Date: 2026-08-03 Scope: First full-execution pass after Phase 0's GO verdict, picking up the two items ranked highest in phase0-findings.md §8: sizing one of the news-desert count's two documented biases (§3), and the civic-harms literature independence audit that Phase 0 flagged but didn't attempt (§4, the outstanding kill-condition test named in the protocol's own execution notes).
§4: The civic-harms literature is uneven — not uniformly weak, but its most-quoted number is the weakest link
Four claimed civic harms were audited against the two-source, different-root independence rule, the same standard applied to Auckland's rent literature in GBMT-2.
1. Municipal borrowing costs — SINGLE-LINEAGE (the claim most quoted in advocacy is also the least independently checked)
The canonical paper is Gao, Lee & Murphy, "Financing Dies in Darkness?" (Journal of Financial Economics, 2019/2020): newspaper closures raise offering yields ~5.5bps and secondary-market yields ~6.4bps within 3 years, using a Craigslist-rollout instrument for exogenous closure timing. This is one peer-reviewed paper from one author team.
No independent academic team has replicated or extended this specific bond-yield finding. The only "update" circulating in 2026 policy discourse — the $1.1B/year national figure now being cited across press coverage (Poynter, Ohio Capital Journal, Local Media Association) — is co-authored by Dermot Murphy, one of the three original 2019 authors, together with Rebuild Local News's research director. Independently re-verified: the June 2026 report's authorship is confirmed as Baker (Rebuild Local News) + Murphy, applying the original paper's own coefficient to 2025 municipal-bond data. This is the same lineage re-applying its own estimate through an advocacy organization, not an independent replication — structurally the same failure mode Auckland's rent claim showed (one author's estimate laundered through repeated citation), except here it is a self-citation rather than a third party's, which is arguably a sharper version of the same problem.
2. Turnout / political knowledge / straight-ticket voting — INDEPENDENTLY CORROBORATED (direction, not magnitude)
Three disciplinarily separate teams, no shared coauthors, converge directionally:
- Schulhofer-Wohl & Garrido, Journal of Media Economics (2013) — Cincinnati Post closure, DiD design.
- Shaker, "Dead Newspapers and Citizens' Civic Engagement" (2014) — CPS survey comparison, Seattle/Denver.
- Darr, Hitt & Dunaway, Journal of Communication (2018) — genetic matching across 110 closures nationally; split-ticket voting falls ~1.9%.
Different data, different methods, different disciplines (economics, communication studies, political communication), same direction (less engagement, more nationalized/straight-ticket voting after closure). Exact effect sizes are not comparable across studies, so press coverage that flattens these into one headline number is overclaiming precision the literature doesn't have — but the directional claim itself passes the independence test cleanly.
3. Corporate/government accountability — PARTIALLY CORROBORATED, with a shared-root risk not fully ruled out
Heese, Pérez Cavazos & Peter, Journal of Financial Economics (2022): newspaper closures → facility violations +1.1%, penalties +15.2% over 3 years (federal/state regulatory data). Two later, fully separate teams extended the same "watchdog" mechanism to different domains: Li, Peng & Zhang (JFQA, 2025) find financial-advisor misconduct +19.28% post-closure; Khavis, Ha & Norris (American Journal of Criminal Justice, 2026) find police crime-statistic misreporting (rape reclassification) rises post-closure. Three separate author teams, three journals, three outcome variables — but all plausibly draw on overlapping newspaper-closure event lists, and no team has re-tested another team's specific outcome. The general mechanism is corroborated across domains; no single effect size has been independently re-checked.
4. Polarization / trust — INDEPENDENTLY CORROBORATED on polarization; no causal evidence found on trust
Darr/Hitt/Dunaway (US, 2018) and Ellger, Hilbig, Riaz & Tillmann (British Journal of Political Science, 2024 — Germany, 402 counties, 1980–2009, county fixed-effects) converge on the same mechanism from two countries with no shared authors or data: local-news decline pushes consumption toward national/tabloid outlets, raising partisan/nationalized voting. No causal (as opposed to correlational-survey) evidence was found for a generalized "media trust" effect — that narrower claim should be dropped from the deliverable or clearly marked as correlational-only.
Net assessment for §11 scoring
The strongest, most policy-relevant number in this literature — the municipal-borrowing-cost figure, which is the one every "local news pays for itself" argument reaches for — is also the one that fails the independence test. The whitepaper should either drop the $1.1B/year figure or present it explicitly as "one team's estimate, self-extrapolated," not as a settled fact on par with the turnout/polarization findings, which do pass. This is the same shape of lesson GBMT-2 drew from Auckland: the most-quoted number in a literature is not automatically its best-supported one, and advocacy amplification tends to launder exactly the numbers that need the most scrutiny.
§3 pass 2: one bias sized cleanly, one sized only via existing literature
Phase 0 found Medill's news-desert count has two documented, opposite-direction biases nobody had sized: TV/radio exclusion (undercounts coverage) and ghost papers counted as served (overcounts it).
