Adding interesting but irrelevant extras to a lesson (seductive details: off-topic stories, facts, pictures or other media) lowers how well learners can apply what they learned to new problems (transfer), compared with the same lesson without them in the same instruction time, on average in controlled experiments, but the harm is not consistent: in a randomised trial inside a large university genetics course, adding short off-topic video clips to the pre-class lecture videos, with students told the clips were only for interest and would not be tested, did not lower scores on problem-solving quizzes or exams (Maloy et al. 2019), and a laboratory experiment that matched the extras to the control text for length and reading level found no consistent harm to transfer (Tislar & Steelman 2021). The harm also depends on how the extras are presented: in a randomised online experiment, extras woven into a learning text lowered transfer, but the same extras set apart in grey boxes and announced beforehand as a separate topic did not lower it compared with the text without them (Wesenberg et al. 2026).
Status
weakened
Last change
2026-10-27
Serves
Q1+Q2 (How do humans actually learn, and what helps them learn best? / How does that work in practice: in school, with screens, and with AI as one of the tools?)
Ages
overgripande (reviews and meta-analyses across ages); vuxen-livslangt for the two studies that narrowed it on 2026-10-08 (university students: Maloy et al. 2019; Tislar & Steelman 2021); German-speaking students recruited online, mostly university students and about a quarter in secondary school (mean age 25), for the study that narrowed it on 2026-10-13 (Wesenberg et al. 2026)
Area
Screens, devices and attention
Relates to
C8 relates-to C6 (both concern extraneous load: C8 from material inside the lesson, C6 from devices in the room; Sweller et al. 2019 count environmental distraction as extraneous load. Thematic, not a dependency).
Bounds
Scope: interesting but irrelevant additions to instruction, compared with the same instruction without them in the same time, measured on transfer. Evidence in the project: Rey 2012 (review and meta-analysis: seductive details lowered retention and, more strongly, transfer, d ≈ 0.30 and d ≈ 0.48 per source note; moderators such as time pressure, type of detail and domain, so the effect is robust but not uniform in size); Noetel et al. 2022 (overview of reviews of multimedia design: coherence, removing seductive details, is among the principles with robust support; good design matters more for complex material and teacher-paced settings such as lectures than for self-paced ones); Mayer & Moreno 2003 (removing extraneous material as one of nine ways to reduce cognitive overload); Sweller et al. 2019 (cognitive load theory: extraneous load from poor instructional design). Consistency bound (added 2026-10-08, nightly attack): the average harm comes from controlled experiments and is not consistent across studies. Maloy, Fries, Laski & Ramirez 2019 (randomised, N = 754, two sections of an undergraduate genetics course at UCLA, flipped classroom; full text read): two clips of about 1 minute each per class session for the first 4 weeks (about 17 minutes in total, on top of 7 h 41 min of lecture video), introduced as "Genetics Applied Around Us" segments that were only for interest and would not be tested; pre-class quiz scores (92 % apply-level problems) 1.67 vs 1.68 out of 2, midterm 74.66 vs 75.76 (p = .389), final exam 164.38 vs 165.22 (p = .716). Quizzes were online with retakes allowed, so answer sharing cannot be ruled out; the clips were flagged as irrelevant, which in one lab study removed the harm (Eitel, Bender & Renkl 2019, as reported by Tislar & Steelman 2021). Tislar & Steelman 2021 (lab, 132 undergraduates, 110 analysed, plate-tectonics lesson; extras matched for word count and reading level; full text read): no main effect of detail type on transfer (F(2,103) = 2.11, p = .13), but a detail type × objectives interaction (F(2,103) = 3.39, p = .04). Within the group given learning objectives the overall test of detail type was not significant (F(2,52) = 2.80, p = .07); the harm there rests on a follow-up comparison (no details vs high-interest details 4.11 vs 3.36, t(34) = −2.15, p = .04, d = 0.72, reported as Bonferroni-adjusted). Without objectives there was no harm, and low-interest details even scored higher