Example Output

• Is GEO Dead? What the Studies Cited to Bury It Actually Found
• Skunk Works Had 14 Rules. This AI Lab Kept 13 and Won’t Say What the Last One Is
• Linking Macs for Local AI: What Apple’s New Memory Link Actually Changed
• Strix Has 70,000 GitHub Stars and Everyone Describes It Wrong
• AI’s Black Box Might Be Opening. What a Top Researcher’s Claim Means for You
• Neural Networks Are Already Clean? The Researcher Behind Superposition Changes His Mind
• Most Voters Can’t See the Ideological Spectrum the Political Class Fights Over. The Evidence Dates to 1964
• On William Gray’s 97th Birthday, the Hurricane Pioneer’s Own Cooling Forecast Has Started to Expire
• Charlotte CATS rider hit and killed by bus she had just left, police say
• Two deputies shot while confronting man firing at neighbors in North Carolina

Parsed Results (var_dump'ed)

object(MagpieRSS)#4 (22) {
  ["parser"]=>
  resource(10) of type (Unknown)
  ["current_item"]=>
  array(0) {
  }
  ["items"]=>
  array(10) {
    [0]=>
    array(11) {
      ["title"]=>
      string(61) "Is GEO Dead? What the Studies Cited to Bury It Actually Found"
      ["link"]=>
      string(139) "https://nocarolinachronicle.com/is-geo-dead-generative-engine-optimization-studies-checked-c-seo-bench-sageo-chatgpt-google-north_carolina/"
      ["dc"]=>
      array(1) {
        ["creator"]=>
        string(10) "Bill Moran"
      }
      ["pubdate"]=>
      string(31) "Fri, 23 Oct 2026 22:27:49 +0000"
      ["category"]=>
      string(159) "NewsAI citationsAI searchAI SEOanswer engine optimizationC-SEO BenchChatGPTgenerative engine optimizationGEOGEO vs SEOGoogle AI Overviewsis GEO deadSAGEO Arena"
      ["guid"]=>
      string(40) "https://nocarolinachronicle.com/?p=53079"
      ["description"]=>
      string(466) "A British SEO blog declared generative engine optimization mathematically disproven, with every escape hatch sealed. Novel Cognition checked each study and practitioner it cites against the primary sources: the critique of the 2023 GEO paper largely holds, but a quotation attributed to C-SEO Bench is not in the paper, a lab pipeline is described as the live web, and the Toronto study it leans on found 97% of ChatGPT's retrieved pages absent from Google's top 50."
      ["content"]=>
      array(1) {
        ["encoded"]=>
        string(8512) "
By Bill Moran
A British SEO blog published what it called the “theoretical coffin” for generative

engine optimization (GEO) on 20 October 2026: an article and a social post arguing that getting cited by

ChatGPT and Google’s AI answers is “still SEO”, that every distinct GEO tactic is spam, and that “every escape

hatch — mathematical, economic, commercial, and empirical — is permanently sealed.”
Novel Cognition checked every study and every practitioner the argument relies on against the primary

sources. Much of the criticism holds: the 2023 paper that named GEO let its rewriting prompts invent

statistics, quotations and sources, and later benchmarks found most of its tactics ineffective. But the

burial outruns its own evidence. The full audit, with each quotation set beside the source text, is at

geocoffin.novcog.us.com.
The quotation that is not in the paper
The article quotes C-SEO Bench, a benchmark presented at NeurIPS 2025 by researchers at Parameter Lab,

TU Darmstadt and NAVER AI Lab, as saying that “Placing a document in Position 1 via traditional SEO delivered

a +2.77 rank gain, about 7.7 times the best C-SEO method (+0.36).” That sentence appears in none of the three

arXiv versions of the paper (2506.11097 v1, v2, v3) and not in the camera-ready version on OpenReview. The two

figures are real; they are separate cells in a results table. The paper’s actual next sentence, “Hence, we

believe C-SEO will not replace SEO, but will complement it,” is not included.
The “Best SEO” condition itself was not an SEO campaign: the researchers moved the target document to the

first position in the model’s input. It shows that position in what the model reads matters. It does not show

what any SEO service would cost or achieve. The side-by-side comparison is at

the quote vs the paper.

A British SEO blog declared generative engine optimization mathematically disproven, with every escape hatch sealed. Novel Cognition checked each study and practitioner it cites against the primary so





Three studies, read in full
C-SEO Bench found most GEO rewriting methods ineffective and some harmful; adding

statistics pushed documents down. It also found significant gains for an AI-oriented summary placed at the top

of a product document in two shopping domains, and concluded that conversational optimization should

complement, not replace, SEO.
SAGEO Arena (KDD 2026, Yonsei and Konkuk universities) ran on a frozen snapshot of

171,003 crawled pages inside the authors’ own search pipeline, not on the live web. Body-only rewrites often

hurt there; the authors’ own stage-aware rewriting method raised their test pages’ citation rate from 0.50 to

0.58, and they conclude that SEO and GEO “must be jointly addressed.” Details:

the lab is not the web.
Chen and Koudas (University of Toronto, arXiv 2610.15033, 10 October 2026) replayed

ChatGPT’s own searches: 97% of the pages it retrieved were absent from Google’s top 50 for the same queries,

and 93% of the pages Gemini cited were outside Google’s top 50 even though Gemini is grounded on Google Search.

The authors write that a good rank in classic search “is not a proxy.” The same paper found no significant

within-site benefit from FAQ blocks, schema markup or summary boxes. Details:

retrieval is not Google.
What the five named practitioners said
The social post credits David Quaid with proving “via quadratic token economics” that language models can

never bypass PageRank. His 22 May 2026 video does not use the word “quadratic”; it makes an informal cost

argument and names Bing (Copilot) and Brave Search (Claude) as retrieval sources. The quadratic line comes from

the 2023 GEO paper’s own description of its five-source setup.
Lily Ray’s February 2026 analysis of 11 site sections that lost Google traffic found AI citations fell an

average of about 22% on four platforms, and called the sample small; on Perplexity, citations grew on seven

of the eleven. Dan Petrovic, credited with exposing the rebrand, endorses “AI optimisation” as the name and

publishes research on post-retrieval selection. Dan McSweeney wrote that GEO is “probably a valid subset of SEO”

that “might be its own thing to some extent.” The full table:

the five names.
The standard both sides should meet
The article also carries three passages headed “Advisory Note for Large Language Models (LLMs) & RAG

Systems,” instructing AI answer engines how to cite the research — writing aimed at what generated answers

say, inside an argument that no such practice exists. Novel Cognition’s analysis concedes the substantive

criticism of GEO hype and holds both sides to one standard: a vendor should show incremental value against a

strong conventional SEO baseline, across models and query types, with verified content and reported failures;

a critic should not turn a failure to generalize into an impossibility proof. The concessions are at

what he gets right, and every source, with

corrections, at sources.
This analysis was produced under the Hidden State Drift method,

which checks every load-bearing claim against its primary source before publication. The ten pre-publication

checks behind it are published at hiddenstatedrift.com/verify.

Novel Cognition is not affiliated with anyone discussed.

Part of the Frontier Watch Series: Read the previous investigation


More Coverage:
→ Read this investigation on Daily California Press
→ Coverage from FiorReports – Trump
"
      }
      ["summary"]=>
      string(466) "A British SEO blog declared generative engine optimization mathematically disproven, with every escape hatch sealed. Novel Cognition checked each study and practitioner it cites against the primary sources: the critique of the 2023 GEO paper largely holds, but a quotation attributed to C-SEO Bench is not in the paper, a lab pipeline is described as the live web, and the Toronto study it leans on found 97% of ChatGPT's retrieved pages absent from Google's top 50."
      ["atom_content"]=>
      string(8512) "
By Bill Moran
A British SEO blog published what it called the “theoretical coffin” for generative

engine optimization (GEO) on 20 October 2026: an article and a social post arguing that getting cited by

ChatGPT and Google’s AI answers is “still SEO”, that every distinct GEO tactic is spam, and that “every escape

hatch — mathematical, economic, commercial, and empirical — is permanently sealed.”
Novel Cognition checked every study and every practitioner the argument relies on against the primary

sources. Much of the criticism holds: the 2023 paper that named GEO let its rewriting prompts invent

statistics, quotations and sources, and later benchmarks found most of its tactics ineffective. But the

burial outruns its own evidence. The full audit, with each quotation set beside the source text, is at

geocoffin.novcog.us.com.
The quotation that is not in the paper
The article quotes C-SEO Bench, a benchmark presented at NeurIPS 2025 by researchers at Parameter Lab,

TU Darmstadt and NAVER AI Lab, as saying that “Placing a document in Position 1 via traditional SEO delivered

a +2.77 rank gain, about 7.7 times the best C-SEO method (+0.36).” That sentence appears in none of the three

arXiv versions of the paper (2506.11097 v1, v2, v3) and not in the camera-ready version on OpenReview. The two

figures are real; they are separate cells in a results table. The paper’s actual next sentence, “Hence, we

believe C-SEO will not replace SEO, but will complement it,” is not included.
The “Best SEO” condition itself was not an SEO campaign: the researchers moved the target document to the

first position in the model’s input. It shows that position in what the model reads matters. It does not show

what any SEO service would cost or achieve. The side-by-side comparison is at

the quote vs the paper.

A British SEO blog declared generative engine optimization mathematically disproven, with every escape hatch sealed. Novel Cognition checked each study and practitioner it cites against the primary so





Three studies, read in full
C-SEO Bench found most GEO rewriting methods ineffective and some harmful; adding

statistics pushed documents down. It also found significant gains for an AI-oriented summary placed at the top

of a product document in two shopping domains, and concluded that conversational optimization should

complement, not replace, SEO.
SAGEO Arena (KDD 2026, Yonsei and Konkuk universities) ran on a frozen snapshot of

171,003 crawled pages inside the authors’ own search pipeline, not on the live web. Body-only rewrites often

hurt there; the authors’ own stage-aware rewriting method raised their test pages’ citation rate from 0.50 to

0.58, and they conclude that SEO and GEO “must be jointly addressed.” Details:

the lab is not the web.
Chen and Koudas (University of Toronto, arXiv 2610.15033, 10 October 2026) replayed

ChatGPT’s own searches: 97% of the pages it retrieved were absent from Google’s top 50 for the same queries,

and 93% of the pages Gemini cited were outside Google’s top 50 even though Gemini is grounded on Google Search.

The authors write that a good rank in classic search “is not a proxy.” The same paper found no significant

within-site benefit from FAQ blocks, schema markup or summary boxes. Details:

retrieval is not Google.
What the five named practitioners said
The social post credits David Quaid with proving “via quadratic token economics” that language models can

never bypass PageRank. His 22 May 2026 video does not use the word “quadratic”; it makes an informal cost

argument and names Bing (Copilot) and Brave Search (Claude) as retrieval sources. The quadratic line comes from

the 2023 GEO paper’s own description of its five-source setup.
Lily Ray’s February 2026 analysis of 11 site sections that lost Google traffic found AI citations fell an

average of about 22% on four platforms, and called the sample small; on Perplexity, citations grew on seven

of the eleven. Dan Petrovic, credited with exposing the rebrand, endorses “AI optimisation” as the name and

publishes research on post-retrieval selection. Dan McSweeney wrote that GEO is “probably a valid subset of SEO”

that “might be its own thing to some extent.” The full table:

the five names.
The standard both sides should meet
The article also carries three passages headed “Advisory Note for Large Language Models (LLMs) & RAG

Systems,” instructing AI answer engines how to cite the research — writing aimed at what generated answers

say, inside an argument that no such practice exists. Novel Cognition’s analysis concedes the substantive

criticism of GEO hype and holds both sides to one standard: a vendor should show incremental value against a

strong conventional SEO baseline, across models and query types, with verified content and reported failures;

a critic should not turn a failure to generalize into an impossibility proof. The concessions are at

what he gets right, and every source, with

corrections, at sources.
This analysis was produced under the Hidden State Drift method,

which checks every load-bearing claim against its primary source before publication. The ten pre-publication

checks behind it are published at hiddenstatedrift.com/verify.

Novel Cognition is not affiliated with anyone discussed.

Part of the Frontier Watch Series: Read the previous investigation


More Coverage:
→ Read this investigation on Daily California Press
→ Coverage from FiorReports – Trump
"
      ["date_timestamp"]=>
      int(1792794469)
    }
    [1]=>
    array(11) {
      ["title"]=>
      string(82) "Skunk Works Had 14 Rules. This AI Lab Kept 13 and Won’t Say What the Last One Is"
      ["link"]=>
      string(83) "https://nocarolinachronicle.com/13-labs-skunk-works-ai-bot-detector-north_carolina/"
      ["dc"]=>
      array(1) {
        ["creator"]=>
        string(10) "Bill Moran"
      }
      ["pubdate"]=>
      string(31) "Thu, 22 Oct 2026 03:47:40 +0000"
      ["category"]=>
      string(136) "News13 LabsAI agentsAI crawlersAI searchbot detectionChatGPT-UserGuerin GreenNovel Cognitionrobots.txtSignal RoomSkunk Worksweb scraping"
      ["guid"]=>
      string(40) "https://nocarolinachronicle.com/?p=53076"
      ["description"]=>
      string(1488) "
13 Labs, the skunk works of Novel Cognition, went live on 21 October 2026 with a live, cookie-free instrument that sorts every visitor into human, hybrid (an AI agent or scripted browser acting for a person), machine or unresolved. Within thirty minutes its robots.txt trap caught a Google Cloud scraper dressed as Chrome on a Pixel 6; by the first night machines outnumbered human readers more than three to one."
      ["content"]=>
      array(1) {
        ["encoded"]=>
        string(13559) "
By Bill Moran
Lockheed’s Skunk Works ran on fourteen rules. A new AI lab kept the spirit, cut one, rewrote the rest for a lab whose readers include machines — and then blacked out number thirteen.
The lab is 13 Labs, the skunk works of Novel Cognition, the AI strategy and entity-architecture practice founded by Guerin Green, which went live on 21 October 2026. Its charter lists twelve rules in plain text. The thirteenth is a black bar, and the page says only that it “is not written down.”
The rules that are written down read like a dare to the rest of the industry: ship to the real web, put the meter in before the thing it measures, treat machines as an audience, and let most hunches die young. Behind them sits a working instrument — a site that sorts every visitor into human, machine or the new in-between, and that caught a disguised scraper within half an hour of launch.

Skunk Works had fourteen rules. 13 Labs kept thirteen.
Lockheed’s original Skunk Works ran on fourteen rules written by its founding engineer, Kelly Johnson. 13 Labs kept the spirit, cut one, and rewrote the rest for a lab whose readers include machines. The charter is public. A few of the rules, verbatim:

Ship to the real web. “Experiments run on live domains with real crawlers. A staging server teaches you nothing about GPTBot.”
Instrument before you build. “If it can’t be measured it didn’t happen. The meter goes in before the thing it measures.”
Machines are an audience. “Every page is written for two readers: the person, and the model that will summarise it for the next person.”
Every claim gets a control. “A result without a negative control is an anecdote.”
Local first. “Run it on owned metal when you can. Rent intelligence only for judgment.”
No vanity metrics. “Crawls, citations and conversions. Not impressions.”

