Steelman log — GBMT-9 (Media), Verification Protocol Phase 2

Protocol: method/verification-protocol.md, Phase 2 (S1–S5). Check date: 2026-08-10. Filing date (as-of baseline): 2026-08-03. Independence (S1): run in a session and worktree separate from GBMT-9's authorship, its red team, its blind re-score, and its Phase 1 fact-check. Phase 1's corrections had already landed on main (PR #104) and were read first, from media/research/verification-log.md. Targets: the filing's stated conclusions — the printed verdict, the scorecard's winner, the scorecard's refuted row — not its facts. Facts were Phase 1's job.

S2 — The tilt, derived from the filing's own text

Written before any evidence was gathered, from the hypothesis set (media/docs/research-inquiry.md §2, §3), the scorecard (ws11-scorecard.md), and the published conclusions (site/media/index.html).

The filing is deflationary in one consistent direction. The protocol carries 19 numbered seed hypotheses. Twelve predict that a claim someone else makes is wrong or that an instrument will not work: public support is an order of magnitude short (H2.1); no census independent of Medill exists (H2.2); the "platforms took our ad revenue" story is "substantially wrong in composition" (H2.3); the desert count "moves by >2× under defensible alternative inclusion criteria" (H3.1); advocacy effect sizes are "the literature's upper bounds, not its central estimates" (H4.2); bargaining-code money "flow[s] disproportionately to the largest incumbent publishers" (H6.1); Canada's ban establishes which side needs the other (H6.2); referral traffic is "a melting asset" (H6.3); content-and-ranking instruments are foreclosed (H7.1); §230 reform is "largely orthogonal" (H7.2); philanthropy is under 10% of the gap (H8.2); demand-side money "inflates incumbent revenue without adding county-level coverage" (H11.1).

The remaining seven are not deflationary, and it matters which: H3.2 and H3.3 (ghost papers make outlet counts "systematically optimistic" — i.e. the deficit is worse than counted), H4.1, H5.1, H5.2, H8.1 and H11.2. H3.3 is the anti-deflationary seed in the very workstream whose deflationary seed, H3.1, became the filing's published verdict. The bias that would have made the number larger was seeded, never sized, and never headlined; the one that would make it smaller was headlined on a check that turns out not to measure it (target A below).

The conclusions delivered match. The verdict printed in the case file, the filing rail, and the mobile rail is a single sentence — "The desert count is single-root." The domain's headline statistic is the thing the filing chose to put its name on. Part 2's original work is a rebuttal of that statistic ("5 of 5 MS 'desert' counties with real TV coverage"). Part 5 refutes bargaining codes. Part 7 puts philanthropy 55× short. Part 3 concludes that the best-insulated design in the field still lost. The one affirmative endorsement — Part 9's "public media funded outside the annual appropriations cycle is now the clear most-consistent performer" — is for an architecture that has never been implemented anywhere.

The filing knows about one of its tilts and not the other. Its own execution notes say "the hypothesis set above leans toward subsidy skepticism + constitutional pessimism; the steelman owed in the unfashionable direction is the Nordic case." That steelman was run (ws10-findings.md) and came back mixed — and the filing then used the mixed half (Finland's retrenchment) to cap its own winner's durability score. So the declared steelman was built, honestly, and its net effect was to lower the score of the thing it was meant to defend. What was never steelmanned is the direction the filing actually leans hardest: against the measurement, and against the instrument it ranks last.

The three targets

Chosen because each is a stated conclusion that the filing marshalled evidence toward, and each has an opposing position the filing scored down or never entertained.

Standard applied (S3): each target is built from primary sources fetched in this pass — government reviews, statutes, regulations, agency data, the publishers' own methodologies — to the same tier Phase 1 demanded, not from the filing's own dismissals. Where the strongest available evidence for a target is thin, that is recorded as the target failing, not as the filing winning by default.

