Today on The Primary Source: the expanded Treasury bond buyback we've been tracking faced its first major market test and failed to suppress yields. We also review a class-action lawsuit over Anthropic's subscription math, a new federal scholarship mechanism bringing real dollars to Jewish day schools, and a critical Lean 4 bug that briefly made Fermat's Last Theorem spuriously provable.
A class-action lawsuit filed September 8 accuses Anthropic of misleading advertising on Claude Max subscriptions, alleging the '5x' ($100/month) and '20x' ($200/month) multipliers apply only per-session while a separate weekly usage cap — introduced in late July 2025 — goes undisclosed at purchase. One Reddit-sourced data point claims the $200 plan delivers only 1.7× effective usage versus the $100 plan despite costing twice as much. Plaintiffs argue that material constraints require clicking through multiple hyperlinks to support documentation rather than being presented at the point of sale.
Why it matters
If courts find that hyperlinked fine-print disclosure is insufficient for subscription terms this opaque, Anthropic and every major AI provider will face pressure to surface session-level versus period-level limits in a single, unambiguous place at checkout. For power users running agentic workflows, the lawsuit also spotlights something practically useful right now: the gap between the headline multiplier and the weekly cap is the right question to ask before upgrading any AI subscription tier, not the nominal price difference.
Continuing the sequence of Claude Code prompt-cache repairs we tracked following Fable 5.1's release, Anthropic shipped v2.1.267 on September 9. The update adds a maxEffortLevel configuration that caps reasoning effort across Bedrock, Vertex, and Foundry deployments, and fixes multiple cache regressions: resumed sessions were rewriting tool definitions and breaking cache reuse; MCP reconnections were re-registering tools; and model-change events were rewriting the full tool list. Large tool output truncation at the 1 GB cap is also corrected.
Why it matters
The volume of cache-fix commits in this single release — resumed-session tool re-sending, MCP re-rendering, model-change tool-list rewrites — suggests the team identified a widespread regression across session-resumption patterns rather than isolated edge cases. For anyone running multi-day agent tasks or MCP-connected workflows on Claude Max, updating to v2.1.267 is load-bearing: silent cache misses in resumed sessions were burning tokens and latency on every turn without surfacing an error.
Following the benchmark performance gap we noted yesterday between GPT-6 Astra and Claude Fable 5.1, Artificial Analysis released Intelligence Index v4.3 on September 7 after two prior revisions in four days (v4.1.1 on September 3, v4.2 on September 4), landing both models at identical 53-point composite scores. The revision trajectory — Astra moving 61→55→53, Fable moving 66→57→53 — was driven by removing GPQA Diamond, adding Terminal-Bench v4.0 and AutomationBench-AA, and increasing private test-set weighting from 20% to 45%. At the tied v4.3 score, independent analysis finds Astra costs $3.26 per task while Fable 5.1 costs $7.63 — a 57% per-task advantage for Astra — because Fable consumes roughly 190 million output tokens against Astra's 60 million on equivalent benchmark suites.
Why it matters
Three revisions in one week to a benchmark that is widely used for model selection decisions is a warning about the stability of composite scores as purchasing signals. The 57% per-task cost gap at identical composite scores is the more durable finding: Astra leads on Terminal-Bench (59.1% vs. 52.0%) and AutomationBench, while Fable 5.1 leads on private knowledge-work evals. For agentic coding and terminal-based engineering workflows, the per-task math favors Astra at parity pricing; for long-horizon research and regulated document work, Fable's verbosity may be productive rather than wasteful. The right question for any Claude Max subscriber is whether their specific workload mix justifies Fable's 3× output-token usage.
DeepSeek's MIT-licensed deepseek-harness repository — which we tracked when it was first open-sourced in mid-August — introduces an agent runtime built on an everything-is-a-plugin architecture that ships sandboxed file/shell execution, append-only session persistence (full conversation restoration after restart), multi-agent subagent invocation, auto-schema tool calling with approval gates, token metering, OpenTelemetry observability, and hot-reload in the core rather than leaving them to integrators. The three-layer architecture (Application, Core Subsystems, Cordis Plugin Framework) allows swapping any component without forking; deployable via npx @deepseek-ai/dsh web for immediate Web UI or headless CLI. The project explicitly benchmarks against LangGraph, AutoGen, Dify, and n8n in a comparison matrix.
Why it matters
Existing frameworks leave sandbox isolation, session restoration, and token metering as build-it-yourself problems, which is why most agentic prototypes don't survive contact with production. DeepSeek Harness shipping all four as first-class core features — not plugins — raises the baseline for what 'production-ready' means and creates direct pressure on LangGraph and AutoGen to follow. The MIT license and explicit comparison matrix make the framework immediately evaluable against your existing stack without vendor negotiation.