than no details (4.39 vs 3.7, p = .04). Study time was self-paced (extra time in proportion to the extra words). Some printed statistics do not add up: the same F(2,50) = 2.80, p = .07 is printed for the no-objectives transfer test, the core-content interaction is printed with one numerator degree of freedom, and the cell sizes in two tables add to 110 and 112 (corrected 2026-10-23). Counts reported by Rey 2012 (as cited by Tislar & Steelman): of 39 experiments 11 supported the effect, 13 were mixed and 15 did not. Meta-analytic averages are still negative (Rey 2012; Sundararajan & Adesope 2020, g ≈ −0.33 over 68 effect sizes, as recorded in the project's Rey 2012 source folder; Noetel et al. 2022 report g = 0.33 for removing such details), so the average claim stands for controlled experiments. Kill-condition test survived (2026-10-10, nightly attack): the newest and largest meta-analysis, Cheng, Wu, Wang & Wang 2026 (Educational Psychology Review 38, article 28; three-level model correcting for dependent effect sizes; 50 studies, 177 effect sizes, 20,683 participants; full text read), finds a smaller average harm than earlier meta-analyses but still a significant one: overall g = −0.16 (95 % CI −0.22 to −0.10), transfer g = −0.12 (95 % CI −0.22 to −0.03; 25 studies). Subgroups, on all outcomes combined rather than transfer alone: no significant harm in paper-and-pencil materials (g = 0.00, CI −0.16 to 0.17; 9 studies) against −0.19 in technology-based ones (moderator p = .033); secondary-school pupils −0.05 (CI −0.19 to 0.10; 9 studies) against college −0.18, though grade level was not a significant moderator (p = .11). The evidence is mostly from university students and on-screen materials, and the average harm is small. Second kill-condition test survived (2026-10-23, nightly attack): Wesenberg, Krieglstein, Schneider & Rey 2025 (Applied Cognitive Psychology 39(3), e70065; open access, full text read): 127 German-speaking students recruited online via Prolific (162 took part, 35 excluded, so below the planned 162; mean age 24.4; about a third at secondary school), the same self-paced coral-reef text as Wesenberg et al. 2026, with five unmarked interesting but irrelevant paragraphs either grouped in one block (position counterbalanced) or interspersed, against no extras; tests right after reading. Transfer (adjusted means, maximum 6): no extras 2.85, grouped 2.02, interspersed 2.18; extras vs none β = −0.45, p = .012; grouped vs interspersed p = .656, Bayesian BF01 = 4.34 for no difference. Retention 6.73, 5.01 and 5.91 of 18 (extras vs none p = .061). Learners with extras spent longer on the page (224 and 230 s vs 182 s, p = .042), so the transfer harm appeared even with more study time. Grouping the extras in one unmarked block did not remove the harm, so the presentation bound below rests on marking and announcing the extras as a separate topic, not on placing them together. Limits: one online experiment from the same lab and with the same text as Wesenberg et al. 2026, immediate tests only, below the planned sample. Presentation bound (added 2026-10-13, nightly attack): Wesenberg, Krieglstein, Jansen, Schneider & Rey 2026 (British Journal of Educational Psychology 96, 382–400; published online 21 August 2025; open access, full text read): 195 German-speaking students recruited online via Prolific (mean age 24.95; 46 % undergraduates, 24 % at secondary school), a self-paced 461-word text on how coral reefs form, test right after reading. Four randomised groups: no extras (n = 45); five interesting but irrelevant paragraphs (37–56 words each) on the uses of corals woven into the text (49); the same paragraphs in grey boxes, announced beforehand as a separate topic (topic signals, 49); and the boxes announced as a separate topic and as not relevant for the test (52). Transfer (six open questions, maximum 6; covariate-adjusted means): 2.90, 1.82, 2.96, 2.40. Unmarked extras lowered transfer against no extras (p < .001, d = .74); boxes announced as a separate topic raised it against unmarked extras (p < .001, d = −.79), and a Bayesian test favoured no difference between that group and the text without extras (BF01 = 4.57, moderate evidence). Also saying the boxes were irrelevant did not help further (2.96 vs 2.40, p = .055, d = .39). Retention did not differ between groups (p = .30). The extras raised interest, and interest partly carried better retention