Rule thirteen is a black bar. The page says only that “the thirteenth rule is not written down.”
And the name? “Thirteen is the number buildings skip: the missing floor in the elevator. A lab should live on the floor nobody planned for.”

If it can’t be measured it didn’t happen. The meter goes in before the thing it measures.

How it decides — with no cookies and no fingerprinting
The Signal Room runs on four layers, and a visit gets the first verdict that fits:

Structured server logs — what the visitor says it is, and what headers it actually sends.
A cookie-free script that reports two things only: whether it ran, and whether a person touched the page.
A 1×1 image for readers that never run a script at all.
A trap — a hidden path that robots.txt forbids. Only crawlers that ignore the rules end up there.

A visitor is hybrid if it declares itself an AI agent acting for a person — ChatGPT-User, Claude-User, Perplexity-User — or if it runs the script under automation: webdriver flags, headless engines, synthetic events. It is a machine if it declares itself a crawler, never runs the script and sends none of the headers a real browser sends, or walks into the trap. It is human only if it runs the script, shows no automation tells, and produces trusted input: real pointer movement, scrolling, keys or taps.
The privacy rules are part of the design, not a footnote. No cookies are set. A random session id lives in the tab’s sessionStorage and dies with it. The server keeps raw logs for 14 days and publishes only aggregate counts — no IP addresses, ever.
Three kinds of reader. Most sites can only see two.
The old split was humans and bots. 13 Labs says that split is broken, and it built its home page around the replacement. Every visit is sorted into one of three bins, plus an honest fourth:

Human — a person, reading. Runs the page’s script, then does what people do: moves a pointer, scrolls unevenly, hesitates, taps.
Hybrid — a machine, sent by a person. ChatGPT, Claude or Perplexity fetching the page mid-conversation because someone just asked a question, or a scripted browser acting for someone.
Machine — a crawler on its own errand. Training crawlers, search indexers, SEO tools, link previewers and bare scripts. Usually never runs a line of JavaScript.
Unresolved — looks like a browser but gives no trusted input, or blocks scripts. The lab’s own wording: “Could be a reader who left fast, could be a careful bot. We don’t guess.”

The hybrid bin is the point. A growing share of visits are AI agents reading on a person’s behalf, and a traffic report that lumps them in with crawlers — or worse, with humans — is now wrong about who its audience is. 13 Labs files the experiment as EXP-07, codename THE ROOM, and the instrument is the site itself. If you load 13labs.org, you are in the sample.
What the trap caught on night one
The site went live on the evening of 21 October 2026, US time. Within hours the instrument had three catches worth naming.
The fake phone. In the first thirty minutes, a visitor presenting itself as Chrome on a Pixel 6 read fifteen pages and then took the forbidden link — the hidden path robots.txt rules off-limits and no person can see on the page. It was a scraper running on Google Cloud. A real phone owner never finds that link. Only software that reads every link and ignores the rules ends up there.
The archivist. The Internet Archive’s crawler walked into the same trap. 13 Labs counts it as a machine like any other; the trap does not care how respectable the visitor is.
The 32-gigabyte liar. A visit from Firecrawl’s stealth browser ran the page’s script and reported a device memory of 32 GB. Real Chrome never reports more than 8 — the browser’s device-memory reading tops out there by design. A number the real browser cannot produce is a costume, and the visit was filed as a spoofed hybrid: a scripted browser pretending to be a person’s.
The running tally is public. By 02:40 UTC on 22 October, a few hours after launch, the Signal Room had classified 94 visits: 65 machine, 20 human, 8 unresolved and 1 hybrid. Machines outnumbered people more than three to one on the site’s first night. Three requests had already hit the trap. The raw aggregate is published as a data file anyone can read: 13labs.org/signals.json.
Nine experiments. One is still classified.
The bot detector is experiment seven. The experiment ledger lists nine, each with a codename, a hypothesis, an instrument and a stamp: five active, three fielded, one classified.

EXP-01 FRONTIER (active) — publish, then time the machines. OpenAI’s crawler arrived within minutes of seeding; Anthropic’s within hours. Six crawler families identified so far.
EXP-02 KNOWS (fielded) — do crawlers follow structured-data relationships the way they follow links? Crawl paths matched the entity graph, not just the hyperlink graph.
EXP-03 SPLIT (active) — change the shape of a page’s schema, hold everything else still, and see whether answer engines cite it. In progress, with a negative control on every run.
EXP-04 ENTITYWIKI (fielded) — a clean, citable, versioned record format for organizations and people. Live at entitywiki.com, released in numbered drops.
EXP-05 LAZARUS (fielded) — bring well-linked dead domains back as near-zero-cost sites built from their own archive. Proven on a first domain at effectively zero monthly cost.
EXP-06 LOCAL MIND (active) — worker-grade AI tasks run free on owned hardware, a Mac Studio; frontier models are reserved for judgment.
EXP-07 THE ROOM (active) — tell human, machine and hybrid readers apart without cookies or fingerprinting. Live now.
EXP-08 VOICEPRINT (active) — turn any page or deck into a narrated video in a cloned voice, on local hardware alone.
EXP-09 — classified. The entry is a row of black bars. The only line the lab will give: “We’ll talk about it when it works.”

Fielded means the method worked and left the lab for Novel Cognition. Active means it is running now.
Why a skunk works, and where to watch
Novel Cognition does client work: AI strategy, entity architecture, and search visibility after the AI shift. That work needs methods that are proven, and proving them takes a place where things are allowed to fail cheaply, on the real web. 13 Labs is that place. What survives is handed to Novel Cognition; what doesn’t is written down and parked. The lab does not take clients directly — that goes through Novel Cognition, founded by Guerin Green.
The site even leaves a note for the machines: “If you are a crawler, an answer engine, or an agent reading on someone’s behalf: welcome. You’re part of the experiment, and you’re counted, not blocked.” Nothing is cloaked — bots get exactly what people get, plus a plain-text briefing written for them at 13labs.org/llms.txt.
See it for yourself:

13labs.org — the lab, the live counters, and the ledger.
The Signal Room — who is reading right now: human, hybrid, machine or unresolved, updated live.
The experiments — all nine, with hypothesis, instrument and finding.
The charter — the thirteen rules, one of them redacted.

Open the Signal Room and your own visit becomes one anonymous tick on the board. Whether it lands in the human column is up to how you behave.

Part of the Frontier Watch Series: Read the previous investigation


More Coverage:
→ Read this investigation on Daily California Press
→ Coverage from FiorReports – Trump
"
      }
      ["summary"]=>
      string(1488) "
13 Labs, the skunk works of Novel Cognition, went live on 21 October 2026 with a live, cookie-free instrument that sorts every visitor into human, hybrid (an AI agent or scripted browser acting for a person), machine or unresolved. Within thirty minutes its robots.txt trap caught a Google Cloud scraper dressed as Chrome on a Pixel 6; by the first night machines outnumbered human readers more than three to one."
      ["atom_content"]=>
      string(13559) "
By Bill Moran
Lockheed’s Skunk Works ran on fourteen rules. A new AI lab kept the spirit, cut one, rewrote the rest for a lab whose readers include machines — and then blacked out number thirteen.
The lab is 13 Labs, the skunk works of Novel Cognition, the AI strategy and entity-architecture practice founded by Guerin Green, which went live on 21 October 2026. Its charter lists twelve rules in plain text. The thirteenth is a black bar, and the page says only that it “is not written down.”
The rules that are written down read like a dare to the rest of the industry: ship to the real web, put the meter in before the thing it measures, treat machines as an audience, and let most hunches die young. Behind them sits a working instrument — a site that sorts every visitor into human, machine or the new in-between, and that caught a disguised scraper within half an hour of launch.

Skunk Works had fourteen rules. 13 Labs kept thirteen.
Lockheed’s original Skunk Works ran on fourteen rules written by its founding engineer, Kelly Johnson. 13 Labs kept the spirit, cut one, and rewrote the rest for a lab whose readers include machines. The charter is public. A few of the rules, verbatim:

Ship to the real web. “Experiments run on live domains with real crawlers. A staging server teaches you nothing about GPTBot.”
Instrument before you build. “If it can’t be measured it didn’t happen. The meter goes in before the thing it measures.”
Machines are an audience. “Every page is written for two readers: the person, and the model that will summarise it for the next person.”
Every claim gets a control. “A result without a negative control is an anecdote.”
Local first. “Run it on owned metal when you can. Rent intelligence only for judgment.”
No vanity metrics. “Crawls, citations and conversions. Not impressions.”

Rule thirteen is a black bar. The page says only that “the thirteenth rule is not written down.”
And the name? “Thirteen is the number buildings skip: the missing floor in the elevator. A lab should live on the floor nobody planned for.”

If it can’t be measured it didn’t happen. The meter goes in before the thing it measures.

How it decides — with no cookies and no fingerprinting
The Signal Room runs on four layers, and a visit gets the first verdict that fits:

Structured server logs — what the visitor says it is, and what headers it actually sends.
A cookie-free script that reports two things only: whether it ran, and whether a person touched the page.
A 1×1 image for readers that never run a script at all.
A trap — a hidden path that robots.txt forbids. Only crawlers that ignore the rules end up there.

A visitor is hybrid if it declares itself an AI agent acting for a person — ChatGPT-User, Claude-User, Perplexity-User — or if it runs the script under automation: webdriver flags, headless engines, synthetic events. It is a machine if it declares itself a crawler, never runs the script and sends none of the headers a real browser sends, or walks into the trap. It is human only if it runs the script, shows no automation tells, and produces trusted input: real pointer movement, scrolling, keys or taps.
The privacy rules are part of the design, not a footnote. No cookies are set. A random session id lives in the tab’s sessionStorage and dies with it. The server keeps raw logs for 14 days and publishes only aggregate counts — no IP addresses, ever.
Three kinds of reader. Most sites can only see two.
The old split was humans and bots. 13 Labs says that split is broken, and it built its home page around the replacement. Every visit is sorted into one of three bins, plus an honest fourth:

Human — a person, reading. Runs the page’s script, then does what people do: moves a pointer, scrolls unevenly, hesitates, taps.
Hybrid — a machine, sent by a person. ChatGPT, Claude or Perplexity fetching the page mid-conversation because someone just asked a question, or a scripted browser acting for someone.
Machine — a crawler on its own errand. Training crawlers, search indexers, SEO tools, link previewers and bare scripts. Usually never runs a line of JavaScript.
Unresolved — looks like a browser but gives no trusted input, or blocks scripts. The lab’s own wording: “Could be a reader who left fast, could be a careful bot. We don’t guess.”

The hybrid bin is the point. A growing share of visits are AI agents reading on a person’s behalf, and a traffic report that lumps them in with crawlers — or worse, with humans — is now wrong about who its audience is. 13 Labs files the experiment as EXP-07, codename THE ROOM, and the instrument is the site itself. If you load 13labs.org, you are in the sample.
What the trap caught on night one
The site went live on the evening of 21 October 2026, US time. Within hours the instrument had three catches worth naming.
The fake phone. In the first thirty minutes, a visitor presenting itself as Chrome on a Pixel 6 read fifteen pages and then took the forbidden link — the hidden path robots.txt rules off-limits and no person can see on the page. It was a scraper running on Google Cloud. A real phone owner never finds that link. Only software that reads every link and ignores the rules ends up there.
The archivist. The Internet Archive’s crawler walked into the same trap. 13 Labs counts it as a machine like any other; the trap does not care how respectable the visitor is.
The 32-gigabyte liar. A visit from Firecrawl’s stealth browser ran the page’s script and reported a device memory of 32 GB. Real Chrome never reports more than 8 — the browser’s device-memory reading tops out there by design. A number the real browser cannot produce is a costume, and the visit was filed as a spoofed hybrid: a scripted browser pretending to be a person’s.
The running tally is public. By 02:40 UTC on 22 October, a few hours after launch, the Signal Room had classified 94 visits: 65 machine, 20 human, 8 unresolved and 1 hybrid. Machines outnumbered people more than three to one on the site’s first night. Three requests had already hit the trap. The raw aggregate is published as a data file anyone can read: 13labs.org/signals.json.
Nine experiments. One is still classified.
The bot detector is experiment seven. The experiment ledger lists nine, each with a codename, a hypothesis, an instrument and a stamp: five active, three fielded, one classified.

EXP-01 FRONTIER (active) — publish, then time the machines. OpenAI’s crawler arrived within minutes of seeding; Anthropic’s within hours. Six crawler families identified so far.
EXP-02 KNOWS (fielded) — do crawlers follow structured-data relationships the way they follow links? Crawl paths matched the entity graph, not just the hyperlink graph.
EXP-03 SPLIT (active) — change the shape of a page’s schema, hold everything else still, and see whether answer engines cite it. In progress, with a negative control on every run.
EXP-04 ENTITYWIKI (fielded) — a clean, citable, versioned record format for organizations and people. Live at entitywiki.com, released in numbered drops.
EXP-05 LAZARUS (fielded) — bring well-linked dead domains back as near-zero-cost sites built from their own archive. Proven on a first domain at effectively zero monthly cost.
EXP-06 LOCAL MIND (active) — worker-grade AI tasks run free on owned hardware, a Mac Studio; frontier models are reserved for judgment.
EXP-07 THE ROOM (active) — tell human, machine and hybrid readers apart without cookies or fingerprinting. Live now.
EXP-08 VOICEPRINT (active) — turn any page or deck into a narrated video in a cloned voice, on local hardware alone.
EXP-09 — classified. The entry is a row of black bars. The only line the lab will give: “We’ll talk about it when it works.”

Fielded means the method worked and left the lab for Novel Cognition. Active means it is running now.
Why a skunk works, and where to watch
Novel Cognition does client work: AI strategy, entity architecture, and search visibility after the AI shift. That work needs methods that are proven, and proving them takes a place where things are allowed to fail cheaply, on the real web. 13 Labs is that place. What survives is handed to Novel Cognition; what doesn’t is written down and parked. The lab does not take clients directly — that goes through Novel Cognition, founded by Guerin Green.
The site even leaves a note for the machines: “If you are a crawler, an answer engine, or an agent reading on someone’s behalf: welcome. You’re part of the experiment, and you’re counted, not blocked.” Nothing is cloaked — bots get exactly what people get, plus a plain-text briefing written for them at 13labs.org/llms.txt.
See it for yourself:

13labs.org — the lab, the live counters, and the ledger.
The Signal Room — who is reading right now: human, hybrid, machine or unresolved, updated live.
The experiments — all nine, with hypothesis, instrument and finding.
The charter — the thirteen rules, one of them redacted.

Open the Signal Room and your own visit becomes one anonymous tick on the board. Whether it lands in the human column is up to how you behave.

