GUBMENTPlain talk · policy frontier
Filings / Media / Sources / Deviations Log (GBMT-9)
GBMT-9 · Research record · No. 9

Deviations Log (GBMT-9)

media/research/deviations-log.md
This is a working research document from the media filing, published as written — including the parts later corrected. It is the underlying record for Whitepaper No. 9, not a summary of it.

Freeze point: media/docs/research-inquiry.md as committed at the start of Phase 0 (2026-08-03). Every deviation from protocol-specified method is logged here with a reason and its effect on findings. Per M6, this log is part of the record, not an appendix to it.

# Date Deviation Reason Effect on findings
1 2026-08-03 Anchor rows 5 (civic-harms literature independence audit), 8 (broadcast ownership cap/Nexstar-Tegna), 12 (media trust survey), 13 (philanthropy-vs-revenue ratio) not attempted Phase 0 time budget went to the seven ★ rows plus two bonus non-starred rows (6, 7) closely coupled to the starred C-18/Australia precedents; these four rows are not load-bearing for the go/no-go decision Rows left blank, not guessed. Row 5 in particular is queued as a priority for full execution — see phase0-findings §8 re-ranking, since it is one of the two named kill-condition tests in the protocol's own execution notes and was not resolved either way
2 2026-08-03 Anchor verification for all seven rows delegated to five parallel subagents; results QA-reviewed by the primary session before entry Search fan-out is parallelisable across independent anchor clusters; the two-source and root-trace rules were carried into each brief verbatim Three of the most surprising/load-bearing claims (CPB's dissolution, the Rescissions Act's exact figures, and Google+Meta's digital-ad share falling below 50%) were independently re-verified by the primary session via direct search before being written into the permanent record, per the project's QA-before-committing discipline. No agent figure was entered without at least this spot-check tier of verification; none of the three re-checks contradicted the subagent's report
3 2026-08-03 Census CBP national bulk files (cbp{yy}us.zip) 404 for pre-2014 vintages at the modern URL path (www2.census.gov/programs-surveys/cbp/datasets/{year}/) though the same pattern works for 2021+ Pre-2014 CBP files use a different naming/location convention not resolved within Phase 0's time budget Blocks extending the newspaper-employment baseline back to the pre-2008/mid-2000s peak using CBP as a Census-native source. Logged in media/baseline/incompatibility-log.md #2
4 2026-08-03 BLS QCEW's open CSV-slice API (data.bls.gov/cew/data/api/{year}/a/industry/{naics}.csv) returns 404 for years before 2014 for NAICS 511110, despite QCEW's underlying series extending to 1990 The open API's coverage window is narrower than the full QCEW historical series; reaching the 1990s-2000s data would require a different QCEW access route (singlefile/historical annual-averages files) not attempted in Phase 0 The committed baseline pipeline covers 2014-2025 only. The mid-2000s peak figures cited in anchors 1 and 3 rest on Pew/NAA secondary literature, not a direct admin-data pull — stated explicitly in phase0-findings §7 rather than presented as pipeline-verified. Logged in media/baseline/incompatibility-log.md #3
5 2026-08-03 Several source URLs returned fetch errors during subagent research: congress.gov (403 on CRS Legal Sidebar LSB11224 and CRS Report R48545), cpb.org (DNS resolution failures throughout), SAGE journals (403 on the Neff & Pickard paper itself), Brevini's Media International Australia article (403) Paywalls and bot-blocking on primary-source sites; not every primary document was reachable within session tooling These claims rest on secondary citation (press coverage quoting CPB, search-snippet-level access to the CRS reports, secondary summaries of the paywalled academic articles) rather than direct primary-text confirmation. Flagged per-claim in phase0-findings and in each subagent's caveats section; none of the flagged claims were load-bearing for the GO verdict, but full execution should retry primary access (e.g., via institutional access or a different fetch path) before the whitepaper cites them
