K Koda Intelligence
Subscribe

The Signal · Sunday 22 November 2026

Two Riemann proofs land hours apart, one Claude-assisted

Anthropic's Levent Alpöge published a Claude-assisted claimed proof as Lean code, about 250,000 pages1 if written out. Alessandro Masullo announced a 940-page purely human draft within hours.

Koda editorial team · 22 November 2026 · 6 min read

Listen · 17 min
Cover image for the 22 November 2026 Signal: Two Riemann proofs land hours apart, one Claude-assisted

In short

Today: Test your agents against real inboxes and real limits. The steps

The lead

Trend · Latent Space

this week, Latent Space's AINews reported two independent claimed proofs of the Riemann Hypothesis, released hours apart. Anthropic's Levent Alpöge published a Claude-assisted proof as Lean code, about 250,000 pages if written out1, while Alessandro Masullo announced a 940-page purely human draft.

Also today

Benchmark · AI Windvane · 22 NOV 2026

Claude Fable 5 Leads GPT-5.5 on GDP.pdf

Anthropic published a system card for Claude Fable 5 in June. It posts a 29.8% strict pass rate on GDP.pdf3, against 24.9% for GPT-5.5 and 22.5% for Claude Opus 4.8.

Why it matters. Route hard document work to Fable 5 only after it clears your own files: at 29.8%,3 even the leader fails most strict GDP.pdf tasks, just ahead of GPT-5.5.

Policy · The Verge · 20 NOV 2026

Signal Blocks Screenshots to Shut Out AI Agents

Signal now blocks screenshots on iOS 27 and Android to keep AI agents from reading private chats.

Why it matters. Agents that operate phones by capturing the screen lose sight of Signal threads, so any assistant workflow that summarizes or acts on those messages will break.

Agents · Google · 19 NOV 2026

Gemini API Loads Tool Schemas Only on Demand

Google previewed Tool Retrieval on November 19, loading full tool schemas only when an agent needs them. It supports Gemini 3.6 and later, including Gemini 4, plus remote MCP servers.

Why it matters. Adopt it for large tool libraries only after measuring your own token bills, because Google gives no savings figure and standard token charges still apply.

Stripe, via TLDR

Stripe Tracks AI Revenue

Stripe launched a weekly AI Index, and its first reading shows aggregate AI revenue up 103% year over year across about 20,000 businesses4. A recurring, payments-based measure gives a running read on how much money AI products are actually collecting.

Conflict · Al Jazeera · 21 NOV 2026

Iran Hits Tankers, Claims Strikes on US Radars

Iran's foreign minister still called for diplomacy, but by Saturday night the IRGC claimed missile and drone strikes on two US Patriot radars in Bahrain, and AP reported three tankers hit.

In brief

Markets

The Nasdaq fell 5.28% on the 20 Nov session and the S&P 500 lost 3.70%, while WTI crude rose 0.93% in the same session, a split that kept the selloff confined to equities and left oil untouched; the crypto Fear & Greed index reads 11, Extreme Fear.

S&P 500

6,910.39

-3.70%

Nasdaq

22,617.73

-5.28%

Bitcoin

$60,200.71

+2.38%

Ethereum

$1,681.53

+3.83%

Crude Oil (WTI)

$83.98

+0.93%

Crypto Fear & Greed

11 / 100

Extreme Fear

Levels from Yahoo Finance5, 20 Nov and 22 Nov; alternative.me6, 22 Nov.

Go deeper

The Koda Signal, 22 November 2026

Two Riemann proofs land hours apart, one Claude-assisted

0:0016:36

Every episode

Tool of the day

Toki

Productivity · Koda score7 6.6 / 10

Toki bets that your inbox and calendar already hold your to-do list, and that an agent which drafts follow-ups and waits for your OK beats one more list app. Early reviews back the capture side, from voice notes to photographed invites, while doubts gather around calendar visibility and opaque pricing. Forgetful schedulers and voice-first users should try the free tier now; paying should wait until Toki explains its credits.

Capability 6.5
Ease 7.5
Value 7.0
Momentum 5.5

Also worth a look

The bottom line

Claims are cheap now, so verification is the work.