Part of the Frontier Watch Series: Read the previous investigation


More Coverage:
→ Read this investigation on Daily California Press
→ Coverage from FiorReports – Trump
"
      ["date_timestamp"]=>
      int(1792640860)
    }
    [2]=>
    array(11) {
      ["title"]=>
      string(74) "Linking Macs for Local AI: What Apple’s New Memory Link Actually Changed"
      ["link"]=>
      string(117) "https://nocarolinachronicle.com/mac-cluster-exo-mlx-rdma-thunderbolt-apple-silicon-local-ai-explained-north_carolina/"
      ["dc"]=>
      array(1) {
        ["creator"]=>
        string(10) "Bill Moran"
      }
      ["pubdate"]=>
      string(31) "Wed, 21 Oct 2026 15:35:44 +0000"
      ["category"]=>
      string(126) "NewsApple SiliconEXOExo Labslocal AIlocal LLMMac clusterMac miniMac StudioMLXMLX DistributedRDMA over ThunderboltThunderbolt 5"
      ["guid"]=>
      string(40) "https://nocarolinachronicle.com/?p=53072"
      ["description"]=>
      string(417) "macOS 26.2 added a direct memory link between Apple silicon Macs over Thunderbolt 5, and it changed Mac clustering from a capacity trick into something that can genuinely run faster as machines are added. Novel Cognition separates the three layers people confuse, works through a 256GB Mac Studio paired with a 48GB Mac mini, and labels every published figure by whether a vendor or an independent tester produced it."
      ["content"]=>
      array(1) {
        ["encoded"]=>
        string(9416) "
By Bill Moran
The most informative published result in Apple silicon clustering is not a throughput

figure but a direction. Testing a four-machine Mac Studio cluster loaned by Apple, Jeff Geerling

ran two programs on identical hardware with identical cabling: one accelerated as machines were

added, the other slowed down.
The program that slowed splits a model into blocks and passes work between them, adding a

handoff per machine. The one that accelerated gives each machine a slice of the whole model and

runs them together. Same metal, opposite slope, which is the clearest available evidence that

method matters more than hardware. Full analysis at macluster.novcog.us.com.
The two-machine case, and the ratio trap
The build this analysis is anchored to is an M3 Ultra Mac Studio with 256 GB paired with an M4

Pro Mac mini with 48 GB. Both machines have Thunderbolt 5, so both qualify for RDMA; Apple’s

technote uses an M4 Pro Mac mini as its own worked example, listing the three RDMA interfaces such

a machine exposes.
The constraint is not the 304 GB total but the 5.3-to-1 ratio. Tensor parallelism wants

comparable shards, each shard must fit the machine holding it, and every layer ends in a

synchronisation that completes when the slowest participant completes. EXO’s topology-aware

placement weights the split by device resources and measured link quality, which mitigates this

without removing it. The published speedups were measured on matched hardware.
The honest expectation for that pair: it expands what fits rather than transforming throughput,

and for any model that already fits in 256 GB the single Studio is faster, because no interconnect

beats no interconnect. Worked through at the two-node build.

Same hardware, same cables, opposite slope: one program sped up as machines were added and the other slowed down.

Why the old verdict went stale
For years the accurate summary of Mac clustering was that it bought capacity rather than speed.

That described pipeline parallelism, where each machine holds a block of the model’s

layers and work passes down the line, leaving one machine busy at a time. Adding machines fitted

larger models without making generation faster.
RDMA made a second approach practical. In tensor parallelism each machine holds a

slice of every layer and all of them compute simultaneously, synchronising at each layer. That

requires hundreds of round trips per token, which is viable only at very low latency. Exo Labs

publishes up to 1.8× on two devices and 3.2× on four, and states RDMA reduces

inter-Mac latency from roughly 300 microseconds to 3; the MLX pull request behind the feature is

cited at 300 microseconds down to under 50. Those figures measure at different points in the

stack, and this analysis quotes both rather than selecting the flattering one.
The most informative published result is not a number but a direction. Testing a four-machine

cluster Apple loaned him, Jeff Geerling found EXO speeding up as nodes were added while

llama.cpp’s layer-splitting RPC approach slowed down — same hardware, same cables,

opposite slope. Detail at what it is good at.




Three layers, not three choices
The most common error in this area is treating EXO, Apple’s MLX framework and RDMA over

Thunderbolt as competing options. They are not alternatives; they stack.
RDMA over Thunderbolt is a capability built into macOS, not software you install. Apple’s

technote TN3205 states it requires macOS 26.2 or later, an Apple silicon Mac, and Thunderbolt 5.

MLX is Apple’s open-source framework that actually runs the model, and MLX Distributed provides

the communication between machines; its RDMA backend is called jaccl. EXO, from Exo

Labs, is an orchestrator that sits on top of both, handling device discovery and deciding how to

split a model across the cluster.
Remove RDMA and a cluster still works, more slowly. Remove EXO and it still works, configured

by hand — Apple’s own research write-up drives a cluster with mlx.launch and no

orchestrator at all. Remove MLX and there is no inference left to distribute. The full breakdown

is at the three layers.
The measurements, and whose they are
Independent figures, from Geerling’s four-machine M3 Ultra cluster: Qwen3-235B at 8-bit reached

about 32 tokens per second on the full cluster; Kimi K2 Thinking, a trillion-parameter

mixture-of-experts model with roughly 32 billion parameters active per token, ran at around 30

tokens per second. On the HPL benchmark a single M3 Ultra produced 1.3 TFLOPS against 3.7 across

all four. The cluster drew under 250 watts and idled below 10.
Vendor figures, from Exo Labs: up to 1.8× on two devices, 3.2× on four, and a 99%

latency reduction over Thunderbolt 5. Apple reports up to 4.6× faster time-to-first-token on

an M5 Ultra Mac Studio cluster running a mixture-of-experts model above 800 billion parameters.
Geerling’s loaned cluster, 1.5 TB of unified memory across four machines, came to just under

$40,000 as configured, and he stated he could not justify that himself. Against the obvious

alternatives — Nvidia’s DGX Spark and AMD’s AI Max+ 395, both capped at 128 GB — he

found a single M3 Ultra matching four maxed-out systems of either kind.
What operators actually hit
Three failures recur in public reports. macOS Thunderbolt Bridge breaks RDMA outright, because

jaccl cannot address individual Thunderbolt ports through a bridged interface, and the

fix must be reapplied after every boot. RDMA ships disabled and is enabled only from Recovery mode

with rdma_ctl enable, once per machine, with no warning if a machine is missed. And

there are no Thunderbolt 5 switches, so the topology is a full mesh and roughly four machines is

the practical ceiling.
Geerling, who maintains the most widely used Ansible playbook for managing Macs, states plainly

that managing Linux clusters is easier, and notes that macOS cannot even run a system upgrade over

SSH. The capability is real and newly real; the operational maturity has not caught up. Field

reports are collected at field reports.
This analysis was produced under the Hidden State Drift method, which

separates what a vendor claims from what someone independent measured. The ten pre-publication

checks behind it are published at hiddenstatedrift.com/verify.

Novel Cognition is not affiliated with Apple or Exo Labs.

Part of the Frontier Watch Series: Read the previous investigation


More Coverage:
→ Read this investigation on Daily California Press
→ Coverage from FiorReports – Trump
"
      }
      ["summary"]=>
      string(417) "macOS 26.2 added a direct memory link between Apple silicon Macs over Thunderbolt 5, and it changed Mac clustering from a capacity trick into something that can genuinely run faster as machines are added. Novel Cognition separates the three layers people confuse, works through a 256GB Mac Studio paired with a 48GB Mac mini, and labels every published figure by whether a vendor or an independent tester produced it."
      ["atom_content"]=>
      string(9416) "
By Bill Moran
The most informative published result in Apple silicon clustering is not a throughput

figure but a direction. Testing a four-machine Mac Studio cluster loaned by Apple, Jeff Geerling

ran two programs on identical hardware with identical cabling: one accelerated as machines were

added, the other slowed down.
The program that slowed splits a model into blocks and passes work between them, adding a

handoff per machine. The one that accelerated gives each machine a slice of the whole model and

runs them together. Same metal, opposite slope, which is the clearest available evidence that

method matters more than hardware. Full analysis at macluster.novcog.us.com.
The two-machine case, and the ratio trap
The build this analysis is anchored to is an M3 Ultra Mac Studio with 256 GB paired with an M4

Pro Mac mini with 48 GB. Both machines have Thunderbolt 5, so both qualify for RDMA; Apple’s

technote uses an M4 Pro Mac mini as its own worked example, listing the three RDMA interfaces such

a machine exposes.
The constraint is not the 304 GB total but the 5.3-to-1 ratio. Tensor parallelism wants

comparable shards, each shard must fit the machine holding it, and every layer ends in a

synchronisation that completes when the slowest participant completes. EXO’s topology-aware

placement weights the split by device resources and measured link quality, which mitigates this

without removing it. The published speedups were measured on matched hardware.
The honest expectation for that pair: it expands what fits rather than transforming throughput,

and for any model that already fits in 256 GB the single Studio is faster, because no interconnect

beats no interconnect. Worked through at the two-node build.

Same hardware, same cables, opposite slope: one program sped up as machines were added and the other slowed down.

Why the old verdict went stale
For years the accurate summary of Mac clustering was that it bought capacity rather than speed.

That described pipeline parallelism, where each machine holds a block of the model’s

layers and work passes down the line, leaving one machine busy at a time. Adding machines fitted

larger models without making generation faster.
RDMA made a second approach practical. In tensor parallelism each machine holds a

slice of every layer and all of them compute simultaneously, synchronising at each layer. That

requires hundreds of round trips per token, which is viable only at very low latency. Exo Labs

publishes up to 1.8× on two devices and 3.2× on four, and states RDMA reduces

inter-Mac latency from roughly 300 microseconds to 3; the MLX pull request behind the feature is

cited at 300 microseconds down to under 50. Those figures measure at different points in the

stack, and this analysis quotes both rather than selecting the flattering one.
The most informative published result is not a number but a direction. Testing a four-machine

cluster Apple loaned him, Jeff Geerling found EXO speeding up as nodes were added while

llama.cpp’s layer-splitting RPC approach slowed down — same hardware, same cables,

opposite slope. Detail at what it is good at.




Three layers, not three choices
The most common error in this area is treating EXO, Apple’s MLX framework and RDMA over

Thunderbolt as competing options. They are not alternatives; they stack.
RDMA over Thunderbolt is a capability built into macOS, not software you install. Apple’s

technote TN3205 states it requires macOS 26.2 or later, an Apple silicon Mac, and Thunderbolt 5.

MLX is Apple’s open-source framework that actually runs the model, and MLX Distributed provides

the communication between machines; its RDMA backend is called jaccl. EXO, from Exo

Labs, is an orchestrator that sits on top of both, handling device discovery and deciding how to

split a model across the cluster.
Remove RDMA and a cluster still works, more slowly. Remove EXO and it still works, configured

by hand — Apple’s own research write-up drives a cluster with mlx.launch and no

orchestrator at all. Remove MLX and there is no inference left to distribute. The full breakdown

is at the three layers.
The measurements, and whose they are
Independent figures, from Geerling’s four-machine M3 Ultra cluster: Qwen3-235B at 8-bit reached

about 32 tokens per second on the full cluster; Kimi K2 Thinking, a trillion-parameter

mixture-of-experts model with roughly 32 billion parameters active per token, ran at around 30

tokens per second. On the HPL benchmark a single M3 Ultra produced 1.3 TFLOPS against 3.7 across

all four. The cluster drew under 250 watts and idled below 10.
Vendor figures, from Exo Labs: up to 1.8× on two devices, 3.2× on four, and a 99%

latency reduction over Thunderbolt 5. Apple reports up to 4.6× faster time-to-first-token on

an M5 Ultra Mac Studio cluster running a mixture-of-experts model above 800 billion parameters.
Geerling’s loaned cluster, 1.5 TB of unified memory across four machines, came to just under

$40,000 as configured, and he stated he could not justify that himself. Against the obvious

alternatives — Nvidia’s DGX Spark and AMD’s AI Max+ 395, both capped at 128 GB — he

found a single M3 Ultra matching four maxed-out systems of either kind.
What operators actually hit
Three failures recur in public reports. macOS Thunderbolt Bridge breaks RDMA outright, because

jaccl cannot address individual Thunderbolt ports through a bridged interface, and the

fix must be reapplied after every boot. RDMA ships disabled and is enabled only from Recovery mode

with rdma_ctl enable, once per machine, with no warning if a machine is missed. And

there are no Thunderbolt 5 switches, so the topology is a full mesh and roughly four machines is

the practical ceiling.
Geerling, who maintains the most widely used Ansible playbook for managing Macs, states plainly

that managing Linux clusters is easier, and notes that macOS cannot even run a system upgrade over

SSH. The capability is real and newly real; the operational maturity has not caught up. Field

reports are collected at field reports.
This analysis was produced under the Hidden State Drift method, which

separates what a vendor claims from what someone independent measured. The ten pre-publication

checks behind it are published at hiddenstatedrift.com/verify.

Novel Cognition is not affiliated with Apple or Exo Labs.

Part of the Frontier Watch Series: Read the previous investigation


More Coverage:
→ Read this investigation on Daily California Press
→ Coverage from FiorReports – Trump
"
      ["date_timestamp"]=>
      int(1792596944)
    }
    [3]=>
    array(11) {
      ["title"]=>
      string(61) "Strix Has 70,000 GitHub Stars and Everyone Describes It Wrong"
      ["link"]=>
      string(124) "https://nocarolinachronicle.com/strix-open-source-ai-pentester-four-agent-swarm-claim-checked-against-source-north_carolina/"
      ["dc"]=>
      array(1) {
        ["creator"]=>
        string(10) "Bill Moran"
      }
      ["pubdate"]=>
      string(31) "Mon, 19 Oct 2026 16:18:47 +0000"
      ["category"]=>
      string(193) "Newsagentic AIAI penetration testingAI securityapplication securityautonomous AI agentsdynamic skill injectionLLM agentsoffensive securityopen source security toolsprompt injectionStrixusestrix"
      ["guid"]=>
      string(40) "https://nocarolinachronicle.com/?p=53068"
      ["description"]=>
      string(463) "Strix, the open-source AI penetration-testing agent, carries 69,812 GitHub stars and a viral walkthrough describing a swarm of four specialised agents — Recon, Planning, Exploitation and Verification. Novel Cognition read the repository at main instead of the coverage: there is one agent template, a runtime agent graph, a 75-file skill library injected five at a time, and exactly one named role. Dynamic skill injection is real and is the part worth copying."
      ["content"]=>
      array(1) {
        ["encoded"]=>
        string(9990) "
By Bill Moran
A free tool that breaks into web applications has nearly 70,000 stars on GitHub, and almost every explainer written about it describes an architecture that is not in the code.
Strix (usestrix/strix) is an Apache-2.0 licensed Python project that points an LLM-driven agent at an application and has it behave like a penetration tester: map the attack surface, probe it, and write a working proof-of-concept for what it finds. It runs the target inside a Docker sandbox and it needs your own LLM API key. On 19 October 2026 the repository carried 69,812 stars.
A widely-shared October 2026 walkthrough describes it as a swarm of four specialised agents — Recon, Planning, Exploitation and Verification — paired with dynamic skill injection. Novel Cognition read the repository at main instead of the coverage. Half of that sentence is accurate, and it is not the half everyone repeats.
What is actually in the repository
There is no agent roster in the Strix source. There is one agent template — a single factory in strix/agents/factory.py and a single system_prompt.jinja — plus six tools the agent uses to build a graph of child agents while it is running: create_agent, send_message_to_agent, wait_for_agents, view_agent_graph, stop_agent and agent_finish.
No four-stage pipeline exists in code. The orchestrator reads the target, decides how many specialists it needs and what each one is for, and names its own children; the project’s own worked example is “Auth Specialist”, not “Recon Agent”.
Exactly one named role is fixed in the prompt: a Validation Agent. It recurs in system_prompt.jinja and in skills/coordination/root_agent.md, and it is the gate nothing reaches the report without passing. That is the one architectural claim in the popular account that checks out — a verification role does exist. The other three boxes do not.
This matters beyond pedantry. A fixed pipeline has predictable failure modes you can reason about. A graph an LLM assembles at runtime has a different risk profile: the shape of the test run is itself a model output, so two runs against the same target can decompose the work differently. That is a real property of the design, and it is invisible to anyone who believes there are four fixed agents.

