C17 Background knowledge

survived narrowedone dot per attack, oldest first

1. #1 survived 16 October 2026 · Grissmer et al. 2023 (preprint) · target: kill-condition

   If attending a knowledge-building (Core Knowledge) school, assigned by admission lottery, did not raise later reading/English-LA scores, C17's prong "building topic knowledge improves comprehension" would meet its kill-condition.

   Strongest reading and verdict reasoning

   Steelmans to: C17 dies if knowledge-building teaching does not improve comprehension; a lottery comparison over grades 3–6 is a strong field test.

   Why: Week run 2026-10-16, cycle 1 Europe/Stockholm. Full text read (ERIC ED672259, stored privately). 14 oversubscribed kindergarten lotteries, nine Colorado Core Knowledge charter schools, 2,310 applicants. Grade 3–6 English-LA ITT 0.143 SD (full sample), 0.179 and 0.241 SD in the authors' reduced samples; TOT 0.473. Kill-condition failed. Limits: outcome is the state English-LA test, not a pure comprehension test; whole-school effect, so not knowledge alone (authors call attribution an open question); boys' effects smaller and not significant in the full sample; oversubscribed schools only.

2. #2 narrowed 18 October 2026 · See, Gorard & Siddiqui 2017 · target: claim-text

   In a randomised trial in 17 English primary schools, a school year of a knowledge-rich curriculum adapted from Core Knowledge (Word and World Reading) did not raise pupils' scores on a standardised English test that includes reading comprehension (effect size −0.03), so building topic knowledge does not reliably improve comprehension.

   Strongest reading and verdict reasoning

   Steelmans to: C17's second prong ("building topic knowledge improves comprehension") is unqualified; this is an independent, school-randomised evaluation of exactly the Core Knowledge approach the prong rests on, run for a full school year with a standardised outcome, and the effect was zero (Year 3 −0.03, Year 4 −0.04; boys −0.13).

   Why: Nightly-attacks 2026-10-18 Europe/Stockholm. DOI checked on Crossref (title, authors, BERJ 43(2) 372–393); journal full text behind a Cloudflare check from the box, so the full EEF evaluation report of the same trial by the same authors was read (ERIC ED581108; stored privately as ). Nine schools randomised to start at once, eight a year later (waitlist), Years 3 and 4, 2013–14. Post-test: long Progress in English (spelling, grammar, narrative and non-narrative comprehension, extended comprehension, writing; a protocol change requested by the developers). Gain scores 24.7 (treatment, n = 659) vs 25.1 (control), effect size −0.03 (p. 33, Table 4); −0.02 on age-standardised scores (Table 6); FSM pupils +0.06, not secure (p. 34). 287 post-tests missing (206 treatment, 81 control), 43% of them from one treatment school; baseline imbalance in FSM, EAL and SEN (p. 32). Kill withheld: a pilot that "did not seek to assess impact on attainment in a robust way" (p. 3), 17 schools, about 18% missing, uneven implementation, a general English outcome rather than comprehension of texts on the taught topics, one year only; Kim et al. 2021 and Grissmer et al. 2023 found gains. Narrowed: knowledge-building has improved comprehension in some trials, not consistently. Not a duplicate (only prior entry: Grissmer et al. 2023, a positive kill-condition test). Not a preprint. Status alive → weakened. Claim and Bounds rewritten same turn; brief and the practical rule rewritten.

3. #3 narrowed 22 October 2026 · Cabell, Kim, White, Gale, Edwards, Hwang, Petscher & Raines 2025 · target: kill-condition

   In two school-randomised trials, the second a replication of the first (47 schools in two large urban US districts, 1,194 kindergarteners), a semester of the most widely used knowledge-building curriculum (Core Knowledge Language Arts: Knowledge Strand, one hour of topic read-alouds a day) taught the words and topics but left standardised listening comprehension unchanged in both trials (pooled Test of Narrative Language ES 0.027, 95% CI −0.092 to 0.147; CELF-4 Sentence Structure −0.031, −0.121 to 0.059), so knowledge-building teaching does not reliably improve comprehension.

   Strongest reading and verdict reasoning

   Steelmans to: C17's kill-condition includes "knowledge-building teaching does not improve comprehension"; unlike the small English pilot already in the claim, this is a replicated, well-powered, school-randomised test with a true pre-randomisation baseline, minimal attrition and standardised comprehension outcomes, of Core Knowledge, the very approach the knowledge-building prong rests on, and the knowledge itself was clearly learned (taught vocabulary ES 0.56), yet comprehension did not move.