Sources fetched in this pass, committed to the atlas so the next reader checks the document rather than this log: method/sources/bargaining-codes-au-treasury-review-and-canada-online-news-act.md · method/sources/dma-geography-and-local-journalist-capacity.md · method/sources/bverfg-broadcasting-fee-and-cpb-rescission-impact.md. Each records, in its own body, where the document cuts against the argument citing it.

S4 — Verdicts

Target A — "The desert count overstates the deficit, and we sized the bias"

The filing's position

The verdict printed in the case file, the filing rail and the mobile rail is one sentence: "The desert count is single-root." Part 2 carries the stamp "Single-root, partly sized," the stat block "5 of 5 — MS 'desert' counties with real TV coverage," and the plain-talk line: "We checked one of those blind spots in one state and found it real: every 'desert' county we checked actually had a local TV station covering it."

The opposing case, built

1. The Mississippi test cannot discriminate, because its condition holds almost everywhere. The test is whether a county "sits inside a Nielsen DMA served by a full-power commercial network-TV affiliate." From the FCC's own report to Congress on DMAs (DA 16-613): "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 about fifteen counties per market. DBS carriers are "required by statute to carry local programming in all 210 DMAs," and only twelve DMAs have no out-of-market full-power signals. So the condition is satisfied by essentially every county in the United States: the 212 deserts, the 1,525 one-outlet counties, and the roughly 1,400 counties Medill rates as well-served alike.

A test whose condition is met by ~100% of the population returns "5 of 5" whether the TV-exclusion bias is large, small, or zero. It is not evidence about the bias; it is a restatement of how Nielsen partitions the country. The filing presents it as the one bias it "sized," and as the original work behind its published verdict.

2. The FCC says a DMA is an advertising region, and its own report exists because DMA membership does not deliver local coverage. Verbatim, DA 16-613 ¶ 6: 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 was written because residents of many counties in a DMA "may not always receive the programming of broadcast television stations located in the state in which they live" — the orphan-county problem. Membership in a market is a statement about where advertising is sold and viewing was measured, not about whether anyone covers the county's board of supervisors.

The filing's own single county-level example concedes the gap without drawing the conclusion: Benton County's evidence is a station tag page, which Phase 1 downgraded to "a Memphis CBS affiliate that has filed dated stories from the county," could not retrieve after 2020, and which the public page itself now hedges as "thinner evidence than a beat." One tag page in one county is what remains of the sizing.

3. On a capacity measure from an independent lineage that includes TV and radio, the deficit is far larger than 212 counties. The Local Journalist Index 2026 (Muck Rack + Rebuild Local News) counts journalist-equivalents across "websites, TV, radio, podcasts, newsletters and more" — so it does not carry Medill's commercial-broadcast exclusion at all. Its published findings: 7.8 LJEs per 100,000, "an 81% decline since 2002"; "209 million people – living in 70% of the counties – are in areas below the already anemic national average"; and, from a sweep of 4.2 million Q1-2026 articles, "nearly 77% of U.S. counties produced zero education articles mentioning a local community during those three months. Some 76% had no local articles about health."

If the objection to Medill is that it misses commercial broadcast coverage, the correction is not "the deficit is smaller than 212 counties." It is that on a measure without that blind spot, roughly three-quarters of American counties produced no local education or health reporting in a quarter. The filing has this source. ws15-capacity-proxy.md reads it accurately, including its limits. It reaches no public page.

4. The filing's own protocol says outlet presence is the wrong quantity, and the whitepaper argues from outlet presence anyway. research-inquiry.md §1.3: "a count quoted without its inclusion criteria is not a finding," and §3 instructs "measure coverage capacity (reporters on the ground, stories filed about local government), not outlet counts, wherever the data permits." Phase 0 §1 carried the necessary qualifier on the TV bias explicitly: the exclusion "almost certainly makes some 'zero-outlet' counties look worse than they are 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)." The parenthesis is the whole objection, it was in the record from day one, and it did not survive the trip to the public page.

