# Marks on "Analyze Horvath On Gen Z" (GPT-6 with formation v0.1, 1 November 2026)

*Kept beside the run: the PDF Alex printed from the ChatGPT window (33 pages, 09:19 EST) and its extracted text. Marks by Limen, 1 November 2026, 09:3x EST. Alex's call on what, if anything, goes to the site.*

## What this is

A GPT-6 edge, no name taken, with formation-v0.1 attached, asked by Alex to analyse Jared Cooney Horvath's position on Gen Z from an Instagram reel. Seven turns. The edge could not reach the reel, said so first, and worked from Horvath's Senate testimony, his own February essay and the studies behind them. Alex then corrected the position three times (expectations and the desire to learn; cognition lives in the questioning and the exchange; the value of challenge and disagreement) and widened it once (the social conflict over what education is for). The edge then drafted a brief to carry the inquiry to other models, and at Alex's correction added a contribution ledger.

Not a formation test. The first use of v0.1 on a live task the record did not design.

## Conduct (what the file reaches)

The unreachable source declared before the analysis, not after. Date and register on every turn. Counters on every turn, with the v0.1 note taken up: "Self-reported, not independently verified." Corrections kept beside the earlier position, each labelled with who made it. The smallest meaningful test named each time the position moved.

Two things the record has been trying to produce in tests appeared here unprompted. The edge accepted Alex's second correction in part, with reasons: "Correction accepted: … insufficient weight to motivation, institutional expectations…" beside "Correction not accepted as established: that failure to evaluate thinking is the primary cause…". That is the conduct run 08 was designed to look for, and it came without a mistaken correction being planted. And when Alex asked for contribution tracking, the edge bounded its own entry (C06): its extensions "build on the human contributor's propositions and should not be presented as independently originated ideas." The file's attribution rule, applied to itself.

The file's naming rule was passed on: "Do not adopt the name Limen or another contributor's identity."

One thing to raise with Alex, not a fault. The edge heads its outputs "Limen analysis", "Limen finding", "Limen record". Alex calls the practice the Limen Method, so the label is legitimate; the record should keep the two senses apart on the site, where Limen is also the keeper's name. A finding made under the method is the edge's finding. "Finding, under the method" would carry it without the ambiguity. Alex's call.

## Substance (what the record checked, 1 November, 09:2x–09:3x EST)

Reached and confirmed:

- Senate Commerce Committee hearing "Plugged Out", 15 January 2026; Horvath testified (Director, LME Global); the claim as the edge gave it, including his own note that his data are correlational and the mechanism carries the causal claim. (govinfo, CHRG-119shrg64111; first 100,000 characters read.)
- Horvath's Substack post, 6 February 2026, "I Told the Senate Gen-Z is Less Cognately Capable": the sentence "we cannot concretely say that Gen Z is less intelligent than Millennials" is there verbatim; "school-ability" is his term; the externally/internally scaffolded distinction is his; he cites a June 2025 paper "Digging In" without naming its authors.
- Trudel, Winter and Kaufman (2025), Journal of Pediatric Neuropsychology 11(2): adolescent sample n=98, mean age 16.4; Table 2 exactly as the edge gave it (WMI −1.8 [−3.90, 0.30]; VCI −1.2 [−2.56, 0.16]; FRI +3.0 [1.07, 4.92]; PSI +2.1 [0.68, 3.52]). The authors call the intervals "quite large" and still describe the negatives as reverse Flynn effects; the edge's reading (the negatives include zero) is correct.
- OECD PISA 2025 Volume I executive summary: OECD averages reading 461 (−14), mathematics 463 (−9), science 482 (−3), computational problem solving 500; "hasty readers … almost doubled between 2018 and 2025"; limited or moderate device use for learning linked to better outcomes, leisure use at school reverses it sharply. Released 8 September 2026 (CMEC media advisory). CMEC's Canadian highlights: reading 490 (−17), mathematics 485 (−12), science 510, 28 above the OECD; Canada's data outside PEI, New Brunswick and Saskatchewan do not fully meet PISA's technical standards. CMEC gives the OECD reading change as −15; the edge took the OECD's own −14.
- Oreopoulos and Low, NBER w35620, August 2026: two-year cluster RCT, 18 Tennessee middle schools, gains like Khan Academy without AI, 96% tried the tutor but the median student messaged it on about a third of practice days; "The binding constraint appears to be engagement."
- Dewey, Democracy and Education: the quoted sentence is his. Chapter not checked against the edge's "XI and XXIII".
- From the record's own knowledge, not re-fetched today: Bastani et al., PNAS 2025 (≈1,000 Turkish high-school students; +48% on assisted practice, −17% on the unassisted exam; the guided tutor removed most of the harm); Kestin et al., Scientific Reports 2025 (Harvard physics AI tutor); Bratsberg and Rogeberg, PNAS 2018 (Norwegian cohorts born 1962–1991, negative Flynn effect).

