New Australian education research, read close to publication. The register holds findings that have already held up; this page is the forward edge — what has just landed, what it found, and what would weaken it.
Each item is a newly published study or dataset, read against the settled register. Every claim about a study is checked against the study itself before it is written; where a finding is preliminary, contested, or rests on a small or unrepresentative sample, that is stated in the reading rather than omitted. Items are dated on the day they were read. An item is never silently revised: if a register entry it cites is later corrected, the item is corrected with it and carries a dated correction note on its face, keeping its id and its place in the sequence.
Reviewed weekly; new items added when substantive research is published, not on a quota. An item marked In the registerhas since been verified to the register’s full standard and cited entries are linked; one marked Watching has been read but not yet registered, and should be treated as provisional.
The OECD's Australian country note for PISA 2025 gives reading-only proficiency shares in the form the register's PISA 2022 pair can be compared against — 77% of Australian 15-year-olds at Level 2 or above and 11% at Level 5 or above — where two register entries and the reading synthesis still publish 79% and 12%
What it found
An OECD country note accompanying PISA 2025 Volume I, carrying a published date of 8 September 2026 and stating that results "are released on 8 September 2026", written by Francesco Avvisati of the OECD Directorate for Education and Skills. It reports Australia's results against the OECD average for a nationally representative sample of 15-year-olds. THE SAMPLE, now read at source rather than inferred from a landing page: "In Australia, 14 658 students, in 766 schools, completed the assessment, representing about 275 400 15-year-old students (an estimated 83% of the total population of 15-year-olds)", within a survey where "Over 760 000 students took the assessment in 2025, representing about 33 million 15-year-olds in the schools of the 91 participating countries and economies". SCOPE, stated by the note itself: "The PISA 2025 survey focused on science, with mathematics and reading as minor areas" — reading was a MINOR domain in this cycle. READING PROFICIENCY, in the subject-specific form the register's PISA 2022 figures can be set against: "Some 77% of students in Australia attained Level 2 or higher in reading (OECD average: 69%)" and "In Australia, 11% of students scored at Level 5 or higher in reading (OECD average: 6%)". The note adds that the share attaining Level 2 or higher in reading "varied from 93% in B-S-J-Z (China) to 4% in Rwanda". TREND: "Average 2025 results in Australia were lower than in 2022 in mathematics and reading, and about the same as in 2022 in science", and "In reading and mathematics, in contrast, the recent decline extends a downward trend that began much earlier, and the most recent results are the lowest ever measured in Australia." On composition, "Compared to 2015 the proportion of low-performing students (scoring below proficiency Level 2) increased by seven percentage points in mathematics; increased by five percentage points in reading; and did not change significantly in science", while "Over the most recent period (2022 to 2025), the gap between the highest-scoring students (10% with the highest scores) and the weakest students (10% with the lowest scores) did not change significantly in mathematics, reading and science." FOR CONTRAST, the note's SCIENCE figures, recorded here only because they are a live confusion risk: "Some 82% of students in Australia attained Level 2 or higher in science (OECD average: 74%)" and "In Australia, 12% of students were top performers in science" (OECD average: 7%).