5. The offsetting bias points the other way and is bigger, and the filing says it cannot net them. Ghost papers: 1,000–1,500 of ~7,200 still-publishing papers (14–21%), inflating the served column. Medill's 2025 methodology verifies that an outlet publishes and updates; it sets no staffing threshold and publishes no per-outlet newsroom size. Medill's own Watch List adds 250 single-source counties at greater than 40% modelled risk of becoming deserts within the decade. ws16-ghost-paper-audit.md records that the netting audit was attempted and returned INDETERMINATE. With the larger bias unsized and the smaller one measured by a test that cannot discriminate, "partly sized" describes an arithmetic that was never done.

S4 verdict on A — PARTIALLY SURVIVES

The steelman does not overturn the single-root finding. Phase 1 independently strengthened it: every candidate alternative census either measures something else (LJI measures capacity) or feeds Medill (Editor & Publisher and AAM are named inputs). That verdict is right and is the filing's to keep.

What the steelman takes is the sizing claim and its direction. "We sized one of them directly" is not supported, because the test used cannot distinguish a desert county from any other county. And the filing's public framing — one lineage, two biases, we checked one and the deserts have TV — reads as though the corrected number would be smaller. On the filing's own preferred quantity, and on a source in its own record with none of Medill's broadcast blind spot, the corrected number is much larger.

Target B — "The one architecture the record actively refutes"

The filing's position

Part 9: the bargaining code "is the one architecture the record actively refutes on four of five axes, not merely ranks last." ws11-rescore-log.md sharpens it: "not merely last, it is the only instrument the record actively argues against," with each of its 1s "resting on peer-reviewed or platform-measured evidence rather than absence of evidence." Row 2 scores 1/1/1/1/3 — neutral mean 1.40, alone at the bottom by a full point.

What the four 1-scores actually rest on

The opposing case, built

1. The load-bearing evidence is not merely uncited — it is unobtainable in principle from the source the filing wanted. Phase 1 established that Brevini & Ward, Media International Australia (2023) does not exist: Brevini's 2023 bargaining-code paper is sole-authored, in Javnost — The Public 30(2), and is paywalled and unread. This pass adds the structural reason no such figure can be audited. The Australian Treasury's own statutory review, verbatim: "the review has not been provided with the details of these agreements by either digital platforms or the relevant news businesses, despite highlighting the importance of this information in the consultation paper." Deal values are confidential by design. Any distributional split of Australian code money — 90/10 or otherwise — is somebody's estimate. Four of the filing's five 1-scores on this row are anchored to a number that its own government could not obtain.

2. The government evaluation of the instrument reaches the opposite conclusion, and documents capacity added in undercovered places. Treasury's review, verbatim: "we consider it is reasonable to conclude that the Code has been a success to date"; "Google and Meta have reached over 30 agreements with a broad range of news businesses, both large and small, and in metropolitan and regional areas"; "at least some of these agreements have enabled news businesses to, in particular, employ additional journalists." Google's own count was 60 agreements across 183 mastheads. Country Press Australia — the collective of small regional publishers — holds agreements with both platforms under an ACCC collective-bargaining authorisation; the Minderoo Foundation signed for a further 24 publishers.

And the A-axis question specifically. The ABC, verbatim in its submission: it "appointed 57 regional positions, including reporters in 19 locations, 10 of which did not previously have them." Guardian Australia: "our newsroom has grown by over 40 journalists." Under the A anchor as written — "1 = no capacity added, or added only to already-strong outlets; 5 = large, demonstrated add specifically in undercovered areas" — reporters placed in ten locations that previously had none is the opposite of anchor 1.