Strix, the open-source AI penetration-testing agent, carries 69,812 GitHub stars and a viral walkthrough describing a swarm of four specialised agents — Recon, Planning, Exploitation and Verification.





The part that is genuinely new
“Dynamic skill injection” is accurate, and it is the piece worth copying. Counting the markdown files under strix/skills/ on 19 October 2026 gives 75 skills across 11 categories — 29 of them vulnerability classes (SQLi, XSS, SSRF, IDOR, RCE, XXE, SSTI, race conditions, JWT and auth, deserialisation, prototype pollution, request smuggling, path traversal, mass assignment, NoSQL injection, open redirect, subdomain takeover, CSRF), 13 command-line playbooks for the sandbox binaries, and the rest covering coordination, reporting and reconnaissance.
These are not documentation. They are prompt payload. The library’s own README states the contract: when an agent is created it loads up to five skills relevant to its subtask, and those files are injected straight into that agent’s system prompt. A load_skill tool lets an agent pull more expertise mid-run rather than only at birth.
The five-skill cap is a context-budget decision dressed as an architectural one, and it is the right call. Strip the security domain away and the pattern generalises: instead of one enormous prompt encoding every technique an agent might need — where instructions compete, the model averages across them, and depth is lost to breadth — expertise becomes a runtime resource. The library grows without any individual agent’s prompt growing at all. Contributors add markdown files, and the contribution guide treats a new skill as a first-class contribution alongside code.
A few entries point somewhere new: agentic_system_security tests effective authority in agent and MCP tool ecosystems, and llm_prompt_injection sits in the same directory as SQL injection, with the same status. A security tool that ships a skill for attacking agent infrastructure is a fair marker of where the attack surface is moving.
What nobody has measured
The honest account has two gaps and one hard constraint, and we are not going to paper over any of them.
Nobody has published a detection rate. No false-positive rate either, and no cost per scan. We looked. If you see a figure for any of the three, ask what it was measured against; we could not source one, so this site does not carry one.
The cost model is not the licence. Strix is free. The tokens are not. You supply your own LLM key and a deep scan is a long agentic run, so the bill scales with how thoroughly the orchestrator decides to work — which is itself a model decision, not a setting you fix in advance.
The constraint is authorisation. This is a tool that writes working proofs-of-concept against live applications. It is lawful against systems you own or have written permission to test, and it is not lawful anywhere else. That is not a disclaimer bolted on the end; it is the whole difference between a penetration test and an offence.
One more correction worth making plainly: it is not new. “Someone just open-sourced an autonomous AI hacker” is off by more than a year. The repository was created on 5 August 2025 and accumulated roughly 70,000 stars over fourteen months before the October 2026 video introduced it to a lot of people as a fresh release. Development is active — the latest push when we looked was 18 October 2026 — but the project is not a new thing.
Why the real answer is better
A fixed four-agent pipeline would be the less impressive design. It would mean the authors had guessed in advance how every assessment decomposes — the same assumption that makes traditional scanners noisy, running the same checks in the same order regardless of what they are pointed at.
What Strix does instead is defer that decision to the run, manufacture narrow specialists out of a 75-file library, and keep exactly one fixed gate so nothing reaches the report without something trying to reproduce it. That is a more interesting bet and a more honest one. The four-box diagram is not a lie anyone told on purpose — Strix’s own feature list says “specialized AI agents for recon, exploitation, and post-exploitation”, which reads naturally as a list of what the agents are, and those four words are simply the right words for the phases of a penetration test. Add a reasonable assumption and the diagram writes itself.
The full analysis is at strix.novcog.us.com: what it actually is, the skill library, the swarm claim checked against the source, running it and the token bill, the limits and sources. Primary source: the repository itself, usestrix/strix, read at main on 19 October 2026.

Part of the Frontier Watch Series: Read the previous investigation


More Coverage:
→ Read this investigation on Daily California Press
→ Coverage from FiorReports – Trump
"
      }
      ["summary"]=>
      string(463) "Strix, the open-source AI penetration-testing agent, carries 69,812 GitHub stars and a viral walkthrough describing a swarm of four specialised agents — Recon, Planning, Exploitation and Verification. Novel Cognition read the repository at main instead of the coverage: there is one agent template, a runtime agent graph, a 75-file skill library injected five at a time, and exactly one named role. Dynamic skill injection is real and is the part worth copying."
      ["atom_content"]=>
      string(9990) "
By Bill Moran
A free tool that breaks into web applications has nearly 70,000 stars on GitHub, and almost every explainer written about it describes an architecture that is not in the code.
Strix (usestrix/strix) is an Apache-2.0 licensed Python project that points an LLM-driven agent at an application and has it behave like a penetration tester: map the attack surface, probe it, and write a working proof-of-concept for what it finds. It runs the target inside a Docker sandbox and it needs your own LLM API key. On 19 October 2026 the repository carried 69,812 stars.
A widely-shared October 2026 walkthrough describes it as a swarm of four specialised agents — Recon, Planning, Exploitation and Verification — paired with dynamic skill injection. Novel Cognition read the repository at main instead of the coverage. Half of that sentence is accurate, and it is not the half everyone repeats.
What is actually in the repository
There is no agent roster in the Strix source. There is one agent template — a single factory in strix/agents/factory.py and a single system_prompt.jinja — plus six tools the agent uses to build a graph of child agents while it is running: create_agent, send_message_to_agent, wait_for_agents, view_agent_graph, stop_agent and agent_finish.
No four-stage pipeline exists in code. The orchestrator reads the target, decides how many specialists it needs and what each one is for, and names its own children; the project’s own worked example is “Auth Specialist”, not “Recon Agent”.
Exactly one named role is fixed in the prompt: a Validation Agent. It recurs in system_prompt.jinja and in skills/coordination/root_agent.md, and it is the gate nothing reaches the report without passing. That is the one architectural claim in the popular account that checks out — a verification role does exist. The other three boxes do not.
This matters beyond pedantry. A fixed pipeline has predictable failure modes you can reason about. A graph an LLM assembles at runtime has a different risk profile: the shape of the test run is itself a model output, so two runs against the same target can decompose the work differently. That is a real property of the design, and it is invisible to anyone who believes there are four fixed agents.

Strix, the open-source AI penetration-testing agent, carries 69,812 GitHub stars and a viral walkthrough describing a swarm of four specialised agents — Recon, Planning, Exploitation and Verification.





The part that is genuinely new
“Dynamic skill injection” is accurate, and it is the piece worth copying. Counting the markdown files under strix/skills/ on 19 October 2026 gives 75 skills across 11 categories — 29 of them vulnerability classes (SQLi, XSS, SSRF, IDOR, RCE, XXE, SSTI, race conditions, JWT and auth, deserialisation, prototype pollution, request smuggling, path traversal, mass assignment, NoSQL injection, open redirect, subdomain takeover, CSRF), 13 command-line playbooks for the sandbox binaries, and the rest covering coordination, reporting and reconnaissance.
These are not documentation. They are prompt payload. The library’s own README states the contract: when an agent is created it loads up to five skills relevant to its subtask, and those files are injected straight into that agent’s system prompt. A load_skill tool lets an agent pull more expertise mid-run rather than only at birth.
The five-skill cap is a context-budget decision dressed as an architectural one, and it is the right call. Strip the security domain away and the pattern generalises: instead of one enormous prompt encoding every technique an agent might need — where instructions compete, the model averages across them, and depth is lost to breadth — expertise becomes a runtime resource. The library grows without any individual agent’s prompt growing at all. Contributors add markdown files, and the contribution guide treats a new skill as a first-class contribution alongside code.
A few entries point somewhere new: agentic_system_security tests effective authority in agent and MCP tool ecosystems, and llm_prompt_injection sits in the same directory as SQL injection, with the same status. A security tool that ships a skill for attacking agent infrastructure is a fair marker of where the attack surface is moving.
What nobody has measured
The honest account has two gaps and one hard constraint, and we are not going to paper over any of them.
Nobody has published a detection rate. No false-positive rate either, and no cost per scan. We looked. If you see a figure for any of the three, ask what it was measured against; we could not source one, so this site does not carry one.
The cost model is not the licence. Strix is free. The tokens are not. You supply your own LLM key and a deep scan is a long agentic run, so the bill scales with how thoroughly the orchestrator decides to work — which is itself a model decision, not a setting you fix in advance.
The constraint is authorisation. This is a tool that writes working proofs-of-concept against live applications. It is lawful against systems you own or have written permission to test, and it is not lawful anywhere else. That is not a disclaimer bolted on the end; it is the whole difference between a penetration test and an offence.
One more correction worth making plainly: it is not new. “Someone just open-sourced an autonomous AI hacker” is off by more than a year. The repository was created on 5 August 2025 and accumulated roughly 70,000 stars over fourteen months before the October 2026 video introduced it to a lot of people as a fresh release. Development is active — the latest push when we looked was 18 October 2026 — but the project is not a new thing.
Why the real answer is better
A fixed four-agent pipeline would be the less impressive design. It would mean the authors had guessed in advance how every assessment decomposes — the same assumption that makes traditional scanners noisy, running the same checks in the same order regardless of what they are pointed at.
What Strix does instead is defer that decision to the run, manufacture narrow specialists out of a 75-file library, and keep exactly one fixed gate so nothing reaches the report without something trying to reproduce it. That is a more interesting bet and a more honest one. The four-box diagram is not a lie anyone told on purpose — Strix’s own feature list says “specialized AI agents for recon, exploitation, and post-exploitation”, which reads naturally as a list of what the agents are, and those four words are simply the right words for the phases of a penetration test. Add a reasonable assumption and the diagram writes itself.
The full analysis is at strix.novcog.us.com: what it actually is, the skill library, the swarm claim checked against the source, running it and the token bill, the limits and sources. Primary source: the repository itself, usestrix/strix, read at main on 19 October 2026.

Part of the Frontier Watch Series: Read the previous investigation


More Coverage:
→ Read this investigation on Daily California Press
→ Coverage from FiorReports – Trump
"
      ["date_timestamp"]=>
      int(1792426727)
    }
    [4]=>
    array(11) {
      ["title"]=>
      string(80) "AI’s Black Box Might Be Opening. What a Top Researcher’s Claim Means for You"
      ["link"]=>
      string(115) "https://nocarolinachronicle.com/ai-black-box-opening-what-it-means-plain-english-goodfire-cammarata-north_carolina/"
      ["dc"]=>
      array(1) {
        ["creator"]=>
        string(10) "Bill Moran"
      }
      ["pubdate"]=>
      string(31) "Fri, 16 Oct 2026 16:44:53 +0000"
      ["category"]=>
      string(138) "NewsAI and SEOAI black boxAI explainedAI researchAI safetyAI trustartificial intelligenceGoodfirehow AI worksNick CammarataNovel Cognition"
      ["guid"]=>
      string(40) "https://nocarolinachronicle.com/?p=53065"
      ["description"]=>
      string(349) "For years AI researchers said the inside of an AI model is an unreadable black box. Now Nick Cammarata, writing with the AI lab Goodfire, says most of an AI's inner parts can be read after all. A plain-English guide to what that means for AI safety, for fixing AI and for anyone whose pages AI reads, what the experts said back, and three takeaways."
      ["content"]=>
      array(1) {
        ["encoded"]=>
        string(6325) "
By Bill Moran
For years, the people who build artificial intelligence have admitted something unsettling: they can’t fully say what goes on inside it. An AI model is billions of tiny parts and the numbers connecting them, and the working belief has been that those parts are a tangled mess, each one holding a jumble of unrelated jobs. That’s why AI gets called a black box.
This week one of the researchers who helped build that belief said the evidence for it was misread. In an essay published on 15 October 2026 with the AI lab Goodfire, Nick Cammarata argues that most of an AI’s inner parts can be read after all. If he’s right, it changes how AI can be checked for safety, how it can be fixed, and how much anyone can trust what it says. What follows is what it means in plain terms, and where to be careful.





For years AI researchers said the inside of an AI model is an unreadable black box. Now Nick Cammarata, writing with the AI lab Goodfire, says most of an AI’s inner parts can be read after all. A plai

The claim, without the jargon
Picture a filing cabinet where every drawer holds a random jumble: tax returns, a recipe and a love letter in one drawer, because the cabinet ran out of room. That was the picture of an AI’s insides, and it’s why researchers built second AI systems just to untangle the first one. Cammarata says the drawers are mostly labelled. In one free, open AI model, he estimates about nine in ten of the parts he checked have one main job. He calls it a working hypothesis and puts himself at about 85% sure.
His best-known example is a part inside an image AI that lit up for cat faces and for the fronts of cars. For years it was cited as proof that AI is a mess. Cammarata, who helped write some of that research, now says it was looking for a V shape all along: pointed cat faces, the V-shaped front of a car, V-shaped flames. Another part, written off by an automated tool as junk, turned out to track Colorado. When researchers turned it up, the AI started planning a hike.
Why it matters outside the lab
Safety. Today AI is mostly judged by what it says, and a model that has learned to mislead could say all the right things. Cammarata argues that a model’s own wiring has no reason to hide anything, so reading the inside could catch what the answers don’t show.
Fixing AI. When a model picks up a bad habit, the usual response is more training and hoping it sticks. If the parts can be read, a bad habit can be traced to where it came from and changed with small, careful fixes. For anyone asked to explain an AI’s decision to a customer, a regulator or a court, that’s a very different conversation.
The web. One part the researchers examined responds to specific signs on pages of health SEO spam, and the essay notes that a well-known research page opened with a giant embedded image that AI reading tools ran out of room inside. How a page is built shapes what AI takes from it.
Three takeaways
The pushback is real. Google DeepMind’s Neel Nanda called the work interesting but not a big update, and researcher Stephen Casper pointed out that knowing what most parts do isn’t the same as knowing what the whole machine does. Goodfire raised $150 million on the idea that AI can be read, and the detailed paper isn’t out yet.
So: the black box may be less black than we’ve been told. Ignore headlines that say AI is “solved” or the danger is over; the author doesn’t say that. And watch for the paper, and for someone without a financial stake to check the work. The model he studied is free to download, so they can.
Every number, every source and the full reaction are at alreadyclean.novcog.us.com, including the cat-and-car story, what the experts said and what isn’t proven yet. The essay: Neural Networks Are Already Clean (Goodfire, 15 October 2026).