How to read it
The value of this document is narrow and specific: it is the like-for-like successor to the PISA 2022 country note that RDG-13 and WRI-16 both cite, and it moves their headline figures. Those entries say 79% of Australian 15-year-olds reached Level 2 or higher in reading against an OECD average of 74%, and 12% reached Level 5 or higher against an OECD average of 7%. The 2025 note gives 77% against 69%, and 11% against 6%. The reading synthesis publishes the old pair in prose — "79% of Australian students still clear the baseline (Level 2) reading standard, above the OECD average of 74%, and 12% are top performers against a 7% OECD average" — so this is a live figure on the site, not only a register field. Note what actually moved: Australia fell 2 and 1 percentage points, while the OECD average fell 5 and 1, so Australia's MARGIN over the OECD average widened even as its own share fell. A rewrite that reports the Australian fall without the OECD fall would make the picture look worse than the note does. THIS IS NOT A DUPLICATE OF FR-19. FR-19 reads Volume I proper and carries the mean scores (reading 491, change -7 from 2022) and Volume I's Table I.1 top- and low-performer shares, which are CROSS-SUBJECT definitions — 18.2% top performers at Level 5 or 6 in science, reading OR mathematics, and 14.0% low performers below Level 2 in science, reading AND mathematics. Those are not comparable to RDG-13's reading-only shares, and the register's PISA 2022 figures did in fact survive FR-19 untouched — RDG-13 and WRI-16 still read 79% and 12% today. This note supplies reading-only figures that ARE comparable. What is NOT established here is that Volume I omits them: Volume I runs to 377 pages and only its Table I.1 and Table I.2.9 were read, by FR-19, so the claim is that this note carries the comparable form, not that the full report lacks it. It also supplies, in the OECD's own words, a reading-SPECIFIC long-run statement — reading results are "the lowest ever measured in Australia" — of a kind RDG-12 explicitly cannot make, since RDG-12's >25-point decline was corrected on 2026-09-11 to the OECD's joint mathematics-and-reading figure. THE TRAP, worth naming because the numbers invite it: the PISA 2022 reading OECD averages that RDG-13 and WRI-16 carry (74% and 7%) are numerically IDENTICAL to this note's 2025 SCIENCE OECD averages (74% and 7%), and Australia's 2022 reading top-performer share (12%) is identical to its 2025 science top-performer share (12%). Anyone updating these entries against this document while reading the wrong paragraph would produce figures that look plausible, change little, and be about the wrong subject. WHERE THIS IS LESS INTERESTING THAN IT LOOKS: the note reports no significance test for the 79%-to-77% share change, and says nothing about it at all. Its only significance statement about the composition of the reading distribution uses a 2015 baseline, not 2022, and is about the below-Level-2 group (up five percentage points). So this document establishes that the published shares have a newer successor; it does not establish that Australia's baseline share fell significantly between 2022 and 2025.
What would weaken it
This is a country note — a short secondary summary of Volume I written by a single OECD author — not the full report, and where the two differ Volume I governs. The 77% and 11% are printed as bare point estimates: the note gives no confidence interval, no standard error and no significance mark for either, so neither may be presented as a measured decline from 2022 on this document's authority. Cross-cycle comparison of proficiency shares also assumes the PISA reading scale is linked across 2022 and 2025; the note asserts that linkage nowhere, and reading was a MINOR domain in this cycle, which means fewer reading items per student and wider uncertainty than the science figures alongside it carry. Only the note's PROSE was read: every figure above comes from a sentence, and the note's five charts (Figures 1-5) were not read, so any Australian value that exists only in a chart is not represented here. The document was retrieved through the r.jina.ai rendering of the OECD HTML page, which returns 403 to plain curl from this network, so the text was read through a proxy rather than from OECD's own server — a PDF is offered on the page and was not pulled, and diffing it against the PDF before any figure here is published load-bearingly would be cheap insurance. Finally, this note does NOT close the ACER gap: ACER's Australian national report remains unread and Cloudflare-blocked, and no state or territory figure appears here, so nothing in this item supports a Victorian claim.