Matthew Nelson's Active Media Publishing Group in Aurora, Oregon — featured in Editor & Publisher on September 9 — has grown from 13 employees and 35 titles (published six times per year) in 2017 to roughly 75 staff and 60+ publications today, with recent launches including The Canby Chronicle (July 2026), The Sherwood Sun (March 2026), and acquisitions of shuttered titles like The West Side Newspaper. The model: monthly frequency, direct-mail distribution, hyperlocal editorial (school events, community news), and local advertiser dependency — no venture backing, no digital paywall strategy.
Why it matters
Active Media's expansion — bootstrapped, monthly, direct-mail — is the existence proof that the community-anchor model can grow headcount while legacy publishers shrink. The economics that make it work are identical to those that make it fragile: monthly frequency reduces production cost and postal frequency, direct mail makes USPS Periodicals-class reliability a direct P&L variable, and local advertiser dependency concentrates revenue risk in a way that a subscriber base would not. For a niche publisher like Kav, the Active Media case is useful not as a template but as a data point about where the floor actually is: 35 titles at six issues per year in 2017 scaled to 60+ monthly titles in nine years without outside capital.
The Treasury buyback expansion we have been tracking executed Wednesday at $6 billion for 10–20-year bonds — triple the prior long-dated operation, aligning with Secretary Bessent's signaled volume — but failed to hold yields down. The 10-year Treasury yield climbed to 4.857% intraday (its highest since November 2023) even as the Treasury sold $39 billion in 10-year notes at 4.834% with the strongest demand ratio (2.71×) since 2019. The 30-year yield touched 5.285%. Fed funds futures traders were pricing roughly 60% probability of a rate hike at the September 15–16 FOMC. Oil breached $100/barrel (Brent) for the first time since July, adding to inflation concerns and narrowing the Fed's options.
Why it matters
We noted earlier this week that the structural tension between Fed portfolio shortening and Bessent's buybacks would test the program's ceiling. The yield-perverse outcome — yields rising despite a larger-than-expected intervention — confirms that when fiscal deficits, term premium, and supply pressure combine, a $6B operation is noise. HSBC's revised year-end 10-year forecast of 4.65% (up from 4.30%) alongside $8.4T in scheduled Treasury rollovers and $2.3T in corporate issuance before year-end means the refinancing math for any CRE or multifamily debt maturing in 2026–2027 should be stress-tested at current yields, not anticipated cuts. A September hike at 60% odds would push the effective rate higher still; the decisive next signal is Friday's CPI print.
Following the final IRS guidance on Section 530A IRAs we tracked Tuesday, Robinhood CEO Vladimir Tenev announced the company's implementation of the Trump Accounts program at the Goldman Sachs Communacopia + Technology Conference. Targeting 70 million eligible children under 18 with $1,000 in Treasury funding each at birth, the program was outlined alongside Robinhood's $1.3 billion annualized ARR, Robinhood Earn at 7% APY through stablecoin yield, and a 3% IRA contribution match on top of contributed dollars. The Gold Card crossed 1 million cardholders with 3% cash back routed into brokerage accounts; Robinhood Chain stock tokens report billions in daily trading volume.
Why it matters
The 7% Earn APY through stablecoin yield competes directly with USFR and money-market instruments at a yield premium that requires understanding the underlying risk structure — stablecoin yield is not a Treasury instrument, and the source of that 7% deserves scrutiny before treating it as a like-for-like alternative. The 3% IRA match is more structurally straightforward: it compounds on contributed dollars and provides a guaranteed return layer before any investment return. The Trump Accounts program — $1,000 per eligible child from Treasury, 70 million potential accounts — is the largest IRA-adjacent construction by a fintech platform in US history if it reaches scale, and the mechanics (tax-deferred growth from birth) warrant tracking as a new personal-finance baseline.
Agudath Israel of America and Torah Umesorah — whose Moetzes Gedolei HaTorah and Vaad Roshei Hayeshivos deliberated extensively before endorsing — have backed the Opportunity Fund for Jewish Education, a national Scholarship Granting Organization launched by OU, Teach Coalition, JFNA, and Prizmah to channel a new $1,700-per-taxpayer annual federal tax credit (dollar-for-dollar, not a deduction) into yeshiva and day school scholarships beginning January 1, 2027. The Joint Committee on Taxation estimates the program will generate $500 million in national SGO donations. Both organizations concluded that a single national SGO maximizes federal dollars rather than fragmenting them into competing local systems; 100% of designated donations reach participating schools. More than 900 Jewish day schools and yeshivos can begin registering now.