Part of the Frontier Watch Series: Read the previous investigation


More Coverage:
→ Read this investigation on Daily California Press
→ Coverage from FiorReports – Trump
"
      }
      ["summary"]=>
      string(349) "For years AI researchers said the inside of an AI model is an unreadable black box. Now Nick Cammarata, writing with the AI lab Goodfire, says most of an AI's inner parts can be read after all. A plain-English guide to what that means for AI safety, for fixing AI and for anyone whose pages AI reads, what the experts said back, and three takeaways."
      ["atom_content"]=>
      string(6325) "
By Bill Moran
For years, the people who build artificial intelligence have admitted something unsettling: they can’t fully say what goes on inside it. An AI model is billions of tiny parts and the numbers connecting them, and the working belief has been that those parts are a tangled mess, each one holding a jumble of unrelated jobs. That’s why AI gets called a black box.
This week one of the researchers who helped build that belief said the evidence for it was misread. In an essay published on 15 October 2026 with the AI lab Goodfire, Nick Cammarata argues that most of an AI’s inner parts can be read after all. If he’s right, it changes how AI can be checked for safety, how it can be fixed, and how much anyone can trust what it says. What follows is what it means in plain terms, and where to be careful.





For years AI researchers said the inside of an AI model is an unreadable black box. Now Nick Cammarata, writing with the AI lab Goodfire, says most of an AI’s inner parts can be read after all. A plai

The claim, without the jargon
Picture a filing cabinet where every drawer holds a random jumble: tax returns, a recipe and a love letter in one drawer, because the cabinet ran out of room. That was the picture of an AI’s insides, and it’s why researchers built second AI systems just to untangle the first one. Cammarata says the drawers are mostly labelled. In one free, open AI model, he estimates about nine in ten of the parts he checked have one main job. He calls it a working hypothesis and puts himself at about 85% sure.
His best-known example is a part inside an image AI that lit up for cat faces and for the fronts of cars. For years it was cited as proof that AI is a mess. Cammarata, who helped write some of that research, now says it was looking for a V shape all along: pointed cat faces, the V-shaped front of a car, V-shaped flames. Another part, written off by an automated tool as junk, turned out to track Colorado. When researchers turned it up, the AI started planning a hike.
Why it matters outside the lab
Safety. Today AI is mostly judged by what it says, and a model that has learned to mislead could say all the right things. Cammarata argues that a model’s own wiring has no reason to hide anything, so reading the inside could catch what the answers don’t show.
Fixing AI. When a model picks up a bad habit, the usual response is more training and hoping it sticks. If the parts can be read, a bad habit can be traced to where it came from and changed with small, careful fixes. For anyone asked to explain an AI’s decision to a customer, a regulator or a court, that’s a very different conversation.
The web. One part the researchers examined responds to specific signs on pages of health SEO spam, and the essay notes that a well-known research page opened with a giant embedded image that AI reading tools ran out of room inside. How a page is built shapes what AI takes from it.
Three takeaways
The pushback is real. Google DeepMind’s Neel Nanda called the work interesting but not a big update, and researcher Stephen Casper pointed out that knowing what most parts do isn’t the same as knowing what the whole machine does. Goodfire raised $150 million on the idea that AI can be read, and the detailed paper isn’t out yet.
So: the black box may be less black than we’ve been told. Ignore headlines that say AI is “solved” or the danger is over; the author doesn’t say that. And watch for the paper, and for someone without a financial stake to check the work. The model he studied is free to download, so they can.
Every number, every source and the full reaction are at alreadyclean.novcog.us.com, including the cat-and-car story, what the experts said and what isn’t proven yet. The essay: Neural Networks Are Already Clean (Goodfire, 15 October 2026).

Part of the Frontier Watch Series: Read the previous investigation


More Coverage:
→ Read this investigation on Daily California Press
→ Coverage from FiorReports – Trump
"
      ["date_timestamp"]=>
      int(1792169093)
    }
    [5]=>
    array(11) {
      ["title"]=>
      string(87) "Neural Networks Are Already Clean? The Researcher Behind Superposition Changes His Mind"
      ["link"]=>
      string(127) "https://nocarolinachronicle.com/neural-networks-already-clean-cammarata-goodfire-superposition-interpretability-north_carolina/"
      ["dc"]=>
      array(1) {
        ["creator"]=>
        string(10) "Bill Moran"
      }
      ["pubdate"]=>
      string(31) "Fri, 16 Oct 2026 14:48:41 +0000"
      ["category"]=>
      string(185) "NewsAI researchAI safetyGoodfireInceptionV1mechanistic interpretabilityneural networksNeural Networks Are Already CleanNick CammaratapolysemanticityQwen3sparse autoencoderssuperposition"
      ["guid"]=>
      string(40) "https://nocarolinachronicle.com/?p=53062"
      ["description"]=>
      string(417) "Nick Cammarata, who co-authored the 2020 paper that made superposition part of interpretability's canon, argues in a Goodfire essay that the evidence was misread: the famous cat-and-car neuron detects a V shape, and about 90% of neuron poles in Qwen3-32B have one main job. Novel Cognition checked the quotes, the numbers and the reaction from researchers including Neel Nanda, Liv Gorton, Leo Gao and Stephen Casper."
      ["content"]=>
      array(1) {
        ["encoded"]=>
        string(7730) "
By Bill Moran
For most of a decade, interpretability researchers have worked on the assumption that you can’t study an AI model one neuron at a time. Networks run out of room, the theory goes, so they pack unrelated concepts into the same neuron because those concepts rarely collide. That’s superposition, and it’s why so much of the field trains a second model, usually a sparse autoencoder, to untangle the first one.
On 15 October 2026 Nick Cammarata, a co-author of the 2020 paper that made superposition part of the canon, published Neural Networks Are Already Clean with the interpretability lab Goodfire. His argument: the canonical evidence was misread, the famous cat-and-car neuron detects a V shape, and about 90% of neuron poles in the open Qwen3-32B model have one main job. Novel Cognition read all 27,000 words, checked the quotations it rests on, and went through the reaction.
The evidence he re-opened
InceptionV1 neuron 4e:55 has been the field’s standard example of a polysemantic neuron since 2017. It responds to cat faces and to car fronts, and the 2020 paper Zoom In said it “isn’t responding to some commonality of cars and cat faces.” Cammarata, a co-author of that paper, now reads it as a V-shape detector: its strongest inputs include V-shaped flames, fireboat water jets and fox faces, and its negative pole is an O.
The circuit Zoom In used to introduce superposition, a car feature spreading into unrelated dog neurons, was one Cammarata found in 2019. He now says the cars and dogs were grouped by shared pose, and that the visualization tool was reward hacked toward dogs. InceptionV1’s unreadable last layer, he estimates, goes from about 5% interpretable with old tools to about 99% in a ConvNeXt of the same size trained on the same data. When he starved layers of neurons, they got coarser instead of packing unrelated concepts together.
What remains of the mess, he argues, is recursive generalization: a neuron that detects “iy” grows into “diy” and then “do it yourself”, one related offshoot at a time.

Nick Cammarata, who co-authored the 2020 paper that made superposition part of interpretability’s canon, argues in a Goodfire essay that the evidence was misread: the famous cat-and-car neuron detects





What the 90% counts
The headline figure comes from preliminary audits of Qwen3-32B: about 90% of neuron poles with one main job, about 7% with two related jobs, and about 2-4% not yet understood. The essay itself says the share moves with thresholds and counts each neuron’s positive and negative sides separately. Agents did most of the labelling, Kimi K3 appears as a single attribution graph, and the technical companion paper with the language-model detail had not been published as of 16 October.
The comparison point is Anthropic’s 2023 Towards Monosemanticity, where the median neuron in a one-layer model scored 0 on an interpretability rubric. The same paragraph scored the median dictionary feature at 12, a figure the essay doesn’t mention. Goodfire, which raised a $150 million Series B in February 2026, has a commercial interest in models being readable; many of the image-model results can be checked in its public microscope.
How the field reacted
The reach was enormous: the launch post on X passed 2.7 million views in a day, and the Hacker News thread reached 743 points. Working interpretability researchers outside Goodfire were cooler. Neel Nanda, who leads mechanistic interpretability at Google DeepMind, called it interesting work but not a significant update and said he still thinks superposition is real. Liv Gorton argued that superposition is a claim about geometry, more features than dimensions, so recursive generalization doesn’t compete with it. Leo Gao said that if you can’t simulate a neuron you haven’t explained it, and Stephen Casper noted that understanding 90% of the behaviour of 90% of neurons doesn’t add up to 81% of the network.
Cammarata answered critics on X, conceded that his definition of monosemantic was loose, and said the technical paper is maybe a few weeks away. Some amplifiers went further than the essay does; it never calls superposition a hoax, and it labels its biggest conclusion highly speculative.
What would settle it
The re-reading of the old image-model evidence is the strongest part, because it comes with pictures and a public microscope. The language-model numbers wait on the companion paper. Cammarata names his own bar: one convincing superposition circuit in a real network. Qwen3-32B is open weights, so anyone can go looking.
The full file, with every number, every quote check and the reaction by venue, is at alreadyclean.novcog.us.com: the cat-and-car neuron, the numbers, the reaction and what the essay doesn’t show yet. Primary source: Neural Networks Are Already Clean (Goodfire, 15 October 2026).

Part of the Frontier Watch Series: Read the previous investigation


More Coverage:
→ Read this investigation on Daily California Press
→ Coverage from FiorReports – Trump
"
      }
      ["summary"]=>
      string(417) "Nick Cammarata, who co-authored the 2020 paper that made superposition part of interpretability's canon, argues in a Goodfire essay that the evidence was misread: the famous cat-and-car neuron detects a V shape, and about 90% of neuron poles in Qwen3-32B have one main job. Novel Cognition checked the quotes, the numbers and the reaction from researchers including Neel Nanda, Liv Gorton, Leo Gao and Stephen Casper."
      ["atom_content"]=>
      string(7730) "
By Bill Moran
For most of a decade, interpretability researchers have worked on the assumption that you can’t study an AI model one neuron at a time. Networks run out of room, the theory goes, so they pack unrelated concepts into the same neuron because those concepts rarely collide. That’s superposition, and it’s why so much of the field trains a second model, usually a sparse autoencoder, to untangle the first one.
On 15 October 2026 Nick Cammarata, a co-author of the 2020 paper that made superposition part of the canon, published Neural Networks Are Already Clean with the interpretability lab Goodfire. His argument: the canonical evidence was misread, the famous cat-and-car neuron detects a V shape, and about 90% of neuron poles in the open Qwen3-32B model have one main job. Novel Cognition read all 27,000 words, checked the quotations it rests on, and went through the reaction.
The evidence he re-opened
InceptionV1 neuron 4e:55 has been the field’s standard example of a polysemantic neuron since 2017. It responds to cat faces and to car fronts, and the 2020 paper Zoom In said it “isn’t responding to some commonality of cars and cat faces.” Cammarata, a co-author of that paper, now reads it as a V-shape detector: its strongest inputs include V-shaped flames, fireboat water jets and fox faces, and its negative pole is an O.
The circuit Zoom In used to introduce superposition, a car feature spreading into unrelated dog neurons, was one Cammarata found in 2019. He now says the cars and dogs were grouped by shared pose, and that the visualization tool was reward hacked toward dogs. InceptionV1’s unreadable last layer, he estimates, goes from about 5% interpretable with old tools to about 99% in a ConvNeXt of the same size trained on the same data. When he starved layers of neurons, they got coarser instead of packing unrelated concepts together.
What remains of the mess, he argues, is recursive generalization: a neuron that detects “iy” grows into “diy” and then “do it yourself”, one related offshoot at a time.

Nick Cammarata, who co-authored the 2020 paper that made superposition part of interpretability’s canon, argues in a Goodfire essay that the evidence was misread: the famous cat-and-car neuron detects





What the 90% counts
The headline figure comes from preliminary audits of Qwen3-32B: about 90% of neuron poles with one main job, about 7% with two related jobs, and about 2-4% not yet understood. The essay itself says the share moves with thresholds and counts each neuron’s positive and negative sides separately. Agents did most of the labelling, Kimi K3 appears as a single attribution graph, and the technical companion paper with the language-model detail had not been published as of 16 October.
The comparison point is Anthropic’s 2023 Towards Monosemanticity, where the median neuron in a one-layer model scored 0 on an interpretability rubric. The same paragraph scored the median dictionary feature at 12, a figure the essay doesn’t mention. Goodfire, which raised a $150 million Series B in February 2026, has a commercial interest in models being readable; many of the image-model results can be checked in its public microscope.
How the field reacted
The reach was enormous: the launch post on X passed 2.7 million views in a day, and the Hacker News thread reached 743 points. Working interpretability researchers outside Goodfire were cooler. Neel Nanda, who leads mechanistic interpretability at Google DeepMind, called it interesting work but not a significant update and said he still thinks superposition is real. Liv Gorton argued that superposition is a claim about geometry, more features than dimensions, so recursive generalization doesn’t compete with it. Leo Gao said that if you can’t simulate a neuron you haven’t explained it, and Stephen Casper noted that understanding 90% of the behaviour of 90% of neurons doesn’t add up to 81% of the network.
Cammarata answered critics on X, conceded that his definition of monosemantic was loose, and said the technical paper is maybe a few weeks away. Some amplifiers went further than the essay does; it never calls superposition a hoax, and it labels its biggest conclusion highly speculative.
What would settle it
The re-reading of the old image-model evidence is the strongest part, because it comes with pictures and a public microscope. The language-model numbers wait on the companion paper. Cammarata names his own bar: one convincing superposition circuit in a real network. Qwen3-32B is open weights, so anyone can go looking.
The full file, with every number, every quote check and the reaction by venue, is at alreadyclean.novcog.us.com: the cat-and-car neuron, the numbers, the reaction and what the essay doesn’t show yet. Primary source: Neural Networks Are Already Clean (Goodfire, 15 October 2026).