and transfer. Limits: one online experiment, mostly adults, one short text, immediate test, self-paced reading; the authors call it a first finding that needs replication. As reported by Wesenberg et al., marking which material was relevant or irrelevant removed the harm in Eitel, Bender & Renkl 2019 and Bender, Renkl & Eitel 2021 (Journal of Computer Assisted Learning; neither read). Kienitz, Krebs & Eitel 2023 (Instructional Science 51; online, N = 248, mostly undergraduates, fixed 6.5-minute study time, tests right after; full text read) does not fit that summary for transfer (corrected 2026-10-19): the extras lowered recall (F(1,112) = 16.430, p < .001) but not transfer (p = .142; transfer test ω = 0.56), so there was no transfer harm to remove; telling learners the extras were irrelevant removed the recall harm (focal contrast F(1,177) = 11.580, p = .001), and a prompt to process the extras separately did not significantly improve recall (p = .106) or transfer (p = .283). Third kill-condition test survived (2026-10-27, extra2): Bender, Renkl, Scheiter, Richter & Eitel 2026 (Journal of Computer Assisted Learning 42(2), e70205, DOI 10.1002/jcal.70205; open access, full text read): 165 participants in a randomised online chemistry lesson; interesting but irrelevant extras as mouse-over hyperlinks were rated less relevant than integrated extras, but without an irrelevance instruction they still lowered transfer (orthogonal contrast in the hyperlink arms, uninformed vs informed + no-extras control, β = 0.15, p = .042, η²p = .06; transfer z-means hyperlink no-SD 0.19, uninformed −0.12, informed −0.03). With an irrelevance instruction, hyperlink transfer matched the no-extras control. So hyperlink separation alone is not enough: the presentation bound still needs the extras marked or announced as irrelevant/separate. Limits: online, mostly adults, one topic, self-paced; overall transfer main effect only marginal (p = .059); integrated arms did not clearly replicate the classic seductive-details harm. Felt-learning caveat (added 2026-10-27 from candidate K52; not a claim): in Maloy et al. 2019, among students with high self-reported prior knowledge the off-topic clips raised perceived learning with no quiz or exam difference, so felt learning can move without measured learning.
What would kill this
A replication or meta-analysis in which seductive details or extraneous multimedia do not lower transfer (they match or raise it) in the same instruction time. Widened on promotion from the bank's “raise” so that a reliable null also counts against the claim. Bank wording (Swedish, verbatim): “Replikering där seductive details/extraneous multimedia höjer transfer under samma instruktionstid.”
Tutor consequence
none yet
Attack record
5 attacks: 3 survived, 2 narrowed
Note from 27 October 2026 (not an attack, so it is not counted). Added a caution to the limits (from a candidate belief in the library, not promoted to a claim): in Maloy et al. 2019 a subgroup rated learning higher with no quiz or exam change. Claim sentence, Status and kill-condition unchanged.
2026-10-08 narrowed
Attack
In a randomised trial inside a real university course (N = 754), adding short interesting-but-irrelevant video clips to the pre-class lecture videos did not lower scores on problem-solving (apply-level) quizzes, the midterm or the final exam, and a laboratory experiment that removed known confounds (extras matched for length and reading level) found no consistent harm to transfer, so C8's unconditioned claim that such extras lower transfer is not consistent beyond the average of controlled experiments.
Target
kill-condition
Source
Maloy et al. 2019 | DOI 10.1187/cbe.19-01-0004; Tislar & Steelman 2021 | DOI 10.1002/brb3.2322
Steelmans to
C8 dies on "a replication or meta-analysis in which seductive details do not lower transfer (they match or raise it) in the same instruction time"; Maloy et al. is the largest randomised test in an authentic course, with core content held identical and apply-level quizzes plus midterm and final exams, and finds no difference; Tislar & Steelman, built to fix the confounds critics raised (longer texts, unmatched reading level, no learning objectives), find no main effect on transfer; with Rey 2012 counting only 11 of 39 experiments as clearly supporting the effect, the claim as stated (extras lower transfer) overstates what is consistently found.