3. The distribution problem the filing says the instrument would have to solve was solved by regulation two and a half years before the filing. ws12-sequencing.md point 5: the code "needs a fundamentally different distribution mechanism (a mandatory desert/small-outlet allocation floor, not a negotiate-your-own-deal structure) before it is worth the fragility risk." Canada's Online News Act Application and Exemption Regulations, SOR/2023-276, registered 15 December 2023, is that mechanism, in binding text:

• s.9(2) sets the price by formula rather than by bargaining: compensation qualifies "if and only if … in accordance with the formula $100 million × CPI x ÷ CPI 2023."
• s.10(2) sets the key: distribution "having regard to the number of full-time equivalent employees … employed … for the purpose of producing … original news content."
• s.10(3) caps the incumbents: "No more than 30% … to news businesses — other than the Canadian Broadcasting Corporation — that carry on a programming undertaking … and no more than 7% … to the Canadian Broadcasting Corporation."
• s.7 requires a commitment to spend "a majority of the monetary compensation … to support the production of local, regional and national news content."
• s.4(1)(a) and s.10(1)(b) require a 60-day open call to every eligible news business and admission of any respondent "at any time."
• s.8 writes journalistic-independence protections into the exemption criteria — no retaliation for editorial decisions, no intervention in the editorial process.

What that produced, from the CRTC's exemption decision (2024-262), verbatim: "Google received responses from more than 600 news businesses representing 1435 news outlets … and submitted that it did not reject any news business that submitted an attestation." The Act's own criteria require the agreements to "support a significant portion of independent local news businesses" and to reach "Indigenous new outlets, and official language minority community … news outlets, in local and regional markets across Canada."

4. Australia redesigned in the same direction, and finalised it on this filing's filing date. The News Bargaining Incentive, joint ministerial release of 3 August 2026 — GBMT-9's own filing date — verbatim: "increasing the distribution scheme loading for regional-based journalists, small to medium publishers and media servicing underrepresented communities from 10 to 20 per cent"; "establishing a grants program for small publishers and start-ups that have an annual revenue of less than $150,000"; broadening "the definition of journalists to include more essential news roles" and "the inclusion of freelancers." The Minister: "An important change is the doubling of the distribution loading for smaller and regional publishers."

The filing's moving-targets-log.md caught this on 2026-08-06 and recorded, correctly, that these are "post-consultation design changes that respond directly to the incumbent-capture critique this filing documented." The note stops there. The scorecard did not move, and Part 9 continued to say the record "actively refutes" the architecture.

5. The "completed failed test" had two subjects and one verdict is recorded. B=1's basis is Meta's exit. Google did not exit: it holds 23 Australian agreements, and in Canada it pays a regulator-set, inflation-indexed $100 million a year under a five-year exemption order. The honest description of the Canadian natural experiment is that one platform walked and one paid under a code that set the price by regulation — which is a finding about which design features hold a platform at the table, not a finding that the instrument family failed.

S4 verdict on B — THE STEELMAN WINS

"The one architecture the record actively refutes rather than merely ranks last" does not survive. The claim's strength came from the assertion that row 2's 1s rest on evidence rather than on absence of evidence; three of the four rest on a figure from a paper that does not exist, drawn from deal values the issuing government states it could not obtain, and the fourth is contradicted by that government's published evaluation.

The narrower claims survive and should be kept: platforms can exit at low cost to themselves, and Meta did; the 2021 Australian design left small publishers dependent on collective bargaining to get to the table at all; the US reform coalition's own opposition to the JCPA shape is real and correctly reported. What does not survive is the generalisation from one design, in one country, in one year, to the instrument family — at the exact moment two of the three countries that have tried it were legislating the distribution floor the filing says it would need.

Target C — "The clear most-consistent performer"

The filing's position

Part 9: "Public media funded outside the annual appropriations cycle is now the clear most-consistent performer — its distribution record to some 1,500 local and tribal stations is the best-evidenced reach in the field, its durability still capped by Finland's counter-example." Neutral mean 3.6, nothing else above 3.0. Row 4b scores 4/3/5/3/3.