Part of the Frontier Watch Series: Read the previous investigation


More Coverage:
→ Read this investigation on Daily California Press
→ Coverage from FiorReports – Trump
"
      ["date_timestamp"]=>
      int(1792162121)
    }
    [6]=>
    array(11) {
      ["title"]=>
      string(108) "Most Voters Can’t See the Ideological Spectrum the Political Class Fights Over. The Evidence Dates to 1964"
      ["link"]=>
      string(133) "https://nocarolinachronicle.com/strategists-fallacy-ideological-spectrum-most-voters-cant-see-sixty-years-of-evidence-north_carolina/"
      ["dc"]=>
      array(1) {
        ["creator"]=>
        string(10) "Bill Moran"
      }
      ["pubdate"]=>
      string(31) "Sun, 11 Oct 2026 16:27:27 +0000"
      ["category"]=>
      string(189) "News2024 ElectionDeciding to WinDemocracy for RealistsDemocratic PartyG. Elliott Morrismoderationpersuadable votersPhilip Conversepolitical scienceStrategist's Fallacyswing votersWelcomePAC"
      ["guid"]=>
      string(40) "https://nocarolinachronicle.com/?p=53059"
      ["description"]=>
      string(567) "G. Elliott Morris named the Strategist's Fallacy in October 2025: the belief that voters hold positions and labels and match them to candidates like an entree off a menu. Guerin Green, publisher of Daily Political Press, asks why, if a persuadable marginal voter can't see ideological identity, like ultraviolet or infrared light, the political class is obsessed with it. Daily Political Press checked Morris's argument against sixty years of research, from Converse in 1964 to Achen and Bartels, and against the 2024 state canvasses. Every figure carries its source."
      ["content"]=>
      array(1) {
        ["encoded"]=>
        string(13210) "
By Bill Moran
“If a persuadable marginal voter can’t see ideological identity, like they can’t see ultraviolet or infrared light, why are pundits and the political class obsessed with it?” The question is Guerin Green’s, publisher of Daily Political Press, a Broadside News Group title, and it is the question a new seven-page file from Daily Political Press sets out to answer.
The obsession has a name. In October 2025 the data journalist G. Elliott Morris called it the Strategist’s Fallacy: the belief that voters hold concrete issue positions and ideological labels and match them to candidates the way a customer picks an entree off a dinner menu, so that the way to win is to tweak a candidate’s positions. He was answering Deciding to Win, a report published the day before by the centrist group WelcomePAC, which recommends that Democrats “affirmatively moderate” on immigration, crime and energy.
The political science behind Morris’s objection is sixty years old. Philip Converse sorted the 1956 electorate by how people talked about politics and found that only a small minority reasoned in ideological terms. Christopher Achen and Larry Bartels, in Democracy for Realists (Princeton, 2016), assembled the panel evidence that people join a party first and take on its positions afterwards. Morris’s own House model puts the premium for moderation at about 1.4 points of margin, down roughly 80 per cent since 2000, inside a plausible range from near zero to about three.
Daily Political Press has published the file at The Strategist’s Fallacy, checked: Morris’s argument as he made it, the research behind it, the 2024 margins by the state canvasses, and a page for what could and could not be verified, with a video walkthrough. The argument is Morris’s and is credited to him throughout. The spectrum is Green’s.
Sixty years of evidence
The research is laid out paper by paper on sixty years of evidence.
Philip Converse sorted the 1956 electorate’s open-ended interviews by how respondents reasoned about politics. By the figures usually reported from his 1964 chapter, 2.5 per cent qualified as ideologues and another 9 per cent came close: about one voter in nine. Roughly 42 per cent judged the parties by group interest; the rest judged the times, or offered no issue content at all. The file’s sources page flags those percentages as taken from secondary summaries, pending a check of Converse’s original table. Donald Kinder and Nathan Kalmoe found in 2017 that Converse’s ideological innocence holds nearly as well today. Morris, in the Boston Review in February 2026, puts the ideological share at 10 to 15 per cent, and adds from his own surveys that many apparent “moderates” sit at the midpoint because they provide no ideological content at all.
Christopher Achen and Larry Bartels, in Democracy for Realists (Princeton, 2016), call the menu model the folk theory of democracy and argue the causation runs the other way: people take a party from family, region, race and class, then take their positions from the party. Thomas Carsey and Geoffrey Layman (American Journal of Political Science 50(2), 2006) followed the same respondents through the 1992, 1994 and 1996 elections and found that voters who knew the parties differed on abortion, but did not much care, changed their minds to match their party. Michael Barber and Jeremy Pope (American Political Science Review 113(1), 2019) found that Republicans told Donald Trump held a position moved toward it, including toward liberal positions.
Voters do grade the party in power, but on a short window. Andrew Healy and Gabriel Lenz (American Journal of Political Science 58, 2014) found that people say they weigh a president’s whole term and in practice weigh the election year.

G. Elliott Morris named the Strategist’s Fallacy in October 2025: the belief that voters hold positions and labels and match them to candidates like an entree off a menu. Guerin Green, publisher of Da

2024 by the fundamentals
The model, the margins and the state canvasses are on the incumbent.
In May 2026 Morris published a forecasting model with two inputs and nothing about the campaign: the president’s net approval in June, from Gallup, and the two-year average of the University of Michigan’s consumer sentiment index. Fitted to every election from 1956 to 2020, with 2024 held out, it forecast about 48 per cent of the two-party vote for Kamala Harris, with an 80 per cent interval from 46.6 to 49.9. She received about 49.3. On conditions alone, by his model, Trump wins that election about nine times in ten.
It was not only an American result. John Burn-Murdoch of the Financial Times found that governing parties lost vote share in all ten of the major developed democracies in the ParlGov dataset that voted in 2024, the first time that had happened in a record that begins in 1905.
The margin is where the explanations come from. Trump carried Pennsylvania by 120,266 votes, by the Department of State’s certification of 4 December 2024; Michigan by 80,103, as compiled from the state canvass; and Wisconsin by 29,397, by the Wisconsin Elections Commission canvass of 29 November 2024. That is 229,766 votes across three states, out of 155,512,532 cast for president: about 0.15 per cent. Morris rounds the same figures to 230,000 of 152 million; the file uses the state counts and says so.
Morris’s word for a result that close is overdetermined. Presidential approval, inflation, real income and candidate favourability each move a national result by at least twice the point and a half that decided 2024, so no single factor can be shown to have lost it, and the same defeat is cited as proof the party was too liberal and proof it was too timid. The takes, he writes, reaffirm their authors’ priors. His word for the exercise is projection.
What Deciding to Win argues, and what Morris objected to
The report and Morris’s reply sit side by side on the report page of the file.
Deciding to Win, by Simon Bazelon, Lauren Harper Pope and Liam Kerr, was published by WelcomePAC on 27 October 2025 at decidingtowin.org. It draws on election results, public polling and online-panel surveys of more than 500,000 Americans between 13 November 2024 and 18 June 2025. Its thesis is that “since 2012, highly educated staffers, donors, advocacy groups, pundits, and elected officials have reshaped the Democratic Party’s agenda, decreasing our party’s focus on the economic issues that are the top concerns of the American people.” The remedy is to talk about costs and to “affirmatively moderate our positions on high-salience issues where voters distrust us and where progressive policies are unpopular, particularly on immigration, crime, and energy policy.” Part 11.4 sets the test for 2028: “Democrats should pick a nominee who understands what it takes to win elections in difficult terrain—and is willing to run on positions that majorities of Americans support, even if this sometimes requires breaking with the unpopular demands of progressive advocacy organizations, corporate interests, or the Democratic donor class.” Its evidence that moderates outperform is overperformance estimates credited to the elections site Split Ticket.
Morris answered the next day in his newsletter Strength In Numbers, crediting the report’s breadth of data and raising three objections. It ignores current polling showing Democrats recovering with voters who care about the economy. It says little about inflation, and treats 2024 as Democrat against Republican when it was incumbent against opposition. And it reads too much into noisy, unadjusted gaps between “moderate” and “progressive” candidates: in WelcomePAC’s own data, overperformance in one cycle predicts only 25 per cent of a candidate’s overperformance in the next.
His largest objection was the premise: the menu is how strategists decide, mapped onto voters who do not.




The persuasion problem
Both experiments and Morris’s holdout survey are on the persuasion problem.
A better pitch should move a menu voter. Joshua Kalla and David Broockman (American Political Science Review 112(1), 2018) pooled 40 field experiments on campaign contact with nine of their own: across the 49, the best estimate of general-election persuasion is zero, with narrow exceptions for heavily targeted unpopular positions and for early contact that decays. Deciding to Win concedes as much in Part 4.10, field programmes having “minimal impacts on changing voters’ minds,” then recommends changing positions instead.
Broockman and Christopher Skovron (American Political Science Review 112(3), 2018) found the instruments misread too: 3,765 state legislators, compared with their own constituents, overestimated support for conservative policies in both parties.
Morris looked for the persuadable voter in 5,509 registered voters pooled from four of his surveys, January–April 2026. The real holdouts, who disapprove of Trump, will not commit to a Democrat and carry no Republican marker, were 3.4 per cent. Of them, 83 per cent were undecided, about 40 per cent had not voted in 2024, 76.8 per cent were pure independents, and prices (38 per cent) and jobs (21) outranked crime (3). His conclusion: engagement, not persuasion.
Immigration shows the cost of photographing opinion rather than forecasting it. Early in 2025 strategists told Democrats to go quiet on it; ICE’s net approval then fell from +16 in February 2025 to −20 in January 2026, and Democrats won statewide in Georgia, Pennsylvania, Virginia and New Jersey with nominees keeping left-leaning positions on immigration, abortion and health care (Morris, Boston Review, February 2026).
Morris’s charge that the report lists abolishing prisons, DC statehood and Medicare for All as 2024 agenda items could not be found in the live report and sits on the file’s could-not-verify page, not here. Every document the file rests on is on its sources page.

Part of the Verified Series: Read the previous investigation


More Coverage:
→ Read this investigation on Daily California Press
→ Coverage from FiorReports – Trump
"
      }
      ["summary"]=>
      string(567) "G. Elliott Morris named the Strategist's Fallacy in October 2025: the belief that voters hold positions and labels and match them to candidates like an entree off a menu. Guerin Green, publisher of Daily Political Press, asks why, if a persuadable marginal voter can't see ideological identity, like ultraviolet or infrared light, the political class is obsessed with it. Daily Political Press checked Morris's argument against sixty years of research, from Converse in 1964 to Achen and Bartels, and against the 2024 state canvasses. Every figure carries its source."
      ["atom_content"]=>
      string(13210) "
By Bill Moran
“If a persuadable marginal voter can’t see ideological identity, like they can’t see ultraviolet or infrared light, why are pundits and the political class obsessed with it?” The question is Guerin Green’s, publisher of Daily Political Press, a Broadside News Group title, and it is the question a new seven-page file from Daily Political Press sets out to answer.
The obsession has a name. In October 2025 the data journalist G. Elliott Morris called it the Strategist’s Fallacy: the belief that voters hold concrete issue positions and ideological labels and match them to candidates the way a customer picks an entree off a dinner menu, so that the way to win is to tweak a candidate’s positions. He was answering Deciding to Win, a report published the day before by the centrist group WelcomePAC, which recommends that Democrats “affirmatively moderate” on immigration, crime and energy.
The political science behind Morris’s objection is sixty years old. Philip Converse sorted the 1956 electorate by how people talked about politics and found that only a small minority reasoned in ideological terms. Christopher Achen and Larry Bartels, in Democracy for Realists (Princeton, 2016), assembled the panel evidence that people join a party first and take on its positions afterwards. Morris’s own House model puts the premium for moderation at about 1.4 points of margin, down roughly 80 per cent since 2000, inside a plausible range from near zero to about three.
Daily Political Press has published the file at The Strategist’s Fallacy, checked: Morris’s argument as he made it, the research behind it, the 2024 margins by the state canvasses, and a page for what could and could not be verified, with a video walkthrough. The argument is Morris’s and is credited to him throughout. The spectrum is Green’s.
Sixty years of evidence
The research is laid out paper by paper on sixty years of evidence.
Philip Converse sorted the 1956 electorate’s open-ended interviews by how respondents reasoned about politics. By the figures usually reported from his 1964 chapter, 2.5 per cent qualified as ideologues and another 9 per cent came close: about one voter in nine. Roughly 42 per cent judged the parties by group interest; the rest judged the times, or offered no issue content at all. The file’s sources page flags those percentages as taken from secondary summaries, pending a check of Converse’s original table. Donald Kinder and Nathan Kalmoe found in 2017 that Converse’s ideological innocence holds nearly as well today. Morris, in the Boston Review in February 2026, puts the ideological share at 10 to 15 per cent, and adds from his own surveys that many apparent “moderates” sit at the midpoint because they provide no ideological content at all.
Christopher Achen and Larry Bartels, in Democracy for Realists (Princeton, 2016), call the menu model the folk theory of democracy and argue the causation runs the other way: people take a party from family, region, race and class, then take their positions from the party. Thomas Carsey and Geoffrey Layman (American Journal of Political Science 50(2), 2006) followed the same respondents through the 1992, 1994 and 1996 elections and found that voters who knew the parties differed on abortion, but did not much care, changed their minds to match their party. Michael Barber and Jeremy Pope (American Political Science Review 113(1), 2019) found that Republicans told Donald Trump held a position moved toward it, including toward liberal positions.
Voters do grade the party in power, but on a short window. Andrew Healy and Gabriel Lenz (American Journal of Political Science 58, 2014) found that people say they weigh a president’s whole term and in practice weigh the election year.

G. Elliott Morris named the Strategist’s Fallacy in October 2025: the belief that voters hold positions and labels and match them to candidates like an entree off a menu. Guerin Green, publisher of Da

2024 by the fundamentals
The model, the margins and the state canvasses are on the incumbent.
In May 2026 Morris published a forecasting model with two inputs and nothing about the campaign: the president’s net approval in June, from Gallup, and the two-year average of the University of Michigan’s consumer sentiment index. Fitted to every election from 1956 to 2020, with 2024 held out, it forecast about 48 per cent of the two-party vote for Kamala Harris, with an 80 per cent interval from 46.6 to 49.9. She received about 49.3. On conditions alone, by his model, Trump wins that election about nine times in ten.
It was not only an American result. John Burn-Murdoch of the Financial Times found that governing parties lost vote share in all ten of the major developed democracies in the ParlGov dataset that voted in 2024, the first time that had happened in a record that begins in 1905.
The margin is where the explanations come from. Trump carried Pennsylvania by 120,266 votes, by the Department of State’s certification of 4 December 2024; Michigan by 80,103, as compiled from the state canvass; and Wisconsin by 29,397, by the Wisconsin Elections Commission canvass of 29 November 2024. That is 229,766 votes across three states, out of 155,512,532 cast for president: about 0.15 per cent. Morris rounds the same figures to 230,000 of 152 million; the file uses the state counts and says so.
Morris’s word for a result that close is overdetermined. Presidential approval, inflation, real income and candidate favourability each move a national result by at least twice the point and a half that decided 2024, so no single factor can be shown to have lost it, and the same defeat is cited as proof the party was too liberal and proof it was too timid. The takes, he writes, reaffirm their authors’ priors. His word for the exercise is projection.
What Deciding to Win argues, and what Morris objected to
The report and Morris’s reply sit side by side on the report page of the file.
Deciding to Win, by Simon Bazelon, Lauren Harper Pope and Liam Kerr, was published by WelcomePAC on 27 October 2025 at decidingtowin.org. It draws on election results, public polling and online-panel surveys of more than 500,000 Americans between 13 November 2024 and 18 June 2025. Its thesis is that “since 2012, highly educated staffers, donors, advocacy groups, pundits, and elected officials have reshaped the Democratic Party’s agenda, decreasing our party’s focus on the economic issues that are the top concerns of the American people.” The remedy is to talk about costs and to “affirmatively moderate our positions on high-salience issues where voters distrust us and where progressive policies are unpopular, particularly on immigration, crime, and energy policy.” Part 11.4 sets the test for 2028: “Democrats should pick a nominee who understands what it takes to win elections in difficult terrain—and is willing to run on positions that majorities of Americans support, even if this sometimes requires breaking with the unpopular demands of progressive advocacy organizations, corporate interests, or the Democratic donor class.” Its evidence that moderates outperform is overperformance estimates credited to the elections site Split Ticket.
Morris answered the next day in his newsletter Strength In Numbers, crediting the report’s breadth of data and raising three objections. It ignores current polling showing Democrats recovering with voters who care about the economy. It says little about inflation, and treats 2024 as Democrat against Republican when it was incumbent against opposition. And it reads too much into noisy, unadjusted gaps between “moderate” and “progressive” candidates: in WelcomePAC’s own data, overperformance in one cycle predicts only 25 per cent of a candidate’s overperformance in the next.
His largest objection was the premise: the menu is how strategists decide, mapped onto voters who do not.