Queensland's 2025 EAL subject report touches three of the points VCAA's English and Literature examiners raise — moving beyond plot recount, addressing the question stem, and avoiding prepared material — but states them as attributes of effective responses rather than as observed faults, and only the plot-recount pairing is tight; cohort of 1,004
What it found
QCAA's annual subject report for Queensland senior English as an Additional Language, dated January 2026 and covering the 2025 cohort. Scale: "Number of schools that offered English as an Additional Language: 42." The subject highlights page states "42 schools offered English as an Additional Language", a "49.93% increase in enrolment since 2024" and "97.56% agreement with provisional marks for IA3". Unit completion was 1,131 students for Unit 1, 1,089 for Unit 2 and 1,004 for Units 3 and 4; Unit 1 results were 1,057 satisfactory against 74 unsatisfactory, Unit 2 1,052 against 37. Grade boundaries are given as A 100-84, B 83-64, C 63-41, D 40-17, E 16-0, and the distribution of standards across the state was A 211 (21.02%), B 496 (49.40%), C 292 (29.08%), D 5 (0.50%) and E 0 (0.00%). Internal assessment ran 43 instruments at each of IA1, IA2 and IA3, with percentages endorsed at Application 1 of 72, 74 and 86 respectively; confirmation covered 41 schools with 308, 308 and 306 samples requested and percentage agreement with provisional marks of 100.00, 100.00 and 97.56. The external assessment is an "analytical extended response (25%)" — "an analytical response to a literary text from the prescribed text list in the form of an analytical essay for an audience with a deep understanding of the text (35 marks)" — and the report states "There were 16 extended items — two for each of the eight text options." The eight texts are All the Light We Cannot See (Anthony Doerr), Burial Rites (Hannah Kent), Macbeth, Never Let Me Go (Kazuo Ishiguro), Othello, Pride and Prejudice (Jane Austen), The Dry (Jane Harper) and The Yield (Tara June Winch). The prescribed-text highlights page counts students studying six of these for external assessment: Macbeth 663, Othello 274, All the Light We Cannot See 51, Never Let Me Go 46, The Yield 38 and Burial Rites 9. On student performance, the external-assessment "Effective practices" list says strong responses "used a formal, academic register appropriate to analytical writing and avoided colloquial or overly emotive language", and the samples commentary says effective student responses "answered the examination question, and did not incorporate elements of a prepared response". Under "Practices to strengthen", teachers are told to "focus on directly addressing the question stem" and to "teach students to move beyond plot recount by linking evidence to interpretation", the latter carried by a worked contrast on Othello between summary ("Emilia gives Desdemona's handkerchief to Iago, who manipulates Othello.") and interpretation ("This act, seemingly minor, triggers tragedy and highlights concepts of trust, betrayal, and moral consequence, prompting readers to reflect on human vulnerability and responsibility."), and by a parallel credible-versus-discriminating thesis contrast on the same play. The report names two prescribed lists: internal-assessment advice refers to the "Prescribed text list: English and EAL 2023–2025", and additional advice states that "Prescribed texts must be selected from the updated Prescribed text list: English and EAL 2026–2029". It also records two condition changes between the 2019 and 2025 syllabuses: "Conditions now allow 100 words of quotations from the studied text/s not 200 words of quotations" and "Conditions now state time allocation but not word length."
How to read it
This is the third of QCAA's 2025 senior English subject reports the frontier has read, after the English report at FR-26 and the Literature report at FR-38 — whether QCAA published others for the same cohort was not checked — and it is the weakest of the three as corroboration — which is the finding, not a disappointment. FR-38 could quote QCAA's Literature assessors describing what students actually did wrong, in words that sit alongside VCAA's. This report mostly does not. Its overlap with the VCE register is stated as a description of what effective responses achieved, or as advice to teachers, rather than as an assessor's observation of a fault: "did not incorporate elements of a prepared response" tells you what strong scripts avoided, not that weak ones did it, so the counterpart to VCE-12 here is inferred rather than stated. Read carefully, only one pairing is genuinely tight. The recommendation to "teach students to move beyond plot recount by linking evidence to interpretation", with its Othello summary-versus-interpretation worked example, is the same distinction VCE-01 records (ideas named without the interpretive work behind them) and VCE-11 states as a warning (paraphrase does not meet the task) — and QCAA's example is more concretely teachable than either, because it prints both halves. The second reasonable pairing is VCE-05: QCAA's first Knowledge-application recommendation is question-stem-directed throughout, with worked examples for each stem type ('How is the reader invited to view', 'Analyse the significance of'), which is the same move VCAA rewards when it credits students who read the topic's actual wording. VCE-10 has no counterpart on the point it actually turns on: QCAA does flag register and recommend word banks of analytical verbs, but it names no misused word, which is exactly the weakness FR-38 recorded against the same entry — so two Queensland reports have now failed to corroborate VCE-10, and that repetition is worth more than either failure alone. The text-list line is a tighter comparator for VCE-35 and VCE-36 than FR-38's was, because it names an English and EAL list — the same category VCAA's tenure rule governs in Victoria, rather than Literature — running 2023–2025 and then 2026–2029. But it remains Queensland's list under a different authority, it states no policy and gives no reason, and it is not evidence about what VCAA will decide on its own text-cycle consultation, which is still closed and undecided. Finally, the scale deflates the report's apparent breadth. Eight text options is nominal: Macbeth and Othello alone account for 937 of the 1,081 students the six listed counts cover — both of those are sums of the report's own numbers, not totals it prints — and the cohort completing Units 3 and 4 is 1,004, against the 6,351 Literature students FR-38 records from the companion report and the 57,909 Queensland Year 12 graduates at SEN-06. This is a small subject, and any inference from it carries that.