Why
Nightly-attacks 2026-10-08 Europe/Stockholm. Both full texts read (open access: CBE—Life Sciences Education 18(3), ar42; Brain and Behavior 11, e2322; not stored). Maloy et al. 2019: two large genetics course sections randomised between subjects, 389 control / 365 clips; about two ~1-minute clips per session for 4 weeks (17 min 13 s in total; control 7 h 41 min, clips group 7 h 58 min of video); pre-class quiz mean 1.67 vs 1.68 of 2 (92 % of items apply-level); midterm 74.66 vs 75.76 (p = .389); final exam 164.38 vs 165.22 (p = .716); course percentage 84.75 vs 85.05 (p = .730); no difference by prior knowledge in scores. Tislar & Steelman 2021: 3 × 2 between subjects (no / low-interest / high-interest details × learning objectives or not), 110 analysed; transfer: no main effect of details (F(2,103) = 2.11, p = .13), interaction with objectives (p = .04); with objectives, no details 4.11 vs high-interest 3.36 of 5 (p = .04, d = 0.72); without objectives no difference (p = .21). Kill withheld: (1) meta-analytic averages remain negative (Rey 2012, transfer d ≈ 0.48; Sundararajan & Adesope 2020, g ≈ −0.33 over 68 effect sizes, as recorded in the project's Rey 2012 source folder; Noetel et al. 2022 report g = 0.33 for removing such details); (2) in Maloy the clips were flagged as not for the test and added about 17 minutes (~4 %) of video, so the same-time condition held for the core content only, and quizzes were online with retakes; (3) Tislar & Steelman did find a harm when objectives were given; (4) both used university students. Not a restatement: C8's Bounds named moderators of the size of the effect (Rey 2012) but not a classroom null or a confound-controlled null; first ledger entry on C8. Claim and Bounds rewritten same turn; brief and the practical rule rewritten.
Rewrite
claim-field+brief
2026-10-10 survived
Attack
The newest and largest meta-analysis of seductive details, which corrects for the dependence between effect sizes that inflated earlier meta-analyses, finds the harm much smaller than before and no significant harm at all in paper-and-pencil materials or among secondary-school pupils, so a meta-analysis no longer shows that interesting but irrelevant extras lower transfer and C8's kill-condition is met.
Target
kill-condition
Source
Cheng, Wu, Wang & Wang 2026 | DOI 10.1007/s10648-025-10099-z
Steelmans to
C8 dies on "a replication or meta-analysis in which seductive details do not lower transfer (they match or raise it) in the same instruction time"; this three-level meta-analysis (50 studies, 177 effect sizes, 20,683 participants, published to May 2025) nearly doubles the evidence since Sundararajan & Adesope 2020 and halves the overall effect (g = −0.16 against −0.33 and Rey 2012's transfer d ≈ 0.48); if transfer alone came out near zero, the average claim would fall.