The persuasion problem
Both experiments and Morris’s holdout survey are on the persuasion problem.
A better pitch should move a menu voter. Joshua Kalla and David Broockman (American Political Science Review 112(1), 2018) pooled 40 field experiments on campaign contact with nine of their own: across the 49, the best estimate of general-election persuasion is zero, with narrow exceptions for heavily targeted unpopular positions and for early contact that decays. Deciding to Win concedes as much in Part 4.10, field programmes having “minimal impacts on changing voters’ minds,” then recommends changing positions instead.
Broockman and Christopher Skovron (American Political Science Review 112(3), 2018) found the instruments misread too: 3,765 state legislators, compared with their own constituents, overestimated support for conservative policies in both parties.
Morris looked for the persuadable voter in 5,509 registered voters pooled from four of his surveys, January–April 2026. The real holdouts, who disapprove of Trump, will not commit to a Democrat and carry no Republican marker, were 3.4 per cent. Of them, 83 per cent were undecided, about 40 per cent had not voted in 2024, 76.8 per cent were pure independents, and prices (38 per cent) and jobs (21) outranked crime (3). His conclusion: engagement, not persuasion.
Immigration shows the cost of photographing opinion rather than forecasting it. Early in 2025 strategists told Democrats to go quiet on it; ICE’s net approval then fell from +16 in February 2025 to −20 in January 2026, and Democrats won statewide in Georgia, Pennsylvania, Virginia and New Jersey with nominees keeping left-leaning positions on immigration, abortion and health care (Morris, Boston Review, February 2026).
Morris’s charge that the report lists abolishing prisons, DC statehood and Medicare for All as 2024 agenda items could not be found in the live report and sits on the file’s could-not-verify page, not here. Every document the file rests on is on its sources page.

Part of the Verified Series: Read the previous investigation


More Coverage:
→ Read this investigation on Daily California Press
→ Coverage from FiorReports – Trump
"
      ["date_timestamp"]=>
      int(1791736047)
    }
    [7]=>
    array(11) {
      ["title"]=>
      string(103) "On William Gray’s 97th Birthday, the Hurricane Pioneer’s Own Cooling Forecast Has Started to Expire"
      ["link"]=>
      string(110) "https://nocarolinachronicle.com/william-gray-cooling-forecast-2025-2030-climate-record-audited-north_carolina/"
      ["dc"]=>
      array(1) {
        ["creator"]=>
        string(10) "Bill Moran"
      }
      ["pubdate"]=>
      string(31) "Fri, 09 Oct 2026 16:11:31 +0000"
      ["category"]=>
      string(189) "Newsclimate recordclimate sensitivityColorado State Universityglobal coolinghurricane forecastingIPCC AR6NOAAocean circulationseasonal hurricane forecastwater vapour feedbackWilliam GrayWMO"
      ["guid"]=>
      string(40) "https://nocarolinachronicle.com/?p=53056"
      ["description"]=>
      string(594) "William Gray invented seasonal hurricane forecasting at Colorado State University and spent his last twenty years predicting global cooling. North Denver News audited seven of his claims against the observational record: three hold up. The only cooling forecast he put a date on ran to about 2025-2030; the WMO's eight-dataset analysis makes 2015-2025 the eleven warmest years on record, and August 2026 the warmest August measured. The report also corrects five errors in the review document it started from, including two cooling forecasts attributed to Gray that could not be sourced at all."
      ["content"]=>
      array(1) {
        ["encoded"]=>
        string(10711) "
By Bill Moran
“Note the recent 1975-2000 warming trend has not continued, and we are currently in a weak global cooling period that should continue, I believe, until about 2025-2030,” wrote William Mason Gray in a 2012 paper for the Heartland Institute. Gray, the pioneer of seasonal hurricane forecasting, spent his final years predicting a global cooling trend that contradicted mainstream climate science. A new analysis by North Denver News, titled The Colorado Denialist, tested Gray’s claims against observational records through October 2026. The findings reveal that while some of Gray’s assertions held true, his specific prediction for a cooling period ending by 2030 has been debunked by the warmest decade in recorded history.
The number the case rests on
Gray’s own arithmetic is reproduced step by step on the sensitivity page; his water-vapour claim and his ocean-circulation argument have their own pages.
Gray’s climate position reduces to one figure, and to his credit he published the arithmetic behind it rather than asserting it. In the 2012 paper he takes the direct warming from a doubling of carbon dioxide and halves it, to about 0.5 °C. Where climate models find a strongly positive water-vapour and cloud feedback worth around 2 °C, he argues the feedback runs the other way, at about −0.2 °C. The result, in his words, is “only about 0.3°C, which is only about one-tenth of the global warming projected by nearly all of the recent GCM climate simulations.”
The IPCC’s Sixth Assessment Report gives a best estimate for equilibrium climate sensitivity of 3 °C, with a likely range of 2.5 to 4 °C, and assesses a value below 1.5 °C as effectively ruled out — a conclusion drawn from process understanding, the historical record, palaeoclimate evidence and emergent constraints together, not from models alone.
The step that fails is the first one. Standard radiative physics gives roughly 1 °C of surface warming for doubled carbon dioxide with water-vapour feedback left out entirely. Gray’s halving of that term is where the factor of ten begins, and reviewers at the time traced it to the way he balanced top-of-atmosphere forcing against changes in evaporation. It is an energy budget that does not close, not a difference of opinion about clouds.

Every one of those start dates has passed.

Where Gray was right, and where the conclusion does not follow
The three claims that survive the audit are set out on what he got right.
Gray argued for years — loudly, and while the confident public line ran the other way — that you cannot get from a warmer ocean to more hurricanes by arithmetic, and that historical storm counts are contaminated by changes in how storms were observed.
NOAA’s Geophysical Fluid Dynamics Laboratory, in its most recent revision, concludes that the historical Atlantic hurricane data “do not provide compelling evidence” for a substantial greenhouse-driven century-scale increase in tropical storm or hurricane frequency, and that “most modeling studies project a decrease (or little change) in the global frequency of all tropical cyclones combined.” That is Gray’s position, reached by the institution he spent his later years arguing with.
The same assessment projects, with medium-to-high confidence, that rainfall rates increase, that average intensity increases, and that the proportion of storms reaching Category 4 or 5 increases. Rapid intensification is also projected to rise. Fewer storms, each one wetter and stronger, arriving on top of a higher sea level, is not a smaller problem.
Gray’s emphasis on multidecadal Atlantic variability also holds up, with the qualification that aerosols and greenhouse forcing are now understood to contribute to what he treated as purely internal. The Sahel is the clearest case: he tied a warm North Atlantic to wetter Sahel summers and the rains did partly recover, but the recovery is uneven east to west and the literature attributes it to a mixture of Atlantic and Indian Ocean temperatures, aerosols, greenhouse forcing and internal variability. He saw the pattern early. He was too certain about its cause.
The forecast he kept moving
The dated forecasts are set out in full, with the temperature record behind them, on the forecast page of the dossier.
Gray did something most participants in this argument never do: he made claims precise enough to be killed by data, and then he put dates on them. Collected in one place, the dated forecasts show a prediction being reissued rather than tested.

28 September 2005, testimony to the U.S. Senate Committee on Environment and Public Works: the warming trend would end and “in a few years we will start to see a weak cooling trend.”
28 May 2006, The Washington Post: “It may warm another three, five, eight years, and then it will start to cool.”
5 June 2006, The Denver Post: “My belief is that three, four years from now, the globe will start to cool again.”
2007, NBC News: the Earth could begin cooling on its own within five to ten years.
May 2012, Heartland Institute conference paper: the cooling is already running, since about 2000, and will continue “until about 2025-2030.”

Every one of those start dates has passed. None was followed by cooling. The 2012 claim — that a weak cooling had been under way since about 2000 — describes a period that contains the eleven warmest years on record.
A scientist revising a forecast is not a scandal; revision is the work. What the record shows is that every revision pushed the cooling further out, and the 2012 paper does not say the earlier forecasts were wrong.




What this file could not verify
This report began from a supplied review document carrying 37 references. Checking it against primary sources changed five things, four of which were errors that would otherwise have been published, and one of which was unfair to Gray. All are set out in full on the dossier’s corrections and limits page, and every document this report rests on is listed on its sources and method page.
Two cooling forecasts were dropped. The review built its central argument on Gray forecasts dated 1996, 2009 and 2012, and footnoted all three to a single reference: Gray’s curriculum vitae, a document that cannot carry a forecast wording. The 1996 forecast could not be found — not in his publication list, and not in catalogues that record individual newspaper interviews he gave. The 2009 forecast could not be sourced to Gray at all; its pairing of a 2001 start with a 2030 horizon corresponds to work published by other researchers that year. Neither is used here.
One number was corrected downward, against the story. The review credited Atlantic seasonal outlooks with 40 to 60 per cent skill in August, citing a 2017 review chapter. That chapter contains no such table: it reports qualitative bands and a single pair of percentages, skill rising from 3 per cent in the preceding December to 29 per cent in early August.
One criticism was corrected in Gray’s favour. He is commonly said to have confused relative humidity with specific humidity. He did not: his paper claims a reduction in both in the upper troposphere, with a table of observed changes behind it. That is a cleaner empirical claim than the one he is usually accused of making, and it is the one that satellite records contradict.
Two further claims in the review — a characterisation of the peer-review record, and an account of a National Science Foundation proposal review — could not be verified and are reported as unverified rather than asserted.

Part of the Verified Series: Read the previous investigation


More Coverage:
→ Read this investigation on Daily California Press
→ Coverage from FiorReports – Trump
"
      }
      ["summary"]=>
      string(594) "William Gray invented seasonal hurricane forecasting at Colorado State University and spent his last twenty years predicting global cooling. North Denver News audited seven of his claims against the observational record: three hold up. The only cooling forecast he put a date on ran to about 2025-2030; the WMO's eight-dataset analysis makes 2015-2025 the eleven warmest years on record, and August 2026 the warmest August measured. The report also corrects five errors in the review document it started from, including two cooling forecasts attributed to Gray that could not be sourced at all."
      ["atom_content"]=>
      string(10711) "
By Bill Moran
“Note the recent 1975-2000 warming trend has not continued, and we are currently in a weak global cooling period that should continue, I believe, until about 2025-2030,” wrote William Mason Gray in a 2012 paper for the Heartland Institute. Gray, the pioneer of seasonal hurricane forecasting, spent his final years predicting a global cooling trend that contradicted mainstream climate science. A new analysis by North Denver News, titled The Colorado Denialist, tested Gray’s claims against observational records through October 2026. The findings reveal that while some of Gray’s assertions held true, his specific prediction for a cooling period ending by 2030 has been debunked by the warmest decade in recorded history.
The number the case rests on
Gray’s own arithmetic is reproduced step by step on the sensitivity page; his water-vapour claim and his ocean-circulation argument have their own pages.
Gray’s climate position reduces to one figure, and to his credit he published the arithmetic behind it rather than asserting it. In the 2012 paper he takes the direct warming from a doubling of carbon dioxide and halves it, to about 0.5 °C. Where climate models find a strongly positive water-vapour and cloud feedback worth around 2 °C, he argues the feedback runs the other way, at about −0.2 °C. The result, in his words, is “only about 0.3°C, which is only about one-tenth of the global warming projected by nearly all of the recent GCM climate simulations.”
The IPCC’s Sixth Assessment Report gives a best estimate for equilibrium climate sensitivity of 3 °C, with a likely range of 2.5 to 4 °C, and assesses a value below 1.5 °C as effectively ruled out — a conclusion drawn from process understanding, the historical record, palaeoclimate evidence and emergent constraints together, not from models alone.
The step that fails is the first one. Standard radiative physics gives roughly 1 °C of surface warming for doubled carbon dioxide with water-vapour feedback left out entirely. Gray’s halving of that term is where the factor of ten begins, and reviewers at the time traced it to the way he balanced top-of-atmosphere forcing against changes in evaporation. It is an energy budget that does not close, not a difference of opinion about clouds.

Every one of those start dates has passed.

Where Gray was right, and where the conclusion does not follow
The three claims that survive the audit are set out on what he got right.
Gray argued for years — loudly, and while the confident public line ran the other way — that you cannot get from a warmer ocean to more hurricanes by arithmetic, and that historical storm counts are contaminated by changes in how storms were observed.
NOAA’s Geophysical Fluid Dynamics Laboratory, in its most recent revision, concludes that the historical Atlantic hurricane data “do not provide compelling evidence” for a substantial greenhouse-driven century-scale increase in tropical storm or hurricane frequency, and that “most modeling studies project a decrease (or little change) in the global frequency of all tropical cyclones combined.” That is Gray’s position, reached by the institution he spent his later years arguing with.
The same assessment projects, with medium-to-high confidence, that rainfall rates increase, that average intensity increases, and that the proportion of storms reaching Category 4 or 5 increases. Rapid intensification is also projected to rise. Fewer storms, each one wetter and stronger, arriving on top of a higher sea level, is not a smaller problem.
Gray’s emphasis on multidecadal Atlantic variability also holds up, with the qualification that aerosols and greenhouse forcing are now understood to contribute to what he treated as purely internal. The Sahel is the clearest case: he tied a warm North Atlantic to wetter Sahel summers and the rains did partly recover, but the recovery is uneven east to west and the literature attributes it to a mixture of Atlantic and Indian Ocean temperatures, aerosols, greenhouse forcing and internal variability. He saw the pattern early. He was too certain about its cause.
The forecast he kept moving
The dated forecasts are set out in full, with the temperature record behind them, on the forecast page of the dossier.
Gray did something most participants in this argument never do: he made claims precise enough to be killed by data, and then he put dates on them. Collected in one place, the dated forecasts show a prediction being reissued rather than tested.

28 September 2005, testimony to the U.S. Senate Committee on Environment and Public Works: the warming trend would end and “in a few years we will start to see a weak cooling trend.”
28 May 2006, The Washington Post: “It may warm another three, five, eight years, and then it will start to cool.”
5 June 2006, The Denver Post: “My belief is that three, four years from now, the globe will start to cool again.”
2007, NBC News: the Earth could begin cooling on its own within five to ten years.
May 2012, Heartland Institute conference paper: the cooling is already running, since about 2000, and will continue “until about 2025-2030.”

Every one of those start dates has passed. None was followed by cooling. The 2012 claim — that a weak cooling had been under way since about 2000 — describes a period that contains the eleven warmest years on record.
A scientist revising a forecast is not a scandal; revision is the work. What the record shows is that every revision pushed the cooling further out, and the 2012 paper does not say the earlier forecasts were wrong.