   Why: Nightly-attacks 2026-10-22 Europe/Stockholm. DOI checked on Crossref (title, eight authors, JEP 117(2) 153–175, February 2025). The published article is open access at APA but its PDF sits behind an Incapsula bot check from the box and from the fetch tool, so the published abstract (University of Minnesota research portal) and the authors' 2020 structured report of the same two trials (Florida State University, public PDF, 6 pages with methods, results and a pooled table; stored privately as ) were read. Report: Study 1, 23 schools (11/12), mid-Atlantic district, 2017–18; Study 2, 24 schools (12/12), Southern district, 2018–19; 82% and 44% free or reduced-price lunch; pre-test before randomisation, baseline equivalence, minimal missing data, no differential attrition; ITT multilevel models; both studies "no discernable main effects on standardized measures of... listening comprehension" (p. 3); pooled Table 1: TNL 0.027 (p = .655), CELF-4 −0.031 (p = .502), PPVT 0.001, WJ-III Picture Vocabulary 0.085 (p = .016), WJ-III Science 0.116 (p = .044), WJ-III Social Studies 0.01, proximal vocabulary 0.559 (p < .001). The published abstract reports gains on proximal vocabulary and science and social-studies knowledge, and none on listening comprehension. Kill withheld: one semester only; kindergarten listening comprehension, not reading comprehension of texts; pooled numbers are the authors' preliminary estimates (the final article may differ slightly; not read); Kim et al. 2021 (first grade, listening comprehension ES .40) and Grissmer et al. 2023 found gains, and Hwang, Cabell & Joyner 2022 (abstract only) a positive meta-analytic average. Not a restatement: the existing consistency bound rests on one small pilot with about 18% missing results and a general English outcome (See et al. 2017); this is a replicated, low-attrition randomised null on a comprehension measure. Not a duplicate (prior entries: Grissmer et al. 2023, See et al. 2017). Not a preprint. Status stays weakened. Claim and Bounds rewritten same turn; brief and practical rule and limits rewritten (headline unchanged).

4. #5 narrowed 30 October 2026 · Kim, Burkhauser, Mesite, Asher, Relyea, Fitzgerald & Elmore 2021 · target: claim-text

   The upper end of C17's stated gain from teaching topic knowledge ("roughly 0.1 to 0.4 standard deviations" for "a school year or more") rests on a 10-day first-grade unit's four-item listening measure (Kim et al. 2021: ES .40, Cronbach's alpha .35), not on a school year of teaching; the same trial moved a general reading test by only ES .11, and the only school-year-or-longer randomised or lottery gains on file are 0.14 to 0.24 SD (Grissmer et al. 2023).

   Strongest reading and verdict reasoning

   Steelmans to: C17's size range should describe what knowledge teaching does to general reading or comprehension tests over ordinary teaching; the 0.4 figure is a proximal, low-reliability listening measure after ten lessons, so a range topping out at 0.4 for a school year or more overstates the evidence on file.

   Why: Nightly-attacks 2026-10-30 Europe/Stockholm. Full text: "MORE lessons included a 10-day thematic unit" (abstract); 38 first-grade classrooms, N = 674, ten 60-minute lessons within a 3-week window; listening comprehension = four inferential questions on one 189-word passage, reliability .35, ES .40; distal MAP reading ES .11 (Table 5). The school-year-or-longer evidence on file is Grissmer et al. 2023 (ITT 0.143; 0.241 reduced sample; working paper), See et al. 2017 (−0.03) and, for one semester, Cabell et al. 2025 (listening 0.027). The range shrinks, so this is a narrowing, not a correction. Not a duplicate (prior entries: Grissmer et al. 2023, See et al. 2017, Cabell et al. 2025, the 2026-10-29 sharpening); not a restatement of an existing Bound (Bounds gave Kim's ES .40 without its duration or measure). Not a preprint. Status stays weakened. Claim field, Bounds, brief and the practical rule rewritten in the same turn; headline unchanged (it was not stronger than the new claim).

5. #7 narrowed 4 November 2026 · Kim & Cao 2025 · target: kill-condition

   The largest meta-analysis of causal studies of knowledge teaching and comprehension finds that knowledge-building teaching, as opposed to activating knowledge pupils already have, did not significantly improve reading comprehension (g = 0.19, p > .05), and that effects on norm-referenced tests (0.21) and in multi-group designs (0.18) are much smaller than on researcher-made tests or in single-class designs; so C17's "teaching topic knowledge has raised scores" rests on individual trials, not on a secure pooled effect of knowledge building.

   Strongest reading and verdict reasoning

   Steelmans to: C17's kill line includes "knowledge-building teaching does not improve comprehension"; a 55-study causal meta-analysis is the strongest available test of that prong, and its estimate for knowledge building is not distinguishable from zero.

   Why: Nightly-attacks run of 2026-11-04 Europe/Stockholm. Full text: causal data 55 studies, 304 effect sizes, N = 18,540; reading comprehension g = 0.44 (95% CI 0.18–0.70; 51 studies; I² = 97.75%); "knowledge activation instruction produced a statistically significant effect (g = 0.66, p < .01), while knowledge building instruction did not show a significant effect (g = 0.19, p > .05)", difference p = .045; normed tests g = 0.21 vs non-normed 0.51 (p < .01); non-N-of-1 designs 0.18 (p < .01) vs N-of-1 0.67; listening 0.13 (95% CI −0.41 to 0.68; 5 studies); grade, reading skill, SES not moderators; correlational r = .41 (108 studies). Kill withheld: a non-significant moderator estimate under very high heterogeneity is not a finding of no effect, the point estimate (0.19) lies inside the claim's 0.1–0.25 range, and the authors read both kinds of instruction as positive. Not a survival either: the prong the kill line names is not shown to hold in the pooled evidence, so the claim must say so. Not a duplicate (prior entries: Grissmer 2023, See 2017, Cabell 2025, the 2026-10-29 sharpening, Kim 2021, Reynolds 2024 correction); not a restatement (Bounds held only Hwang et al. 2022's abstract average of 0.25). Not a preprint. Status stays weakened. Kill-match check (the routing method): the new hedge concedes that the pooled knowledge-building estimate is not significant, not that knowledge building fails; prongs 1 (topic knowledge does not help once reading skill is accounted for) and 3 (strong readers without knowledge clearly beat weak readers with it) are untouched; not a kill-mismatch, but prong 2 is now close to conceded — flagged for Jonatan.