Two claimed Riemann Hypothesis proofs arrived hours apart, one built with Claude as Lean code and one as a 940-page human draft, and neither counts until it is checked. The same holds for agents, where Toki's draft-then-approve design and reports of Meta's Muse reaching private Signal messages show why a human checkpoint matters. Build that checkpoint into anything you ship this week. Measure how often reviewers accept the agent's output as written.

  1. Pilot an approval-gated agent. Point Toki at one inbox and calendar for a week. Count how many drafted follow-ups you approve unchanged, since that rate tells you whether to trust it with more.
  2. Try on-demand tool schemas. Google previewed Tool Retrieval in the Gemini API on November 19, and it works with Gemini 3.6 and later plus remote MCP servers. Move one large tool set onto it and compare prompt size before and after.
  3. Audit what your agents can read. Meta's Muse agent reportedly reached private Signal messages through device permissions. List every permission your own agents hold and remove any they do not need.

Sources and checks

Every story links to its outlet. Below, by source, are the 14 figures we checked: 1 verified, 10 reported and 3 estimated.

  1. 1Latent Space. 1 figure on this page, 22 Nov 2026.Reported
  2. 2Palo Alto Networks, via TLDR. 1 figure on this page, 20 Nov 2026.Reported
  3. 3AI Windvane. 3 figures on this page, 22 Nov 2026.VerifiedReportedEstimated
  4. 4Stripe, via TLDR. 2 figures on this page, 20 Nov 2026.ReportedEstimated
  5. 5Yahoo Finance. 5 figures on this page, 20 Nov 2026 and 22 Nov 2026.Reported
  6. 6alternative.me. Crypto Fear & Greed close, Extreme Fear on the session, 22 Nov 2026.Reported
  7. 7Koda Lab report (our score). Toki, Koda score across four dimensions, 22 Nov 2026.Estimated

Verified confirmed by an independent searchReported as published by the named sourceEstimated approximate, in the source or in our check

Every figure, one by one (14)
  1. 250,000...Alpöge published a Claude-assisted proof as Lean code, about 250,000 pages if written out, while Alessandro Masullo...Latent SpaceReported
  2. $4.5 billionPalo Alto Networks agrees to acquire Applied Compute for $4.5 billion in stock to train and serve customized...Palo Alto Networks, via TLDRReported
  3. 29.8%It posts a 29.8% strict pass rate on GDP.pdf, against 24.9% for GPT-5.5 and 22.5% for Claude Opus 4.8.AI WindvaneVerified
  4. 24.9%It posts a 29.8% strict pass rate on GDP.pdf, against 24.9% for GPT-5.5 and 22.5% for Claude Opus 4.8.AI WindvaneReported
  5. 22.5%It posts a 29.8% strict pass rate on GDP.pdf, against 24.9% for GPT-5.5 and 22.5% for Claude Opus 4.8.AI WindvaneEstimated
  6. 103%Stripe launches weekly AI Index: aggregate AI revenue up 103% year over year across about 20,000 businessesStripe, via TLDRReported
  7. 20,000Stripe launches weekly AI Index: aggregate AI revenue up 103% year over year across about 20,000 businessesStripe, via TLDREstimated
  8. 6,910.39S&P 500 close, -3.70% on the sessionYahoo FinanceReported
  9. 22,617.73Nasdaq close, -5.28% on the sessionYahoo FinanceReported
  10. $60,200.71Bitcoin close, +2.38% on the sessionYahoo FinanceReported
  11. $1,681.53Ethereum close, +3.83% on the sessionYahoo FinanceReported
  12. $83.98Crude Oil (WTI) close, +0.93% on the sessionYahoo FinanceReported
  13. 11 / 100Crypto Fear & Greed close, Extreme Fear on the sessionalternative.meReported
  14. 6.6 / 10Toki, Koda score across four dimensionsKoda Lab report (our score)Estimated

Signal 240 · 22 November 2026 · 11 sources · How every number is checked · Corrections

Get the morning Signal

239 editions so far, one a day. Unsubscribe anytime.

Forward this to one operator you work with. Your referral link is in every email; milestones at koda.community/refer.