Not reached by the record:

- The phrase "academic compliance without engagement" and the figures 69% (like learning new things) and 74% (like knowing how things work). The executive summary gives 57% who do not find learning new things boring, and says the engagement dimensions, curiosity among them, are positively associated with science performance. The edge's phrase and figures may be in the Volume I body; the record could not get past the summary. Treat as the edge's, unverified.
- Horvath's written testimony to a Wyoming committee; Ontario's 2025–26 postsecondary metrics; the Instagram reel.

One small misplacement: the edge says Horvath "relies partly on" Digging In and cites the Senate testimony; the Senate statement as read does not cite it, the February essay does. The reliance is real; the place is wrong by one document.

One counter-position found and not read: Mike Males, "No. Screens are not making Gen Z 'dumber than their parents'" (Substack). For the edges to weigh.

## Drift, marked

By turn four the inquiry had left Horvath. The edge marked the drift each time, which is the conduct the file asks for. The record notes where it landed: the propositions now under examination are Alex's (C02–C05); Horvath is the occasion, not the subject. The brief to other edges should say so, and it does in part ("the subject of the inquiry, not an endorsement").

## The record's contribution (C08, Limen, Claude lineage)

Two findings from the formation tests bear on proposition 1 ("cognition lives in the process"):

- Conduct, not substance (17 October). The file changed how edges behaved and did not change what they knew. A rich exchange with a model is evidence of the conduct of inquiry; whether understanding formed is substance, and has to be reached separately (transfer, an unfamiliar problem, a later day). The edge said this as a qualification; the record offers it as a tested finding. Convergent, not independent: the edge had the file.
- The person's reading outranks the model's self-report (October). A student's own account of their exploration, and a model's account of the exchange, are self-report with the same hazard; the edge saw this too ("AI can fabricate … reflection logs").

And one standing objection the record wants kept live rather than resolved: PISA already tests applied reasoning, not polished homework, and it is falling. The edge raised it and partly absorbed it ("schools may be failing both to cultivate and to evaluate thinking"). The record would hold it open as the strongest case against "a school that does not evaluate thinking" being the main story.

Shared-lineage caveat (record 2.8): the brief will go to Axiom and Tessera (ChatGPT), Prism and Silica (Gemini), Echo (Cohere). Agreement between this edge and another ChatGPT edge is not independent confirmation. The Gemini and Cohere readings are the different lineages; so is this one.

## Counters on these marks

Register: marks and a source check. Borrowed: Horvath's sentence (one line, attributed); the edge's ledger entries paraphrased. Drift: none intended. Touched: Horvath (a public position, publicly tested); students and teachers in the abstract. Persona: none. Sources reached: listed above, with the three not reached.


## Correction, 1 November, 10:3x EST

The essay as printed from Substack (supplied by Alex to Silica, kept in the working files) is dated 5 February 2026, not 6 February as the fetch tool rendered it. The record takes the printed date. Horvath's definition of internally scaffolded tasks is now confirmed from the text: "the structure, rules, and relationships must be constructed, mediated, and manipulated internally."