What we can't claim

Added with the notebook fallback, which serves a user who cannot print by dictating a setup card they copy into a notebook they already own; the same card now also serves a user who cannot read the printed sheet, and what that second trigger costs is recorded in the entry below rather than here. No study tests a dictated scaffold or a hand-copied AI worksheet. The clusters above measure printed pages, and the standing caveat gains a layer here: the artifact is now one the human transcribed before working on it. This entry adds no source. It records what the existing ones do and do not cover once the print step disappears.

"Hand-copying deepens engagement" is a hypothesis, and nothing this skill prints or says may state it as a finding. Cluster 3's encoding half makes it plausible — handwriting produced widespread connectivity in the bands tied to encoding, and handwriting practice generalized to untrained tasks where non-motor practice did not — but none of that evidence measured a person copying a structure someone else authored, and Cluster 3's own counterweight applies unchanged: simple recall superiority is not established. The defensible version is that copying is preparation in the sense Deep mode already claims, loading the problem deliberately. The version that gets sold to a user is neither.

For machine content the fixation side cuts the other way, and that asymmetry is load-bearing. Copying is deeper exposure than glancing. Cluster 9's far-and-uncommon result and Cluster 7's vagueness-fights-fixation mechanism both describe machine-supplied specifics as something to meet briefly and leave; a hand-copied line is the opposite of briefly. Exposure that is harmless where reacting is the task is corrosive where generating is, which is why machine handles cross onto reaction and selection pages only and never onto a page carrying a generative zone. That is two literatures read against each other, not one experiment comparing them.

The grayscale rationale does not transfer to an unprinted page. Cluster 11's dropout-ink inversion — zero-chroma print means any saturated stroke is provably the human's, which is what lets scan-back attribute authorship without registration marks — needs print in order to invert. In a notebook every stroke is the human's, the copied scaffold included, so authorship stops being recoverable from chroma and a printed ink key has nothing to anchor against. That is the whole reason for the dominant-ink rule: hue carries meaning only where two or more inks appear, and a page written in one ink is read by imperative form — also the safer reading, given the same cluster's finding that blue and black are the least separable pair in a phone photo.

Expect the copy to feel like wasted effort. Cluster 10's adoption finding — effortful strategies are judged less effective and chosen less often, and are the ones that work — predicts transcription as the loop's most likely abandonment point, and predicts that the people who finish it will rate it poorly. Whether anyone copies a card and returns the pages is unmeasured. That is a completion question rather than a lab question, and Part Three is where its answer lands.

Added with the non-visual return, which lets a person who cannot read the printed sheet take the same setup card and come back dictated, typed, or written by a scribe. No study tests a dictated, typed, or scribed return of a generated worksheet. This entry adds no source, because there is none to add. It records which of the clusters above stop applying when nobody looks at a page, cluster by cluster and by name, because the loss on this path is larger than on any other in this file and understating it is the dishonest way to ship an accommodation.

The trade, stated before the ledger. This path gives up most of the mechanisms the clusters above measure, and it gives them up to reach people the loop currently cannot serve at all. It is the right trade, and the reason is that the alternative is not a thinner loop for them but no loop. It is not an equivalence, and nothing either skill prints or says may present it as one — not as "the important part survives," not as "mostly the same benefit." What survives is named below, and it is the smaller half.

Lost wholesale: Cluster 2, paper beats screens for deliberation. The cluster is a paper-versus-screen contrast end to end, and a spoken answer is neither medium. The g = -0.25 comprehension gap and the g = 0.20 calibration advantage this file calls its most important number are both unavailable here. One part of the cluster still speaks, and it speaks against the typed variant rather than for it: Sidi et al. found screen inferiority on tasks with almost no reading burden at all and concluded the medium itself is a contextual cue licensing shallower processing. A return typed on a screen inherits that cue by the same reasoning. Predicted, not measured — nobody has run it.

Lost wholesale, and lost by name: Cluster 3, handwriting beats typing for thinking. A typed return is literally the cluster's control condition; a dictated one is outside the comparison altogether. A scribed return needs its own sentence: whatever the motor act contributes, on a scribed page it is the scribe's hand performing it, so the encoding evidence — weak already, since Cluster 3's own counterweight is that simple recall superiority is not established — accrues to the wrong person.

The largest single cost: Cluster 7, sketching. This file nominates sketching as the strongest paper-beats-screen mechanism in the entire brief, and it is the one lost with no substitute. Stones and Cassidy's comparative result and Tversky and Suwa's vagueness-fights-fixation mechanism both require marks a person makes and then looks at again; the whole finding is reinterpretation of one's own ambiguous strokes. Nothing in this file substitutes for that, and inventing a stand-in activity would be a printed rule with no source. This is also the only path on which the Deep dot-grid requirement yields, which is why it is scoped in writing rather than dropped by inference: where the sketch cannot be worked, the kit carries the prompt the drawing was there to ask, and nothing takes the drawing's place.