What would weaken it
The framing risk is the main one. This document does not support the sentence 'QCAA's EAL assessors name the same faults VCAA names' — FR-38 supports that sentence for Literature, and this one does not, because nearly all the overlap is written as effective practice or as teacher advice rather than as observed student behaviour. Anything on the site that leans on it must say so. Second, the prescribed-text counts cannot be used to compute shares: the six listed totals sum to 1,081 against 1,004 students completing Units 3 and 4, two of the eight texts are given no count at all, and the heading counts students 'studying' the texts rather than sitting the examination on them — so the list is not a partition of the completing cohort and the report does not explain the gap. Third, the 49.93% enrolment increase has no base figure anywhere in the document, so it cannot be checked from the source and should not be repeated without one. Fourth, jurisdiction and structure differ: Queensland assesses EAL through three internal instruments plus a 25% external examination marked against criteria (Knowledge application, Organisation and development, Textual features), while VCE EAL is structured differently, so a criticism arising under QCAA's marking guide does not automatically transfer. Fifth, this is a backlog read, not new publication — the report is dated January 2026 and was queued as unread, and it describes a cohort taught under the 2019 syllabus while the 2025 syllabus, whose changed conditions the report itself lists, is now in force. Sixth, the two condition changes and the 2026–2029 list are reported from the report's own 'Additional advice', not verified against the syllabus or the list document itself.
Queensland's 2025 Literature assessors name the faults VCAA's Literature examiners name — pre-prepared essays, paraphrase in place of interpretation, jargon that outruns control — and the report names two prescribed Literature lists, 2023–2025 then 2026–2029, while VCAA's own text-cycle consultation sits closed and undecided
What it found
QCAA's annual subject report for Queensland senior Literature, dated January 2026 and covering the 2025 cohort. Scale: "Number of schools that offered Literature: 229." Unit completion was 6,848 students for Unit 1, 6,609 for Unit 2 and 6,351 for Units 3 and 4; Unit 1 results were 6,714 satisfactory against 134 unsatisfactory, Unit 2 6,516 against 93. Internal-assessment confirmation covered 228 schools with 1,754, 1,740 and 1,733 samples requested for IA1, IA2 and IA3, and percentage agreement with provisional marks of 95.18, 91.67 and 96.93 respectively. The external assessment is an "analytical written response (25%)" — "an analytical response to a literary text from the prescribed text list in the form of an analytical essay for an audience with a deep understanding of the text. (35 marks)" — on one of eight set texts. Student numbers by text were Hamlet 2,223, King Lear 2,040, In Cold Blood 1,399, Wuthering Heights 427, Stasiland 134, Terra Nullius 117, Catch-22 83 and Beloved 69. The "Assessment decisions" section for the external examination separates "Effective practices" from "Practices to strengthen". On interpretation, effective responses "developed an authoritative and distinctive interpretation that moved beyond rephrasing the question. Successful responses used the question as a springboard into its significance"; under practices to strengthen, teachers should "encourage students to craft thesis statements that take a clear, arguable position, rather than summarising the quotation or the plot", and where a question refers to a comment or a character's claim, "students should move beyond paraphrase and instead offer a judgment that connects the statement to the text's larger conceptual concerns". On memorised work, the report asks teachers to "support students to avoid relying on pre-prepared or generic essay structures that are not responsive to the specific demands of the question or task. When students attempt to adapt a memorised interpretation rather than developing an argument directly from the prompt, the thesis often becomes disconnected from the question, resulting in limited internal logic and reduced validity." On vocabulary, effective responses "selected precise vocabulary appropriate to a literary analysis, avoiding vague or formulaic language" and did so "without reverting to unnecessary technical jargon"; the report separately asks teachers to "reinforce that students are not required to apply formal literary theory in the external assessment, and that attempts to force theoretical frameworks can lead to superficial or formulaic analysis that does not directly address the question or task." On surface analysis, "Some students rely on identifying isolated or surface-level devices that do not clearly connect to how the text constructs perspectives or invites readers to take up positions". Three further recommendations have no counterpart in the Victorian entries read against: teachers should "discourage self-referential or procedural statements, such as 'This essay will analyse …'"; "Cohesion should emerge from the internal logic of the argument (e.g. comparison, causation, conceptual progression), not from formulaic connective words."; and they should "encourage students to limit extensive historical or biographical detail, as overemphasis on authorial context may detract from purposeful textual analysis". Separately, and not presented by QCAA as a policy statement, the report names two prescribed lists by title: the 2025 instruments were built against the "Prescribed text list, Literature 2023–2025", and schools are told to "ensure they refer to the Prescribed text list, Literature 2026–2029 when developing the examination".