What this file could not verify
This report began from a supplied review document carrying 37 references. Checking it against primary sources changed five things, four of which were errors that would otherwise have been published, and one of which was unfair to Gray. All are set out in full on the dossier’s corrections and limits page, and every document this report rests on is listed on its sources and method page.
Two cooling forecasts were dropped. The review built its central argument on Gray forecasts dated 1996, 2009 and 2012, and footnoted all three to a single reference: Gray’s curriculum vitae, a document that cannot carry a forecast wording. The 1996 forecast could not be found — not in his publication list, and not in catalogues that record individual newspaper interviews he gave. The 2009 forecast could not be sourced to Gray at all; its pairing of a 2001 start with a 2030 horizon corresponds to work published by other researchers that year. Neither is used here.
One number was corrected downward, against the story. The review credited Atlantic seasonal outlooks with 40 to 60 per cent skill in August, citing a 2017 review chapter. That chapter contains no such table: it reports qualitative bands and a single pair of percentages, skill rising from 3 per cent in the preceding December to 29 per cent in early August.
One criticism was corrected in Gray’s favour. He is commonly said to have confused relative humidity with specific humidity. He did not: his paper claims a reduction in both in the upper troposphere, with a table of observed changes behind it. That is a cleaner empirical claim than the one he is usually accused of making, and it is the one that satellite records contradict.
Two further claims in the review — a characterisation of the peer-review record, and an account of a National Science Foundation proposal review — could not be verified and are reported as unverified rather than asserted.

Part of the Verified Series: Read the previous investigation


More Coverage:
→ Read this investigation on Daily California Press
→ Coverage from FiorReports – Trump
"
      ["date_timestamp"]=>
      int(1791562291)
    }
    [8]=>
    array(11) {
      ["title"]=>
      string(72) "Charlotte CATS rider hit and killed by bus she had just left, police say"
      ["link"]=>
      string(68) "https://nocarolinachronicle.com/charlotte-cats-rider-hit-killed-bus/"
      ["dc"]=>
      array(1) {
        ["creator"]=>
        string(10) "Bill Moran"
      }
      ["pubdate"]=>
      string(31) "Sat, 03 Oct 2026 19:50:11 +0000"
      ["category"]=>
      string(64) "NewsCharlotte Area TransitNorth Graham StreetPedestrian Accident"
      ["guid"]=>
      string(68) "https://nocarolinachronicle.com/charlotte-cats-rider-hit-killed-bus/"
      ["description"]=>
      string(1013) "
A 49-year-old woman was struck and killed by a Charlotte Area Transit System bus she had just exited on North Graham Street, police report."
      ["content"]=>
      array(1) {
        ["encoded"]=>
        string(3557) "
A 49-year-old woman was struck and killed by a Charlotte Area Transit System bus she had just exited Friday afternoon in the 1300 block of North Graham Street in Charlotte, police said. According to the Charlotte-Mecklenburg Police Department, the woman walked back toward the bus as it was moving away from the stop and was hit in the roadway; the investigation is ongoing.
The victim was identified by the Charlotte-Mecklenburg Police Department as 49-year-old Ebony Wingate. Officers arrived at the scene in the 1300 block of North Graham Street, near the intersection with Dalton Avenue, at approximately 1:10 p.m. on Friday, October 2, 2026. According to police,
Wingate was found unresponsive in the roadway and was pronounced dead at the scene by emergency medical personnel.
Preliminary information from CMPD indicates that Wingate had just exited a Charlotte Area Transit System bus at a designated stop when she walked back toward the bus as it was pulling away. The bus had begun moving as Wingate entered the street, at which point she was struck by the vehicle. The department described this account as preliminary and said the investigation is ongoing.
No other injuries or victims were reported in connection with the incident. CMPD’s Major Crash Investigation Unit has taken over the investigation, categorizing the case as a fatal crash involving a transit bus and pedestrian. As of October 3, police had not announced whether any charges or citations had been issued to the bus driver.
The crash location is situated north of Uptown Charlotte, a detail confirmed by local news sources. The Charlotte Observer reported the incident on the morning of October 3, with an update later that afternoon. WBTV also covered the crash, citing CMPD as the source for the timeline and details.
Authorities have asked anyone with information or who may have witnessed the collision to contact Detective Justin Kupfer at 704-432-2169, extension 1. Anonymous tips can be submitted to Charlotte Crime Stoppers at 704-334-1600. No additional witness statements have been publicly released.
Further details about Wingate’s background or residence have not been provided by the police or in public reports. The investigation remains active as officials continue to gather information and determine the full circumstances surrounding the fatal collision.

"
      }
      ["summary"]=>
      string(1013) "
A 49-year-old woman was struck and killed by a Charlotte Area Transit System bus she had just exited on North Graham Street, police report."
      ["atom_content"]=>
      string(3557) "
A 49-year-old woman was struck and killed by a Charlotte Area Transit System bus she had just exited Friday afternoon in the 1300 block of North Graham Street in Charlotte, police said. According to the Charlotte-Mecklenburg Police Department, the woman walked back toward the bus as it was moving away from the stop and was hit in the roadway; the investigation is ongoing.
The victim was identified by the Charlotte-Mecklenburg Police Department as 49-year-old Ebony Wingate. Officers arrived at the scene in the 1300 block of North Graham Street, near the intersection with Dalton Avenue, at approximately 1:10 p.m. on Friday, October 2, 2026. According to police,
Wingate was found unresponsive in the roadway and was pronounced dead at the scene by emergency medical personnel.
Preliminary information from CMPD indicates that Wingate had just exited a Charlotte Area Transit System bus at a designated stop when she walked back toward the bus as it was pulling away. The bus had begun moving as Wingate entered the street, at which point she was struck by the vehicle. The department described this account as preliminary and said the investigation is ongoing.
No other injuries or victims were reported in connection with the incident. CMPD’s Major Crash Investigation Unit has taken over the investigation, categorizing the case as a fatal crash involving a transit bus and pedestrian. As of October 3, police had not announced whether any charges or citations had been issued to the bus driver.
The crash location is situated north of Uptown Charlotte, a detail confirmed by local news sources. The Charlotte Observer reported the incident on the morning of October 3, with an update later that afternoon. WBTV also covered the crash, citing CMPD as the source for the timeline and details.
Authorities have asked anyone with information or who may have witnessed the collision to contact Detective Justin Kupfer at 704-432-2169, extension 1. Anonymous tips can be submitted to Charlotte Crime Stoppers at 704-334-1600. No additional witness statements have been publicly released.
Further details about Wingate’s background or residence have not been provided by the police or in public reports. The investigation remains active as officials continue to gather information and determine the full circumstances surrounding the fatal collision.

"
      ["date_timestamp"]=>
      int(1791057011)
    }
    [9]=>
    array(11) {
      ["title"]=>
      string(77) "Two deputies shot while confronting man firing at neighbors in North Carolina"
      ["link"]=>
      string(81) "https://nocarolinachronicle.com/two-deputies-shot-confronting-man-north-carolina/"
      ["dc"]=>
      array(1) {
        ["creator"]=>
        string(10) "Bill Moran"
      }
      ["pubdate"]=>
      string(31) "Sat, 03 Oct 2026 19:49:36 +0000"
      ["category"]=>
      string(39) "NewsDeep RunLenoir CountyNorth Carolina"
      ["guid"]=>
      string(81) "https://nocarolinachronicle.com/two-deputies-shot-confronting-man-north-carolina/"
      ["description"]=>
      string(1058) "
Two Lenoir County deputies were shot while responding to a man firing at neighbors in Deep Run, NC; the suspect was later fatally shot."
      ["content"]=>
      array(1) {
        ["encoded"]=>
        string(4528) "
Two Lenoir County deputies were shot Friday morning while responding to reports of a man firing at neighbors in the 700 block of Jonestown Road near Deep Run, North Carolina. According to Sheriff Jackie Rogers, the suspect opened fire on officers upon their arrival and later barricaded himself inside a residence before being fatally shot by deputies.
The two deputies wounded in the shooting were identified as Capt. Jovanni Villagra and Sgt. Cole Davis, according to Lenoir County Sheriff Jackie Rogers. Villagra was shot in the stomach during the initial exchange of gunfire after the suspect, later identified as 58-year-old Cecil James Langston of Deep Run, emerged from the residence and fired at officers upon their arrival. Langston then retreated inside the home, where he fired again, striking Sgt. Davis before officers entered the residence.
Both deputies were transported to ECU Health Medical Center in Greenville, where officials reported their injuries were non-life-threatening and that they remained in stable condition.
The incident began when deputies responded to reports of a man firing at people passing by in the 700 block of Jonestown Road near Deep Run on Friday, October 2, 2026. The call came in between approximately 11 a.m. and 11:20 a.m., according to Sheriff Rogers. Upon arrival, Langston opened fire on the deputies, prompting them to return fire and leading to a standoff at the residence. Additional law enforcement personnel from local and state agencies arrived to assist, and armored vehicles were used to breach the front door of the home.
Authorities deployed a drone to locate Langston inside the residence, where he was found in a bedroom. After officers entered, Langston was pronounced dead at the scene, according to the Lenoir County Sheriff’s Office and local reports. Sheriff Rogers confirmed the suspect’s death and stated that there was no continuing threat to the public following the conclusion of the incident.
Following the shooting, the North Carolina State Bureau of Investigation (SBI) took over the investigation, classifying the case as an officer-involved shooting due to multiple law enforcement officers discharging their weapons. The deputies involved were placed on administrative duty during the inquiry, which focuses on the exchanges of gunfire between Langston and responding officers. As of October 3, 2026, the SBI reported that the deputies remained alive and in stable condition, while the suspect was deceased and the scene was secured.
Sheriff Rogers emphasized that the immediate concern was the safety and well-being of Capt. Villagra and Sgt. Davis. He also advised the public to rely on official law enforcement updates as the investigation continued. Authorities confirmed that no ongoing danger to residents remained after the incident was contained.
The shooting occurred in the Pink Hill area of southern Lenoir County, near Deep Run, a rural community where such incidents are rare. The use of armored vehicles and drone technology highlighted the coordinated response by multiple agencies to ensure the safety of officers and the public. The SBI’s ongoing investigation will determine the full circumstances surrounding the shooting and the actions of all parties involved.

"
      }
      ["summary"]=>
      string(1058) "
Two Lenoir County deputies were shot while responding to a man firing at neighbors in Deep Run, NC; the suspect was later fatally shot."
      ["atom_content"]=>
      string(4528) "
Two Lenoir County deputies were shot Friday morning while responding to reports of a man firing at neighbors in the 700 block of Jonestown Road near Deep Run, North Carolina. According to Sheriff Jackie Rogers, the suspect opened fire on officers upon their arrival and later barricaded himself inside a residence before being fatally shot by deputies.
The two deputies wounded in the shooting were identified as Capt. Jovanni Villagra and Sgt. Cole Davis, according to Lenoir County Sheriff Jackie Rogers. Villagra was shot in the stomach during the initial exchange of gunfire after the suspect, later identified as 58-year-old Cecil James Langston of Deep Run, emerged from the residence and fired at officers upon their arrival. Langston then retreated inside the home, where he fired again, striking Sgt. Davis before officers entered the residence.
Both deputies were transported to ECU Health Medical Center in Greenville, where officials reported their injuries were non-life-threatening and that they remained in stable condition.
The incident began when deputies responded to reports of a man firing at people passing by in the 700 block of Jonestown Road near Deep Run on Friday, October 2, 2026. The call came in between approximately 11 a.m. and 11:20 a.m., according to Sheriff Rogers. Upon arrival, Langston opened fire on the deputies, prompting them to return fire and leading to a standoff at the residence. Additional law enforcement personnel from local and state agencies arrived to assist, and armored vehicles were used to breach the front door of the home.
Authorities deployed a drone to locate Langston inside the residence, where he was found in a bedroom. After officers entered, Langston was pronounced dead at the scene, according to the Lenoir County Sheriff’s Office and local reports. Sheriff Rogers confirmed the suspect’s death and stated that there was no continuing threat to the public following the conclusion of the incident.
Following the shooting, the North Carolina State Bureau of Investigation (SBI) took over the investigation, classifying the case as an officer-involved shooting due to multiple law enforcement officers discharging their weapons. The deputies involved were placed on administrative duty during the inquiry, which focuses on the exchanges of gunfire between Langston and responding officers. As of October 3, 2026, the SBI reported that the deputies remained alive and in stable condition, while the suspect was deceased and the scene was secured.
Sheriff Rogers emphasized that the immediate concern was the safety and well-being of Capt. Villagra and Sgt. Davis. He also advised the public to rely on official law enforcement updates as the investigation continued. Authorities confirmed that no ongoing danger to residents remained after the incident was contained.
The shooting occurred in the Pink Hill area of southern Lenoir County, near Deep Run, a rural community where such incidents are rare. The use of armored vehicles and drone technology highlighted the coordinated response by multiple agencies to ensure the safety of officers and the public. The SBI’s ongoing investigation will determine the full circumstances surrounding the shooting and the actions of all parties involved.

"
      ["date_timestamp"]=>
      int(1791056976)
    }
  }
  ["channel"]=>
  array(8) {
    ["title"]=>
    string(24) "North Carolina Chronicle"
    ["link"]=>
    string(31) "https://nocarolinachronicle.com"
    ["description"]=>
    string(13) "Tarheel Times"
    ["lastbuilddate"]=>
    string(31) "Fri, 23 Oct 2026 21:44:48 +0000"
    ["language"]=>
    string(5) "en-US"
    ["sy"]=>
    array(2) {
      ["updateperiod"]=>
      string(9) "
	hourly	"
      ["updatefrequency"]=>
      string(4) "
	1	"
    }
    ["generator"]=>
    string(30) "https://wordpress.org/?v=7.1.3"
    ["tagline"]=>
    string(13) "Tarheel Times"
  }
  ["textinput"]=>
  array(0) {
  }
  ["image"]=>
  array(0) {
  }
  ["feed_type"]=>
  string(3) "RSS"
  ["feed_version"]=>
  string(3) "2.0"
  ["encoding"]=>
  string(5) "UTF-8"
  ["_source_encoding"]=>
  string(0) ""
  ["ERROR"]=>
  string(0) ""
  ["WARNING"]=>
  string(0) ""
  ["_CONTENT_CONSTRUCTS"]=>
  array(6) {
    [0]=>
    string(7) "content"
    [1]=>
    string(7) "summary"
    [2]=>
    string(4) "info"
    [3]=>
    string(5) "title"
    [4]=>
    string(7) "tagline"
    [5]=>
    string(9) "copyright"
  }
  ["_KNOWN_ENCODINGS"]=>
  array(3) {
    [0]=>
    string(5) "UTF-8"
    [1]=>
    string(8) "US-ASCII"
    [2]=>
    string(10) "ISO-8859-1"
  }
  ["stack"]=>
  array(0) {
  }
  ["inchannel"]=>
  bool(false)
  ["initem"]=>
  bool(false)
  ["incontent"]=>
  bool(false)
  ["intextinput"]=>
  bool(false)
  ["inimage"]=>
  bool(false)
  ["current_namespace"]=>
  bool(false)
  ["last_modified"]=>
  string(31) "Sat, 24 Oct 2026 17:00:14 GMT
"
}