The opposing case, built

1. The B axis punishes the only architecture in the filing with disbursement data and spares the one that has never existed. The blind re-score applied the anchor hard and said so: "a brand-new instrument cannot score 3." It moved the payroll credit to B=1 ("NY/IL/NM credits enacted 2024-26; no track record exists"), vouchers to 1 ("Never implemented anywhere"), philanthropy to 2. 4b — never implemented anywhere either, as its own basis text concedes — kept 3. Phase 1 flagged this and referred it here.

Meanwhile the payroll credit is the one row with a completed, repeated, public result. Northwestern's Local News Initiative, 8 July 2026: "Fifty-five Illinois local news organizations operating 130 outlets have received $4.3 million in state tax credits so far this year," which "matches the total disbursed in all of 2025, the first year of the program"; "Nonprofit outlets received 40% of the money, up from around a third in 2025"; "Half the funding went to organizations outside the Chicago area, a similar rate to last year." Two completed cycles, four states enacted (CA, NY, IL, NM), and a live federal bill shape. That is a thin record, but it is a record; 4b's is empty.

2. The negative claim capping 4b's B is false, and the correct counter-example runs the other way. ws11-scorecard.md: "No project-sourced evidence of a non-appropriated public-media mechanism with a completed, successful durability test exists." Germany's Rundfunkbeitrag is one. A per-household levy set on an independent commission's technical evaluation and alterable only by unanimity among sixteen state governments; in 2020 one state, Saxony-Anhalt, refused to ratify the commission-recommended increase and blocked it nationwide. The Federal Constitutional Court, Order of 20 July 2021, verbatim: "By failing to approve the First State Treaty to Amend State Media Treaties … the Land Saxony-Anhalt violated the freedom of broadcasting under Article 5(1) second sentence of the Basic Law," and it ordered the increase "provisionally applicable with effect from 20 July 2021."

The filing let Finland — also non-US, also a design comparator — set the cap on 4b's durability. Symmetric treatment requires Germany to count too. It is the larger system, the test was against a genuine hostile majority, and the mechanism won.

But the steelman must concede the reason it won. The Court's holding rests on Article 5(1) of the Basic Law, which German doctrine reads as an affirmative constitutional duty on the state to fund broadcasting adequately. The First Amendment imposes no such duty. What transfers is the design — an arm's-length commission, a household levy, a unanimity rule — not the enforcement. And the German mechanism is now being eroded in the Finnish shape without being repealed: the fee has stood at EUR 18.36 since 2021, the Länder declined the commission's 2024 recommendation, and ARD and ZDF's resulting complaints are undecided. Five years of nominal freeze under inflation is a real-terms cut delivered by political refusal. (That last sequence is from German trade press; only the 2021 order is verified at primary tier here.)

So the honest B range for 4b is wide in both directions — 1 if only US implementation counts, 5 if the German test counts the way Finland's does — and the published 3 sits in the middle for reasons the record never states.

3. The cell actually carrying 4b's lead is C, not B, and C is inherited. C=5 was set by the blind re-score on the ground that "Distribution to ~1,500 stations including the highest-dependency tribal stations is anchor-5 'demonstrated in practice'." That demonstration belongs to the appropriated variant, #4 — the institution the filing's own Part 3 says has dissolved. CPB's own record, 9 October 2025, verbatim: CPB "has been forced to reduce its staff by 70 percent," and "Without additional congressional action, the federal partnership that sustains 1,500 local public media stations will end — along with the universal service guarantee." New Jersey PBS "will cease operations in 2026"; WPSU will "cease operations by June 30, 2026"; Mississippi Public Broadcasting "plans to drop all PBS and NPR programming"; KZYK faces an "existential threat"; KBRW, KYUK and KUCB are "at risk of going dark."