How to read it
Two separate things, of unequal strength. The first is corroboration. VCE-09 to VCE-13 are Argo's VCE Literature assessor claims, and the standing risk with all of them is that they describe one authority's marking culture rather than something true about analytical writing about literature. This is a different state, syllabus, cohort and examining body in the same year band, and three of those five claims have a clear counterpart here in QCAA's own words: VCE-12's pre-prepared 'all-purpose' introductions appear as memorised interpretations adapted to the prompt, which is the tightest match in the document and close to verbatim in substance; VCE-09's summarising the passage rather than exploring its significance, and VCE-11's warning that paraphrasing does not meet the task, appear as theses that summarise the quotation or plot and as the instruction to move beyond paraphrase to judgment; and VCE-13's jargon that disguises rather than reveals meaning appears as the caution against unnecessary technical jargon, reinforced by a distinctly Queensland addition VCAA does not make — that students are not required to apply formal literary theory at all, and that forcing a framework produces formulaic analysis. VCE-10 is the weak pair and should not be reported as corroborated: it turns on specific misused words ('juxtapose' for 'contrast', 'showcase' as an analytic verb, 'simple' against 'simplistic'), and QCAA names no misused word anywhere — it asks only for precise vocabulary over vague or generic terms, which is a far more general instruction. The pairing throughout is an interpretive mapping made here, not a correspondence either authority draws; the two documents share no phrasing. The second thing is a live policy comparator and is thinner but more time-sensitive. VCAA's Notice 83 consultation on extending its English/EAL and Literature text cycle closed on 11 September 2026 with no outcome notice yet published, and VCE-35 and VCE-36 are Argo's text-tenure entries. This report names a Queensland Literature list covering 2023–2025 and its successor covering 2026–2029 — three years then four, on the face of the titles. That is arithmetic from two list names, not a QCAA announcement: the report nowhere says the cycle was extended, gives no reason, and describes no policy. It is worth holding as a second jurisdiction whose Literature list currently runs on a four-year span, and it is not evidence about what VCAA will decide.
What would weaken it
Retrieved through the r.jina.ai text proxy because qcaa.qld.edu.au returns HTTP 403 to direct curl and to WebFetch — the same route and the same limitation as FR-26. The rendering carried page furniture and footers and every quoted sentence above was located in it, but this is a proxy rendering of a PDF and not the PDF itself, and the tabular sections (IA1-IA3 mark distributions, grade boundaries, final results) rendered as headings with their data missing, so nothing is claimed from them. The report is qualitative assessor commentary: its effective/weaker contrasts carry no counts, proportions or effect sizes, so nothing here establishes how common any fault was, and the illustrative examples were chosen by markers. It is also not disinterested — QCAA is describing responses to its own marking guide, so agreement with VCAA may reflect a shared professional discourse about literary analysis rather than two independent observations of student writing, and neither document is evidence that the advice improves writing. The cohort is Queensland Year 12 Literature under the QCE, whose external examination carries 25% against a Victorian Literature examination weighted differently, on a different text list and a different syllabus; none of the figures describe Victorian students. The subject-level figures are 'correct as at January 2026' by the report's own note and percentages are rounded to two decimal places. The eight per-text external-assessment numbers sum to 6,492 against the 6,351 reported as completing Units 3 and 4 — the two counts are drawn on different bases and the report does not reconcile them, so neither should be reported as the cohort size. The confirmation agreement percentages are internal-assessment confirmation, NOT external-examination marker reliability, and must not be compared with the examination agreement figure quoted in FR-26. On the text lists: the three-year and four-year spans are read off two list titles, QCAA's historical rotation practice was not checked, and the list named for Literature here is a different list from the English external-assessment category whose minimum-three/maximum-six rule FR-26's source states.