Why
Nightly-attacks 2026-10-10 Europe/Stockholm. Full text read (published version, Educational Psychology Review 38, article 28, published 5 March 2026; author-shared copy on ResearchGate; publisher page closed; not stored). Table 3: overall g = −0.16 (SE 0.03, 95 % CI −0.22 to −0.10, p < .001), I² = 57.8 %. Table 4: transfer g = −0.12 (SE 0.05, 95 % CI −0.22 to −0.03, p < .05; 25 studies, 68 effect sizes); recall −0.17; comprehension −0.19. Kill fails: transfer is still lowered, significantly, on average. Context, on all outcomes combined, not transfer alone (Table 5): paper-and-pencil settings g = 0.00 (CI −0.16 to 0.17; 9 studies) versus technology-based −0.19 (41 studies), moderator F = 3.74, p = .033; secondary-school pupils g = −0.05 (CI −0.19 to 0.10; 9 studies) versus college −0.18, but grade level was not a significant moderator (F = 2.02, p = .11); long interventions (74 minutes or more) −0.12, not significant. These subgroup results are between-study comparisons on combined outcomes and do not show that transfer is unharmed in paper materials or for school pupils; not applied as a narrowing (open-questions list full). Publication-bias checks clean (Egger p = .98; trim-and-fill imputed none). Not a duplicate: Rey 2012 and Sundararajan & Adesope 2020 were the earlier averages; Maloy et al. 2019 and Tislar & Steelman 2021 (2026-10-08) were single-study nulls behind the consistency bound. Claim field unchanged; Bounds record the smaller size and the subgroups. Status remains weakened.
Rewrite
none
2026-10-13 narrowed
Attack
Within one randomised experiment, the same interesting but irrelevant paragraphs that lowered transfer when woven into a learning text (d = .74) did not lower it when they were set apart in grey boxes and announced beforehand as a separate topic (transfer equal to the text without extras; BF01 = 4.57), so the harm C8 asserts depends on how the extras are presented, not only on whether they are there.
Target
claim-text
Source
Wesenberg et al. 2026 | DOI 10.1111/bjep.70022
Steelmans to
C8 says extras lower transfer on average in controlled experiments, with nulls only in a real course with clips flagged as not for the test (Maloy et al. 2019) and a confound-controlled lab study (Tislar & Steelman 2021); Wesenberg et al. reproduce the harm inside the same experiment (unmarked extras vs none, d = .74) and then remove it completely by a light change of presentation that does not tell learners to ignore the extras, with a powered sample (N = 195; a priori N = 179) and a Bayesian test for equivalence, so the unconditioned "extras lower transfer" is too broad: it holds for extras woven into the material, and not shown where they are visibly set apart as a separate topic.
Why
Nightly-attacks 2026-10-13 Europe/Stockholm. Full text read (open access, PMC12879520; British Journal of Educational Psychology 96, 382–400, published online 21 August 2025; not stored). Table 3, covariate-adjusted transfer means (max 6): no extras 2.90 (SE .22), unmarked extras 1.82 (.21), topic signals 2.96 (.21), topic and irrelevance signals 2.40 (.20); ANCOVA F(3, 189) = 6.42, p < .001; planned contrasts: unmarked vs none p < .001, d = .74; topic signals vs unmarked p < .001, d = −.79; topic-and-irrelevance vs topic signals p = .055, d = .39; Bayesian ANCOVA no extras vs topic signals BF01 = 4.57. Retention F(3, 189) = 1.23, p = .301. Interest higher with extras (p = .009), and interest mediated small gains in retention and transfer. Kill withheld: (1) the kill-condition asks for a replication or meta-analysis in which extras do not lower transfer; here unmarked extras did lower it, which supports the average; (2) one online experiment with mostly adult students, a single short text and an immediate test; the authors call it a first finding needing replication; (3) a separation prompt did not reduce the harm in Kienitz, Krebs & Eitel 2023 (as reported by Wesenberg et al.). Not a restatement: the Claim field already names Maloy et al.'s clips flagged as not for the test, and Bounds mention that flagging extras as irrelevant removed the harm in one lab study (Eitel et al. 2019, second-hand); this result is a within-study test in which marking the extras as a separate topic, without calling them irrelevant, removed the transfer harm, while calling them irrelevant added nothing. Not a duplicate: first ledger entry from this study. Claim, Ages and Bounds rewritten same turn; brief and the practical rule rewritten.
Rewrite
claim-field+brief
2026-10-23 survived
Attack
If gathering interesting but irrelevant paragraphs into one block, instead of interspersing them, is enough to stop them lowering transfer, or if the extras do not lower transfer in a new randomised replication at all, C8's kill-condition (a replication in which seductive details do not lower transfer) is met for unmarked extras set apart by position alone.