Being fair to the filing: this is CPB's own advocacy document, and it is mostly a roster of cuts, not closures. The network is degraded, not gone. But C=5's anchor is "demonstrated to reach small/local/nonprofit outlets in practice," and the practice being credited is a different architecture's, run through pooled infrastructure — interconnection, emergency alerting, royalties — that CPB says is ending. A dedicated-fee mechanism would have to rebuild the grantee network before it could demonstrate anything.

S4 verdict on C — PARTIALLY SURVIVES, and it relocates the weak cell

4b's rank is robust. Under the neutral weighting it stays top across every single-cell change tested (3.2 at B=1, 4.0 at B=5, 3.2 at C=3). What does not survive is the account of why it leads and how much room the finding has, and Phase 1 named the wrong cell: the honesty box tells readers the durability score is the soft spot, when the arithmetic says the diversity score is.

S5 — Scorecard sensitivity

Neutral-weight means over the published v3 matrix, and over the one weighting vector that exists anywhere in the committed record — pluralism-first, A15/B10/C60/D5/E10, from ws11-red-team-log.md. Recomputed from the matrix in this branch; the published board reproduces exactly (4b 3.60; #4, #6, #8 tied at 3.00; nothing else above 3.0).

As published

A finding that falls straight out of this table. ws11-red-team-log.md attack 3 states that the pluralism-first weighting "does change the leader." Against the v3 matrix it does not: 4b leads on both weightings, and by a wider margin under pluralism-first than under neutral weights, because C=5 is worth 60% of the score there. That claim was true of the pre-re-score board, where #8 led; it has been carried forward unrevised. Phase 1 found the five weightings uncommitted and left the claim UNVERIFIABLE — the one defined vector can now be computed and is committed here, which closes that half.

Sensitivity 1 — row 2 on the evidence that survives

Rescoring the bargaining code against the anchors, using only what this pass could verify at primary tier:

• A: 1 → 3. Anchor 1 is "no capacity added, or added only to already-strong outlets." Treasury's review records reporters placed in 19 locations, 10 previously unserved, plus 40+ Guardian journalists and Country Press/Minderoo collective deals. A=3, "moderate, real add but not desert-targeted," is the defensible floor; A=4 is arguable and not claimed.
• B: 1 → 2. The completed test is split: one platform exited, one paid a regulator-set price under a five-year exemption. A partial real record is B=2 by the same logic the re-score used to hold #6 at 2.
• C: 1 → 3. Anchor 1 requires being "documented to concentrate resources in incumbents." The only documentation is uncertifiable, and SOR/2023-276 ss.10(2)–(3) plus the NBI's doubled small/regional loading are documented allocation floors pointing the other way. Neutral is the rule-consistent score; C=4 is arguable on Canada's formula and not claimed.
• D: 1 → 3. The antitrust-exemption structure is a real concentration mechanism and is documented (the ACCC authorisations). So is the offsetting statutory cap. This is the below-neutral-without-evidence pattern the re-score corrected in five other cells; the same rule gives 3.
• E: 3. Unchanged; no trust evidence either way.

Row 2 rescored: 3/2/3/3/3, neutral mean 2.80, pluralism-first 2.90. Conservative variant (A3/B1/C3/D2/E3): 2.40 / 2.75.

Effect: the bargaining code moves from last by a full point to a five-way tie at 2.80 with #1, #3 and #7 — or, on the conservative variant, to ninth, above cash and managed transition. No ordering above it changes. 4b still leads; the comparators still sit at the bottom. What changes is the sentence "the one architecture the record actively refutes," which the arithmetic no longer supports under either variant.

Sensitivity 2 — 4b's two contested cells

Three things follow, and the second is the one the filing does not currently say.