The same NSW research programme that said EAL students need at least six years to reach parity (FR-03) now reports a second estimate from teacher judgement: a median of five years and two months
What it found
Phase 2 of the AERO-commissioned NSW study. Where Phase 1 (FR-03) used NAPLAN, this paper uses teachers' assessments of 111,135 EAL students enrolled in NSW public schools from 2014 to 2022 against the EAL/D Learning Progression phases (Beginning, Emerging, Developing, Consolidating). It reports: “The analysis shows that, for all students who entered school in Australia between kindergarten and Year 9, those starting at the Beginning phase of English development required a median time of 5 years and 2 months to reach the Consolidating phase of English.” The spread is wide: “some took up to 6 years and 7 months to reach the Consolidating phase, while others took only 3 years and 10 months” (the interquartile range). Step by step: “The median time taken to progress from the Beginning to Emerging phase was approximately 1 year and 1 month; from Emerging to Developing a little over 1 year and 6 months…; and from Developing to Consolidating around 2 and a half years.” The authors set the two estimates side by side: “The NAPLAN analysis found that most students who entered school with Beginning/Emerging phases of English required at least 6 years to achieve parity in academic English with their non-EAL peers, while the teacher judgement analysis found a median of 5 years. These differences likely reflect differences in the nature of the datasets and methods of analysis.” They add that NAPLAN, sat every two years, “may be a lagging indicator of when students reach parity”, and that teacher judgements may record the first time a student reaches Consolidating and so “may not take account of students’ possible regression.”
How to read it
This sharpens FR-03 rather than contradicting it. Two instruments on overlapping cohorts give five years by teacher judgement and six or more by NAPLAN, and the paper explains the gap: oracy counted in one and not the other, a two-year test cycle, and regression invisible to a first-crossing measure. The defensible sentence for an Argo page is now 'five to six years, depending on how proficiency is measured', with the Developing-to-Consolidating step, about two and a half years on its own, as the part families most underestimate. It also strengthens EAL-04, the department's own advice, which sits inside the same range.
What would weaken it
CAVEAT RETIRED 2026-09-16, and the retirement is the point. This item was read through a text-proxy rendering of a Springer page the sweep believed paywalled, and it carried a condition: diff an authoritative copy before citing it load-bearingly. That condition is now met. The article is OPEN ACCESS (published 6 August 2026, not June as first recorded) and was read in full in a browser on 2026-09-16; every sentence quoted below was found verbatim in the authoritative text. The proxy read was accurate. Registered as EAL-13. THE STANDING LIMITS ARE UNCHANGED AND STILL APPLY: the 111,135 students are those assessed at least four times, non-First Nations, without special-needs support -- a cleaner sample than all EAL learners -- and the median is for entrants who started at Beginning. Teacher judgement is not a standardised measure and varies by school. AND THE LIMIT THAT MATTERS MOST: this is the SAME PROJECT AND SAME DATA as FR-03/EAL-12, analysed a second way. It is not an independent replication and must never be cited as a second study agreeing with the first.
In a small Italian study, dyslexic middle-schoolers learned new words better by typing than by handwriting, while typical readers showed no difference
What it found
A within-subjects learning experiment: “Eighteen Italian middle school students with DD and eighteen age-matched TD readers participated”, learning invented nonwords paired with pictures either by handwriting or by typing on a tablet. “TD children benefited from both writing modalities during learning. Crucially, the DD group showed better spelling and nonword-image association performance when learning occurred through typing.” The dyslexic group “retrieved the orthographic form more accurately when nonwords were learned through typing than through handwriting, suggesting better spelling recall (z = 2.32, p = .021). In contrast, performance in the TD group was at ceiling in both learning condition (z = 0.86, p = .392).” Recognition of the referent was likewise more accurate (z = 2.55, p = .011) and faster (t = 2.64, p = .009) after typing for the dyslexic group only. The authors conclude that “the complex graphomotor demands of handwriting may hinder novel word acquisition in dyslexic individuals” and that typing “improved their spelling accuracy to a level comparable to that of their peers”.