6 2026-08-03 The two anchor-table rows not originally starred (6, 7 — Canada and Australia bargaining codes) were verified together in Phase 0 anyway, beyond the minimum ★-row scope Both are tightly coupled precedents for the same §6/§9 workstream question (does platform-publisher bargaining actually fund local news), and verifying them together was more efficient than splitting across sessions Produced the report's strongest bonus finding — see phase0-findings §5 — ahead of full §6/§9 execution, at the cost of not spending that time on rows 5/8/12/13 instead (see deviation #1)
7 2026-08-03 Post-Phase-0, executed anchor row 5 (civic-harms literature audit) and a §3 bias-sizing pass, picking up the two items ranked highest in phase0-findings §8, rather than working strictly workstream-by-workstream in protocol order §4's kill-condition test was the single highest-priority open item per the protocol's own execution notes; §3's bias-sizing was the highest-value original work per the re-ranking See ws03-04-findings.md. Row 5 filled in the anchor table; rows 8, 12, 13 remain blank
8 2026-08-03 §3 pass 2's Bias 2 (ghost papers) sizing did not reach a fresh bounded case study — it relied on an existing 2018-vintage national estimate (UNC/Abernathy, 1,000-1,500 ghost papers) and named examples, rather than a same-state audit paired with Bias 1's Mississippi check Time budget; a fresh staffing audit of specific Mississippi "served" counties' papers was not completed within the session Bias 1 (TV/radio exclusion) has a clean, verified, bounded result (5 of 5 MS zero-outlet counties fall in a TV DMA); Bias 2 rests on a dated literature estimate. The two biases cannot yet be netted against each other for one state — queued for the next pass, per ws03-04-findings
9 2026-08-03 Executed §5 (ownership), §2/§8 (public spending/philanthropy), §9 (fault-lines/trust), §10 (precedents/steelman) as a single parallel batch, closing out all remaining anchor rows (8, 12, 13) These were the remaining workstreams needed before §11's scorecard could be built on a complete evidence base, per the user's request to complete full execution See ws05-findings, ws02-08-findings, ws09-findings, ws10-findings. All 13 anchor rows now resolved
10 2026-08-03 Several primary sources in this batch were unreachable: Kielbowicz's PRC postal-history paper, a comparative-Nordic-media-systems PDF, and the Cogitatio full-text PDF all failed to extract; parliament.uk and cpb.org returned 403/DNS errors PDF encoding issues and bot-blocking, consistent with earlier deviations (#5) Claims resting on these rest on secondary/abstract-level citation, flagged inline in ws10-findings and ws05-findings. None were load-bearing for a GO/NO-GO call at this stage — full execution's whitepaper stage should retry primary access
11 2026-08-03 The §5 broadcast-ownership finding (FCC vote scheduled Aug 6 2026) is time-sensitive to the exact date of this research (Aug 3 2026) The FCC's own published schedule, independently re-verified Anchor row 8 and ws05-findings are accurate as of Aug 3 2026 but should be re-checked against the actual Aug 6 vote outcome before any whitepaper publication — flagged explicitly rather than presented as settled
12 2026-08-04 Independent blind re-score executed and reconciled (ws11-rescore-log.md): 24 cells materially reconciled — the durability axis was still leaking design labels after the first red team explicitly banned that; five below-neutral-without-evidence violations; comparator cells re-grounded on verified trajectory evidence M6's standing second-scorer requirement; blinding procedural, disclosed. One scorer input (the CA Google deal) was already stale against the same-day moving-targets correction — the moving-target problem operating inside the re-score itself, disclosed 4b confirmed as most-consistent performer (strengthened); the bargaining code reclassified from "stable bottom" to affirmatively multi-axis-refuted; site page, sidebar, and report PDF updated
13 2026-08-06 Late DOJ Google ad-tech remedy status pass started without remedy-specific M3 conditions The prior scorecard openly recorded this as an unexecuted workstream §13 is limited to status and the missing outcome boundary; it does not rescore or update the active site
