's own methodology names Editor & Publisher and AAM as input sources, which makes them upstream of Medill rather than independent of it. The negative claim is stronger than the filing states.
• Gallup confirmed exactly (news.gallup.com, published 2026-10-02 for a Sept 2–16 2025 poll): 28% overall, "the first time the measure has fallen below 30%"; Republicans 8%, Independents 27%, Democrats 51%.
• Peterson & Dunaway confirmed to the digit (Crossref DOI 10.1177/00027162231211426): "over 13,000 digitized media directory pages … 211 major newspapers from 2005 to 2022 … reduced the paper's newsroom by nine reporters and editors … a cut equivalent to 14 percent of the average newspaper's staff … concentrated among positions focused on general assignment and political reporting."
• Gao, Lee & Murphy confirmed (+5.5 bps offering, +6.4 bps secondary), and the filing's negative claim survives a real search: all 233 OpenAlex-indexed citing works were scanned and none replicates the bond-yield result with a non-overlapping author team. The $1.1B/yr restatement is confirmed as Baker & Murphy, 2026-06-10, applying the original 8.6 bps weighted coefficient to 2025 issuance — a self-citation, exactly as ws03-04-findings.md says. The whitepaper correctly declines to use the $1.1B figure.
• Maryland confirmed and sharper than published. Chamber of Commerce v. Lierman, No. 24-1727 (4th Cir. Aug. 15, 2025): "We therefore hold that the pass-through provision is facially unconstitutional," reversed and remanded for a remedy. The court applied intermediate scrutiny and found the provision failed even that. The gross-receipts tax itself was not disturbed — the filing's framing is right.
• Leathers v. Medlock supports the screen's use of it. 499 U.S. 439 (1991): a generally applicable, content-neutral sales tax reaching cable television is constitutional; differential taxation becomes suspect when it singles out the press, targets a small group, or turns on content. That is good authority for the payroll credit's "SAFE (settled)" label — and, unremarked in the filing, better authority than it currently claims for the earmarked ad tax's elevated risk.
• IAB/PwC confirmed: "$258.6 billion in 2024."
• The QCEW baseline pipeline is sound. The committed splice check reproduces: 2021 NAICS 511110 = 102,521 → 2022 NAICS 513110 = 104,153, a +1.6% discontinuity, well under the 15% threshold. The incompatibility log's three entries are accurate and the pre-2014 gaps are real.

Section A — published whitepaper (site/media/index.html)

211 claims. Verdict key: C = CONFIRMED, X = CORRECTED, O = OVERSTATED, U = UNVERIFIABLE, S = STALE. "As-of" column: F = correct as of filing date (2026-08-03), K = correct as of check date (2026-08-10); "—" = not time-sensitive.

Header, case file and abstract

Part 1 — the collapse, measured

Part 2 — the desert map

Part 3 — the institution that dissolved itself

Part 4 — who owns what's left

Part 5 — what platforms actually pay for

Part 6 — the constitutional screen

Part 7 — why the money isn't enough

Part 8 — other eras and countries

Part 9 — the scorecard

Published scorecard cells (A-186 – A-235)

50 cells, each a claim under V3.1. All 50 reproduce the reconciled matrix in ws11-rescore-log.md exactly — verified cell by cell. Verdicts below are on the published value's fidelity to the record and to its stated basis, not a re-score.

• A-191, A-193, A-194 (bargaining code A/C/D = 1): each rests on the ~90%-to-three-incumbents figure, whose citation does not exist (finding 18). The scores may well be right; their stated basis is unverified.
• A-207 (4b durability = 3): violates axis B's own anchor for a never-implemented instrument (finding 16).
• All other cells match the record and their cited workstream findings.

Honesty box and receipts

Section A totals: 255 claims — 199 CONFIRMED, 12 CORRECTED, 23 OVERSTATED, 19 UNVERIFIABLE, 2 STALE.

Section B — published sources page

71 claims (B-1 – B-71). Claims duplicating Section A are cross-referenced, not re-adjudicated.

Section C — research record not on the public pages

230 claims (C-1 – C-230). Adjudicated in full; only entries whose verdict is not CONFIRMED are itemised, per the protocol's "same register as the existing red-team logs." The remaining 191 are CONFIRMED against the sources named in Sections A and B and are listed by document below.

Itemised non-CONFIRMED entries

What this pass did not reach

Per V5 and the protocol's "record any authority you could not obtain":

• Brevini (2023), Javnost – The Public 30(2) — paywalled; the ~90% distributional figure remains unread by anyone on this project.
• Rod Sims / Judith Neilson Institute report on Australian deal values — JNI 503, SSRN 403, Wayback 429.
• Richard R. John, Spreading the News (1995) — no open copy; Kielbowicz's Postal Rate Commission history was obtained in full and contains none of the cited figures.
• Nordicom Factsheet 2022:1 — the cited URL 404s and no replacement was found on nordicom.gu.se.
• Neff & Pickard's comparative-spending paper — SAGE 403, as it was for the filing.
• The FCC's full Report and Order text (FCC 26-53) — the news release, all three commissioner statements and the Federal Register notice were obtained; the order's regulatory text was not.
• wreg.com — 403 to automated fetch, no Wayback capture after 2020.
• CPB's board-meeting record for December 2025 — cpb.org DNS-fails; Wayback holds the January 5 2026 release but no vote minutes.

These belong in method/sources/UNOBTAINED.md as known-unknowns; the 39% cap statute listed there is now obtained and should be moved.