Target
kill-condition
Source
Wesenberg, Krieglstein, Schneider & Rey 2025 | Applied Cognitive Psychology 39(3), e70065, DOI 10.1002/acp.70065
Steelmans to
C8's presentation bound shows that setting extras apart in grey boxes announced as a separate topic removed the harm (Wesenberg et al. 2026); Harp & Mayer 1998 found placement mattered, and disruption accounts predict that one interruption costs less than five; a randomised experiment with a no-extras control, a grouped arm and an interspersed arm, using the same text as the study behind the presentation bound, is a fair new test of whether the harm survives when the extras are simply moved out of the flow of the text.
Why
Nightly-attacks 2026-10-23 Europe/Stockholm (widened-search slot for C8, re-checked tonight). DOI checked on Crossref (title, four authors, Applied Cognitive Psychology 39(3), e70065, May 2025). Full text read (open access, CC BY 4.0, stored privately as ). 162 took part; 26 excluded for notes or aids, 2 extreme time outliers, 7 failed attention checks; N = 127 (no extras 46, grouped 43, interspersed 38), below the planned 162. Adjusted transfer means (max 6): 2.85 (SE 0.23), 2.02 (0.24), 2.18 (0.26); extras vs none β = −0.45, t = −2.60, p = 0.012; grouped vs interspersed β = 0.09, t = 0.45, p = 0.656; Bayesian ANCOVA BF01 = 4.34. Retention 6.73, 5.01, 5.91 of 18; extras vs none β = −0.34, p = 0.061. Time on page 224.07 and 229.55 s with extras vs 182.48 s without (p = 0.042). Kill fails: unmarked extras lowered transfer in a new randomised sample, whether grouped or interspersed, even with more time on the page. Caveats: same lab and text as Wesenberg et al. 2026; online, self-paced, immediate tests; transfer α = 0.69; below planned N; not reported as preregistered. Not a duplicate (prior C8 survival: Cheng et al. 2026 meta-analysis; Wesenberg et al. 2026 is a different sample and design). Not a restatement: the presentation bound says nothing about extras set apart by position alone. Claim field unchanged; Bounds add the survival; brief C8 paragraph and the practical rule Limits record it. Status unchanged (weakened). Not a preprint.
Rewrite
none
2026-10-27 survived
Attack
If presenting interesting but irrelevant extras as mouse-over hyperlinks — visibly separated and rated less relevant than integrated extras — were enough that transfer matched a no-extras control without an irrelevance instruction, C8's kill-condition (a replication in which seductive details do not lower transfer) would be met for that format. It was not: without the irrelevance instruction, hyperlink extras still lowered transfer.
Target
kill-condition
Source
Bender, Renkl, Scheiter, Richter & Eitel 2026 | Journal of Computer Assisted Learning 42(2), e70205, DOI 10.1002/jcal.70205
Steelmans to
C8 dies on a replication in which seductive details do not lower transfer in the same instruction time; authors predicted that hyperlink format alone would cue irrelevance and remove the harm (informed-use / presentation bound); a randomised 3 × 2 online experiment with a no-extras control is a fair new test of that kill path.
Why
Extra attack pass 2026-10-27 Europe/Stockholm (week-2 push; FreiDok OA full text). DOI checked on Crossref (title, five authors, JCAL 42(2), e70205, 2026). Full text read (CC BY; stored privately). N = 165 analysed; chemistry atomic-models unit; retention and transfer (11 items, α = .75); spatial ability and prior chemistry experience as covariates. Multivariate main/interaction effects ns (all p > .092). Univariate transfer main effect of SD condition F(2,157) = 2.89, p = .059, η²p = .04. In the hyperlink arms, the orthogonal contrast (uninformed vs informed + control) predicted transfer, β = 0.15, p = .042, η²p = .06 — uninformed hyperlink worse than informed hyperlink and no-extras control (who did not differ; residual p = .095). Table 2 transfer z-means: hyperlink no-SD 0.19, uninformed −0.12, informed −0.03. Integrated arms did not replicate the classic informed-use pattern (transfer contrast β = 0.01, p = .890). Kill fails: unmarked hyperlink extras still lowered transfer; an explicit irrelevance instruction was still needed. Caveats: online, mostly university students, self-paced, one topic; overall MANOVA transfer only marginal; integrated SD harm itself weak in this sample. Not a duplicate (prior C8 survivals: Cheng et al. 2026 meta; Wesenberg et al. 2025 grouped). Not a preprint. Claim field unchanged; Bounds and the practical rule Limits record the survival. Status unchanged (weakened). C8 now at 3 of 3 survivals → floor; the promotion rule promotion of K46 as C19 in the same turn.