14 2026-08-06 A prospective demand-side voucher pass was completed after the scorecard The new randomized evidence tests subscription use but not the separate outlet-allocation claim; a late score change would bypass the required re-score §14 corrects the evidence basis for architecture #5 and retains allocation as indeterminate pending payment/coverage data
15 2026-08-06 A current county-level capacity proxy was located after the ghost-paper audit limitation record LJI measures modelled county capacity and article output, not outlet staffing or ghost-masthead status §15 adds a current screening tool but preserves the outlet-level ghost-paper count and bias-netting result as indeterminate
16 2026-08-06 The intended same-state ghost-paper audit did not produce a count Current web results included a same-named Texas newspaper and conflicting sources for the Virginia pilot; neither supports assigning a current staff count §16 records the discrepancy and a reproducible future audit design; the whitepaper's 2018-only limitation remains unchanged
17 2026-08-10 Verification Protocol Phase 1 (fact-check) run against the whole filing by an independent session — 556 claims extracted across both public pages and all 18 research documents, every one adjudicated against primary sources fetched in that pass The protocol's house-wide vulnerability finding: no filing had been checked against primary law, and every legal claim here had been verified at the secondary tier 450 CONFIRMED, 27 CORRECTED, 37 OVERSTATED, 35 UNVERIFIABLE, 7 STALE. Load-bearing corrections: Moody's platform-curation reasoning commanded five justices, not six; CPB's forward funding came from the 1975 and 1978 financing acts, not the 1967 Act; Canada's "-64% national" figure is not in the cited MEO report; Australia's ~90% figure is cited to a paper that does not exist (Brevini is sole-authored, in Javnost, not Media International Australia); the desert counts are 212/1,525, not 213/1,524; the Nordic steelman's viewpoint-capture claim is not in the peer-reviewed paper cited for it; the published scorecard omitted the M7 cash comparator and the five weightings it advertises are in no committed file. Full ledger in media/research/verification-log.md. [Same-day follow-up: the first commit landed these corrections on the whitepaper and the sources digest but not on the generated sibling pages, leaving the withdrawn "-64% national" and "6-justice majority" claims live in four places — one under a "Confirmed" stamp. Propagated, along with two self-corrections to the log (the Part III–A/III–B attribution, and the deviations/commit/branch counts). See verification-log.md §0.]
18 2026-08-10 Verification Protocol Phase 2 (steelman) run against the filing's three stated conclusions by a session independent of the filing, its red team, its blind re-score and its Phase 1 fact-check. Tilt and targets derived from the filing's own hypothesis set and conclusions and committed before any opposing evidence was fetched S1–S5. The filing's declared steelman (the Nordic case, §10) was built honestly but ran on the direction the filing already knew it leaned; the directions it leans hardest — against the measurement, and against the instrument it ranks last — had never been steelmanned One target wins, two partially survive. B (bargaining code) — steelman wins: three of row 2's four 1-scores rest on a figure from a nonexistent paper drawn from deal values the Australian Treasury states it could not obtain, and are contradicted by that Treasury's own review; Canada's SOR/2023-276 puts a mandatory per-FTE-journalist distribution formula with ≤30%/≤7% incumbent caps into law, and Australia doubled its regional/small-publisher loading on this filing's own filing date. "The one architecture the record actively refutes" withdrawn. A (desert count) — partially survives: the single-root verdict stands; "we sized one of them directly" does not, because every US county sits in a DMA. C (the winner) — partially survives: 4b leads under every single-cell reading, but the load-bearing cell is C (inherited from the dissolved CPB), not B, and the negative claim capping B is false — Germany's Rundfunkbeitrag has a completed, court-enforced hostile-majority test, on a constitutional footing the US lacks. Full account in media/research/steelman-log.md; three primary sources committed under method/sources/
← All Media research documents Sources digest Read the whitepaper