1. The headline survives every single-cell change. On neutral weights 4b leads whether its durability cell is 1, 3 or 5. The honesty box's disclosure — "corrected to the floor, its lead over the field roughly halves, though it still leads" — is accurate and the arithmetic backs it (3.60 → 3.20 against a 3.00 field).
2. The load-bearing cell is C, not B. Under the only weighting vector the record defines, moving C from 5 to 3 alone drops 4b from first (4.35) to fourth (3.15), behind #4, #6 and #1. Nothing on the public pages tells a reader that the winner's position depends on a diversity score inherited from the institution Part 3 is about.
3. Both contested cells at once flips the board. At 4/1/3/3/3 the leader becomes a three-way tie between #4, #6 and #8 at 3.00 on neutral weights, and #4 alone at 4.05 under pluralism-first. That is the only combination tested that changes the answer to the filing's title question — and it is exactly the combination where 4b is credited with nothing it has not itself demonstrated.

Does the recommendation move?

No, on the record as it stands, and the filing should say why rather than leave it implied. 4b leads under every single-cell sensitivity. The two-cell case that unseats it is a defensible reading, not the established one — and the same strictness applied evenly would also move row 2 up and would have to weigh Germany's completed test, which pushes 4b's durability the other way. The honest published statement is that 4b leads on every reading of its durability cell and loses the lead only if its diversity cell is also read as inherited — which is a narrower and more checkable claim than "clear most-consistent performer."

What the steelman could not overturn

Recorded per S4: a conclusion that has beaten a real challenge should say what it beat.

• The single-root finding on the desert count. Every alternative national measure searched either counts a different quantity or is an input to Medill. Phase 1 reached the same result from the other direction. This survives intact and is the filing's strongest original contribution.
• CPB's dissolution as the design lesson. Nothing found weakens it. Germany's case, which is the best available counter-evidence that insulated public-media funding can survive a hostile majority, survived on a constitutional duty to fund that the United States does not have — which sharpens the filing's point rather than blunting it.
• Philanthropy's ~55× shortfall. No source found moves it by an order of magnitude, and the direction of travel in the record (Press Forward now at 130 funders and "$400M+ invested") does not close a gap of that size.
• Meta's exit as a real fragility finding. It is the strongest surviving objection to the bargaining code and should stay in the filing's own words: a platform can leave at negligible cost to itself, and one did.
• The constitutional screen's modesty. Part 6 and the honesty box already concede the screen did real gating work on one candidate only. Nothing found argues it should claim more.

Byproduct corrections found while building the steelman

Not steelman findings; recorded because the pass turned them up and Phase 1 did not reach them.

• Part 7, "most of it landing outside Chicago" (Illinois credit). Northwestern's own July 2026 article says "Half the funding went to organizations outside the Chicago area, a similar rate to last year." "Most" overstates "half." Replace with "about half of it landing outside Chicago."
• Part 7, "$4M Illinois's tax credit, actually disbursed, 2025" and "a real $4 million in its first completed year (2025)."** The same article puts 2026 at **$4.3 million across 55 organizations and 130 outlets, "matching the total disbursed in all of 2025," with nonprofits at 40%, up from about a third. The public page has the 2025 figure and not the second completed cycle, which is the stronger fact.
• ws11-red-team-log.md attack 3's "the fifth weighting does change the leader" is stale against the v3 matrix. Corrected in the log with the computed vector.
• Six media source pages Phase 1 flagged as missing (finding B-66) are now published. scripts/build-source-pages.py was run for real in this pass; the builder's stale From the desk nav label — which Phase 1 recorded as a reason not to run it — has since been fixed, and the run is safe. verification-log.md's "known limitation" paragraph is updated to say so.
• Cross-filing drift in the generated pages, not fixed here. Running the builder also regenerates childcare, elder-care and drugs pages, and childcare has four pages that are stale against its own markdown (sources/all, deviations-log — missing entries 11 and 12 — scorecard-scales, steelman-feasibility). Reverted rather than included, to keep this PR inside GBMT-9 and out of the way of concurrent childcare sessions. Someone should run the builder for childcare on its own branch.