Rewrite
none
Corrections to our notes
These fix a number or a fact in our notes. They are not attacks and do not count toward the floor of 3 survivals.
— C8's limits relay Wesenberg et al.'s summary that in Kienitz, Krebs & Eitel 2023 marking the extras' relevance removed the harm and a separation prompt did not reduce it, but in that study the extras did not significantly lower transfer at all (p = .142), the irrelevance prompt removed a recall harm, and the separation prompt's effect on recall was not significant (p = .106).
— C8's limits say Tislar & Steelman 2021 found a transfer harm when learning objectives were given (p = .04), but within that group the overall test of detail type was not significant (F(2,52) = 2.80, p = .07): the p = .04 is a follow-up comparison, and several of the paper's printed statistics do not add up.
Three-level meta-analysis (50 studies, 177 effect sizes): seductive details still lower transfer on average (g = −0.12, 95 % CI −0.22 to −0.03), smaller than earlier meta-analyses; no significant harm in paper-and-pencil materials on combined outcomes. C8 kill-condition test survived 2026-10-10.
Randomised online experiment (N = 127): unmarked extras lowered transfer (2.02 and 2.18 vs 2.85 of 6; p = .012), and grouping them in one block did no better than interspersing them (BF01 = 4.34). Kill-condition test survived 2026-10-23.
Lab experiment, 113 university students, fixed 6.5-minute study time: extras did not significantly lower transfer overall (p = .160; p = .055, d = −0.37, without the spatial-ability covariate); they did for the half with weaker school-leaving grades (d = −0.60), not significantly for the better half (d = −0.21). Cited in open question 12.
School-age (UK high school) randomised video experiment, N = 101: seductive details hurt retention for one physics topic but not another. Retention, not transfer; abstract only.
Online experiment (N = 248, mostly undergraduates): seductive details lowered recall (p < .001), removable by telling learners they are irrelevant; no significant harm to transfer (p = .14, low-reliability test). Another single-study transfer null: consistent with C8's existing consistency bound, not a new attack.
Randomised trial in a real university course (N = 754): short off-topic video clips flagged as for interest only did not change apply-level quiz, midterm or final-exam scores. C8 weakened 2026-10-08 (consistency bound).
Lab experiment, 100 university students: seductive text passages lowered transfer under a 70 bpm false heartbeat but not under 90 bpm (interaction p < .001). One study, 25 per cell; held as open question 15 (2026-10-23), not a narrowing.
Confound-controlled lab study (extras matched for length and reading level): no main effect of seductive details on transfer; a harm only in a follow-up comparison when learning objectives were given, after a non-significant overall test (p = .07). C8 weakened 2026-10-08 (consistency bound); description corrected 2026-10-23.
Two randomised experiments: more seductive text lowered recall linearly, but transfer was not significantly affected (low reliability in Study 1). Restates C8's consistency bound; not emitted as an attack (nightly 2026-10-15).
Randomised online experiment (N = 195): unmarked extras woven into a text lowered transfer (d = .74), but the same extras in grey boxes announced as a separate topic did not (equal to no extras, BF01 = 4.57); calling them irrelevant added nothing. C8 narrowed 2026-10-13 (presentation bound).
Review state: first-pass draft, pending human review.