Why it matters
The Moetzes and Vaad Roshei Hayeshivos endorsement resolves what could have been a coordination problem — competing local SGOs would have fragmented the donor base and reduced total scholarship dollars reaching families. The $500M JCT estimate, spread across the Orthodox, Conservative, and other Jewish day school network, represents a structural change in tuition economics that has a direct stabilizing effect on enrollment and community demographics in high-density areas like Rockland County. Practically: eligible taxpayers in your network can donate up to $1,700 to the Fund starting January 1 and receive a federal credit dollar-for-dollar — this is not a deduction, it is a credit, which changes the economics substantially for middle-income donors.
Rockland County formally adopted Envision Rockland — its first comprehensive plan update since 2011 — following a 20-month process led by County Executive Ed Day and Legislature Chairman Jay Hood Jr. The plan covers the next 10–15 years and explicitly prioritizes expanding affordable and diverse housing options, directing growth toward town and village centers while protecting open space and environmentally sensitive areas. It does not directly regulate zoning but establishes the policy foundation municipalities will reference for local land-use decisions.
Why it matters
A comprehensive plan doesn't rezone anything by itself, but it sets the political baseline for every variance, rezoning petition, and municipal land-use decision over the next decade. The explicit emphasis on housing affordability and growth directed toward existing town and village centers — which in Rockland means Monsey, Spring Valley, Haverstraw, and Suffern — signals where county-level political cover exists for density increases. For a landlord with existing Rockland holdings, the plan creates a policy environment that is more favorable to multifamily additions near established centers than the 2011 baseline was; watch for municipal zoning updates that follow.
Trail of Bits security researchers discovered a critical bug in Lean 4 versions up to 4.33.1 (fixed in v4.34.0-rc1) in the String.Pos.Raw.extract function: extracting a one-byte slice at an astronomically large position returns an empty string under logical evaluation but returns the full original string in compiled native code, creating a contradiction that allows proving false statements — including a spurious proof of Fermat's Last Theorem. The Lean team patched the memory-safety issue in 90 minutes and resolved the semantic mismatch within five days.
Why it matters
This lands directly after we covered Anthropic's AI formalizing Fermat's Last Theorem in Lean in 11 days, and Kevin Buzzard's independent verification of the formalization. The bug establishes that the trusted boundary in theorem provers extends beyond the kernel to the compiler and external functions — a soundness gap that is independent of the quality of any particular proof. Any formal verification result that relies on native_decide or similar compiled mechanisms needs to be re-examined against v4.34.0-rc1 or later. The 90-minute patch turnaround is reassuring; the fact that the vulnerability existed and could produce a false Fermat proof through an innocuous-looking string function is the part that should update priors about formal verification as a gold standard.
City Council member Chris Banks introduced Bill T2026-1780 on September 10 to authorize electronic filing of annual building registrations for owners of three or more residential units — replacing the manual mail-in requirement in place since 1953. Late or missing registrations currently trigger fines of $500–$5,000 depending on property size and create downstream permit delays. Small Property Owners of New York and REBNY both backed the bill; Mayor Mamdani had endorsed the change, though his administration frames it as improving access to owner information for enforcement purposes.
Why it matters
The administrative efficiency case is real — lost mail and registration lapses are a recurring compliance trap for small-portfolio landlords, and electronic filing with confirmation receipts removes an unnecessary failure mode. The Mamdani administration's framing — digitized data for holding bad landlords accountable — signals that HPD intends to operationalize the centralized registration database for proactive enforcement rather than just modernizing back-office filing. The practical upside (no more fines from misdelivered mail) and the enforcement downside (easier cross-referencing of registration status against 311 complaints) are both real; the bill is likely to pass with broad support.
A Korean scholar, reported in Seoul Economic Daily on September 9, challenges the National Institute of Korean Language's recommendation to replace dakdoritang ('spicy braised chicken soup') with dakbokkeumtang, arguing that 'dori' derives from the Korean verb dorida ('to cut with a knife') — consistent with the productive morphological pattern underlying all Korean dish names (tang, guk, jjim, bap) — rather than from Japanese. The scholar cites pre-Joseon Dynasty records of dakdoritang and notes that the National Institute's own position acknowledges it will reconsider the replacement recommendation if pre-colonial Korean documents confirming 'dori' are found.
Why it matters
The methodological issue is worth the attention: the National Institute recommended replacement on phonological surface similarity to Japanese without producing documentary evidence of borrowing, and the scholar's counter-argument rests on morphological productivity (Korean verb classes and dish-naming rules), historical attestation, and absence of semantic transfer evidence — the same tools that distinguish native Hebrew roots from Aramaic or Romance borrowings in biblical and rabbinic philology. Institutional language bodies recommending etymological reanalysis based on ideological discomfort rather than primary attestation is a recurring problem; this case documents the full methodological debate in a compressed and unusually clear form.