Bias 1 (TV/radio exclusion) — sized: Mississippi, 5 for 5
Medill's own 2024 methodology states plainly that its database "does not include commercial television or radio stations." [Verification Phase 1, 2026-08-10: quote confirmed verbatim in the 2024 methodology, which continues: "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." The 2025 methodology drops the sentence, so keep this citation pinned to 2024.] Confirmed directly against Mississippi's five Medill-designated zero-outlet counties (Benton, Issaquena, Tunica, Lamar, Jasper — identified from the interactive map's underlying FIPS data, not estimated): all five sit inside a Nielsen DMA served by a full-power commercial network-TV affiliate (Memphis, Jackson, or Hattiesburg-Laurel markets). [Verification Phase 1, 2026-08-10: the five-county list is independently reproduced from Medill's own county API. The market list is corrected: per the FCC's county-to-DMA listing (DA 16-613 App. A), Benton and Tunica are Memphis, Jasper and Lamar are Hattiesburg-Laurel, and Issaquena is Greenwood-Greenville — no county is in the Jackson DMA. The substantive finding is unaffected.] One county (Benton) has a directly-confirmed, dated example of ongoing local coverage — WREG's Benton County tag page — that Medill's "zero outlet" label misses entirely.
This is a real, bounded result for one state (5 of 213 national zero-outlet counties), not a national estimate — a full-scope version would need to run the same DMA-overlay check against all 213 counties, plus verify actual dated local coverage (not just affiliate/DMA membership) systematically rather than by spot-check.
[Verification Phase 2, 2026-08-10 — this check does not size the bias, because its condition holds almost everywhere.] The test asks whether a county sits inside a Nielsen DMA served by a full-power commercial network affiliate. From the FCC's own report to Congress on DMAs (DA 16-613, MB Docket 15-43): "Nielsen delineates television markets by assigning each U.S. county (except for certain counties in Alaska) to one market," and "Nielsen divides the United States into 210 [DMAs]" — a mean of roughly fifteen counties per market, with DBS carriers "required by statute to carry local programming in all 210." The condition is therefore satisfied by essentially every county in the country: the 212 deserts, the 1,525 one-outlet counties, and the ~1,400 Medill rates as well-served alike. A test met by ~100% of the population returns "5 of 5" whether the TV-exclusion bias is large, small or zero.
The FCC also characterises what DMA membership means, verbatim: DMA counties "are clustered geographically around the major metropolitan area or areas in that DMA, where the majority of the market's television stations usually are located," and DMAs "are in part primarily designed to facilitate commercial purposes — such as program acquisition, the sale of advertising, and network compensation." The report exists because DMA membership routinely fails to deliver even in-state programming (the "orphan counties" problem). Membership is a fact about where advertising is sold, not about who covers the county.
What survives is the direction of the bias — commercial TV is excluded and some of it does cover some desert counties, as the one Benton County example shows — and Phase 0's own qualifier, which the whitepaper dropped: the exclusion makes counties look worse "on a coverage-availability basis (though not necessarily on a civic-accountability-reporting basis, which is the harder question TV news often doesn't answer either)." What does not survive is "we sized one of them directly." A sizing needs a measure that discriminates. See steelman-log.md, target A, and method/sources/dma-geography-and-local-journalist-capacity.md.
Bias 2 (ghost papers) — sized only via existing literature, not a fresh case study
A citable national estimate already exists from the same UNC/Abernathy lineage underlying Medill's own project: 1,000–1,500 "ghost newspapers" out of ~7,200 still publishing (≈14-21%) as of 2018, defined as papers that lost "significantly more than half" their newsroom staff since 2004. Medill's stated methodology has no staffing-level check, so this bias is structural, not incidental. Concrete named examples with reporter counts confirm the mechanism (Providence Journal 300+→<100; Denver Post 180+→<70; several Massachusetts papers reported in Medill's own December 2024 reporting as having zero remaining reporters).
This did not reach the same bounded-case-study standard as Bias 1 — no fresh audit was completed cross-checking specific "served" counties (ideally in Mississippi, to let both biases net against each other for one state) against current 2025/2026 ghost-paper status. The 1,000–1,500 figure is 2018-vintage; no newer national ghost-paper recount was found. This is queued, not resolved.
Effect on §3/§10
The two biases now have asymmetric evidentiary weight: Bias 1 has a clean, verified, bounded result; Bias 2 has a citable but dated literature estimate and no fresh check. Full execution should prioritize a same-state ghost-paper audit (ideally Mississippi, to complete the paired comparison) before the whitepaper claims anything about the two biases netting out in either direction.
Effect on the protocol and deviations
Anchor row 5 (civic-harms literature) is no longer blank — see the updated research-inquiry.md anchor table. Rows 8, 12, and 13 remain blank and queued, per deviations-log.md.