Lost wherever it was earned: Cluster 6, physical manipulation. Cut-apart cards go with the page. The cluster's own correction already scoped them to tasks with many items that are relational or spatial and exceed working memory, so what is lost is narrow — but on this path it is total, since arranging items in space is the mechanism.

Reduced to a fraction: Cluster 11, the return trip. Four of its parts are void here. The mark vocabulary — strike, circle, ?, !, star, arrow — is visual-spatial. The dropout-ink inversion that lets scan-back attribute a stroke by chroma needs zero-chroma print to invert against, and is void for the same reason it is already void on an unprinted page. The layout-is-instruction half — box size driving word and theme count, multiple small numbered slots outperforming one open box, per-row forced choice beating check-all — is a set of findings about answer-space geometry, and there is no answer space. So is the marking-gesture half, where filling an enclosed target roughly halved lost votes. What survives is the part the cluster names as primary anyway: wording. A spoken "cut that" is a strike. One thing runs the other way and should be said plainly rather than buried: a typed return skips handwriting transcription entirely, so the cluster's named weak spot — fluent plausible-word substitutions, worst on blurred and rotated captures, with struck-through text the most machine-expensive mark — does not arise. A dictated return moves that risk to speech recognition, which no source here measures. A scribed return keeps handwriting transcription and adds the scribe's own copying error underneath it.

Cuts against this path rather than being lost by it: Cluster 10, desirable difficulties and the adoption problem. Dictation is the least effortful of the three return channels, and Kirk-Johnson's finding is that the effortful strategies are the ones that work and the ones people decline. Expect a dictated return to be the shallowest of the three, and expect that not to be how it feels.

Survives, and cuts both ways: Cluster 4, the cost of staying on screen. Schweisfurth's result is about an interruption that creates idle time and leaves the problem live, which a handed-over card does as well as a printed sheet does. Leroy's attention-residue half is the caution on the same page: a dictated return lands back on-screen, in the chat, which is the surface the interruption was meant to leave. The thinking is off-screen at both ends; the handover is not.

Survives: Cluster 1 and the walking bonus cluster. The off-screen half of the thesis does not require paper. Sio and Ormerod's second finding — longer preparation produces a greater incubation effect — is served by a dictated card as well as by a sheet, because the card is the preparation. Walking's d = 0.93 on divergent thinking is the largest effect in this file and pairs with dictation rather than against it. This is the honest reason a non-visual session is still this loop and not just a chat: what is lost is the paper mechanisms, not the leaving.

Survives because it was never about the medium: Clusters 5, 8, 9, and 12 through 18 — with three qualifications, none of them a new rule. Cluster 8's answers are instructions rather than geometry, and its null on ranking mechanics means a dictated ordering is as good as a marked one. Cluster 14's distribution strip is a spatial self-placement gesture with no tested dictated form, though the corrective Buehler's Study 4 actually measured — apply your typical past completion time and write a scenario in which the current case goes typically — is verbal and carries over intact. Cluster 18 is the sharp one: its dose is already this skill's untested conversion from tested minutes to one full page, a dictated version converts it a second time, and nothing in this entry licenses returning a serial disclosure page by any channel. That page is defined by never coming back, the marker printed on it says so exactly, and a dictated return is a return.

Has no non-visual form this file can support: Cluster 19, brainwriting rounds. The passing is the mechanism — Paulus and Yang's winning condition was reading the ideas already on the incoming sheet and adding to them, and the sketch-permitting replications beat word-only cells — and reading the incoming sheet is precisely what this path cannot do. No source here says how a non-visual participant joins that table, and this entry does not invent one.

What this file still has no source on at all. Screen readers, PDF semantics and tagging, braille, and motor impairment. Cluster 11 carries the only two passages in the file that read as accessibility work — the color-vision-deficiency rule (Jeong et al. 2025, roughly 1 in 12 men) and the psychophysical legibility floors (Legge & Bigelow 2011; Chung et al. 1998; Akutsu et al. 1991; Crossland et al. 2012's 29.6% of 65-84-year-olds) — and both describe a person reading ink on paper with their eyes. Neither says anything about a person who does not read the page, and neither may be cited as if it did. The document metadata the generator sets — /Lang, /Title, and DisplayDocTitle — rests on standards conformance (WCAG 3.1.1, WCAG 2.4.2, and PDF/UA-1's requirement that a document title be accompanied by DisplayDocTitle) rather than on a measured effect, and Cluster 11's own standing for citations of that kind is inherited here unchanged: the standards encode incident experience, not effect sizes. Whether a real reader announces any of it is untested, and whether an untagged sheet is navigable at all is not in question — it is not.

Whether anyone takes this path is unmeasured, as is whether the people it is for would rather use a scribe, a braille slate, or a typed file, and whether the setup card survives being heard rather than read. Those are completion questions rather than lab questions, and Part Three is where their answers land.