Advertised Unit Prices Have Become a Poor Proxy for What Users Actually Get Three separate developments this edition illustrate the same structural problem: Claude Max's advertised session multipliers obscure weekly caps (now in litigation), Artificial Analysis revised its composite intelligence score three times in one week before landing GPT-6 Astra and Fable 5.1 at a tied 53 — while Fable consumes 3× more output tokens to reach that score — and the Treasury buyback tripled in size while yields moved the wrong direction. In each case, the headline metric and the operational reality have decoupled enough to mislead decision-makers.
Governance Infrastructure Is Arriving as a Formal Discipline, Not a Retrofit DeepSeek Harness ships sandbox isolation and session persistence in the core; the MDPI cybersecurity taxonomy formalizes six agent architecture patterns for comparative evaluation; Consort enforces test-driven development through deterministic orchestration rather than prompt discipline; and the enterprise-workforce model's 60-vs-40 outcome split keys explicitly to whether governance precedes or follows pilot launch. The convergence across frameworks, academic taxonomy, and practitioner tooling on the same structural questions — identity, authorization, state persistence, stopping conditions — signals that agent governance has moved from advice to artifact.
Federal Policy Is Writing Checks to Orthodox Communities on Multiple Fronts Simultaneously The Opportunity Fund for Jewish Education — endorsed by Agudath Israel's Moetzes Gedolei HaTorah and Torah Umesorah's Vaad Roshei Hayeshivos — channels up to $500M in estimated annual SGO donations into yeshiva scholarships via the One Big Beautiful Bill's $1,700 federal tax credit starting January 2027. Rockland County simultaneously adopted the Envision Rockland comprehensive plan explicitly addressing housing affordability and land-use density near established community centers. Two distinct federal and county policy tracks are now directing resources toward the same demographic — worth tracking whether they compound or create conflicting incentives.
The RSA Factorization Timeline Has Compressed Further Than Most Infrastructure Operators Realize RSA-260 fell in September 2026 at roughly $400,000 in GPU compute; a legacy Certificate Authority's 1990s-era RSA keys were factored separately; and a Lean 4 kernel bug — now patched — briefly allowed a spurious proof of Fermat's Last Theorem, exposing the gap between logical evaluation and native compilation in trusted proof systems. Together these three items make the same argument: cryptographic and formal-verification infrastructure that was 'probably fine' a year ago now has a concrete cost of attack, and institutional inertia around legacy keys or unaudited compiled code is measurable risk, not theoretical risk.
Print Survival Is Bifurcating Between Scale-Consolidation and Community-Anchor Models Transcontinental's Q3 results show Canada's largest printer pivoting aggressively into retail-media and in-store marketing via acquisition; Goubault Imprimeur's acquisition of ValPG explicitly cites industrial-landscape pressure requiring consolidation; while Active Media Publishing Group in Oregon runs 60+ local titles on a direct-mail, local-advertiser, community-anchor model that has grown headcount from 13 to 75 since 2017. The two survival strategies require opposite organizational structures: scale-consolidators need M&A capital and distribution infrastructure, while community-anchors need low churn, hyperlocal editorial discipline, and USPS rate predictability. Kav sits squarely in the second camp.
What to Expect
2026-09-14—Anthropic's Claude Code permanent 25% limit increase takes effect — replacing the temporary summer promotion and representing a net 17% reduction from the prior baseline for Claude Max subscribers.
2026-09-15—Public comment deadline at the Postal Regulatory Commission on USPS dockets MC2026-377/K2026-367 (Priority Mail Express, Priority Mail, and USPS Ground Advantage Contract 1511 additions to the Competitive Product List).
2026-09-15-16—Federal Open Market Committee meeting; markets currently pricing approximately 60% probability of a rate hike, with the 10-year Treasury already at its highest since November 2023.
2026-09-27-30—America's Newspapers Family Owners and Senior Leadership Conference in Washington, D.C., including a congressional Fly-In to advocate for the Stealth Bot Prohibition Act and postal rate relief.
2027-01-01—Opportunity Fund for Jewish Education launches, enabling eligible taxpayers to claim the new $1,700 federal tax credit for SGO donations; yeshivas can begin registering with the Fund now.
How We Built This Briefing
Every story, researched.
Every story verified across multiple sources before publication.
🔍
Scanned
Across multiple search engines and news databases
868
📖
Read in full
Every article opened, read, and evaluated
177
⭐
Published today
Ranked by importance and verified across sources
12
— The Primary Source
🎙 Listen as a podcast
Subscribe in your favorite podcast app to get each new briefing delivered automatically as audio.
Apple Podcasts
Library tab → ••• menu → Follow a Show by URL → paste