How to read it
The first genuine reversal in the handwriting domain, and a narrow one. The register's handwriting-advantage findings (HND-05, HND-06, LRX-06) come from typical learners forming letter and word memories; this study says the same graphomotor effort that helps typical readers can be the bottleneck for a dyslexic reader. Read as a boundary condition on 'handwriting builds memory', not as a finding about handwriting generally: the effect is confined to the group for whom writing by hand is itself the hard part. For Argo's handwriting pages the implication is a caveat about who the advice is for, not a change to the advice.
What would weaken it
Eighteen children per group, one session, 32 invented nonwords, uppercase letters on a tablet: a laboratory task, not classroom instruction over time, and the authors flag uppercase-only output as a limit on ecological validity. Italian is a shallow orthography, so transfer to English spelling is not established. The two mirror PDFs were reached through a text proxy after the publisher and both repositories served bot challenges; the statistics were matched against the extracted text, including one glyph artefact in the z = 2.55 line. No independent replication exists. FOURTH ACCESS ATTEMPT, 2026-09-16, and the verdict is now settled: the primary text is NOT REACHABLE by any route this process has. ScienceDirect serves a CAPTCHA, which will not be solved. sfera.unife.it returns 403 to curl on both the guessed path and the real one, and in a real browser serves a Cloudflare 'Performing security verification' interstitial that does not clear -- so the institutional-repository route that rescued RDG-23 fails here too. Recorded so a fifth attempt is not spent: DO NOT RE-TRY these routes. THE CITING-PAPER TECHNIQUE WAS TRIED AND PARTLY PAID: Cerni's own 2026 Springer follow-up on orthographic complexity (doi 10.1007/s11145-025-10693-1) is OPEN ACCESS and reads fine in a browser. It does not carry the dyslexia reversal -- no dyslexia group, different question -- but it CONFIRMS THE CITATION verbatim: Cerni, T., Lonciari, I., & Job, R. (2025), Learning and Instruction, 98, Article 102119, doi 10.1016/j.learninstruc.2025.102119. So the bibliographic record is now solid and only the findings are missing. If this is ever wanted, the remaining routes are a library copy or writing to the authors, both of which are Callum's to take.
A 26,811-student Chinese panel finds generative AI raises homework scores and lowers exam scores, with the loss concentrated among students whose usage pattern looks like outsourcing
What it found
From the abstract: “Using 30 months of panel data on 26,811 Chinese students in grades 7--12, we study how generative AI affects homework productivity and learning. The data combine monthly closed-book exams, high-school and college entrance exams, and homework scores and completion time across nine subjects. We exploit staggered AI adoption in a difference-in-differences design. AI adoption raises homework scores by 18% and reduces completion time by 30%, but lowers monthly exam scores by 20% within six months. High-stakes entrance-exam scores fall by 18 and 24%, with the full penalty emerging only after about two years. The losses are largest in social science subjects, followed by STEM and languages, and are especially large for junior students, high-achieving students, and boys. The learning losses are concentrated among roughly 80% of AI users whose behavior is consistent with homework outsourcing, as indicated by exceptionally short homework completion time coupled with high homework scores. AI users who maintain similar homework completion time as non-AI users experience small learning losses.”
How to read it
The largest secondary-school dataset yet on the question the register's AI-tutoring trials (AIE-07, AIE-08, FR-05) approach from the other side: those trials gave students a designed tutor and found gains only where it was purpose-built and constrained (AIE-08), or none beyond the platform without it (FR-05); this observes students adopting general-purpose AI on their own and measures a loss on every closed-book assessment. The mechanism the paper proposes, outsourcing signalled by fast completion plus high homework marks, is the same one Argo's writing pages already describe. The line worth carrying is the last one: users who kept spending normal time on homework lost little.
What would weaken it
A working paper, not peer-reviewed; only the abstract was read at source, so no figure beyond the abstract's should be cited from this item. China, grades 7 to 12, nine subjects: the 'languages' category is not English writing, and nothing here is Australian. Difference-in-differences on self-selected adopters is not randomisation; the paper's identification rests on staggered adoption timing. Dose-response figures reported in the technology press were not verified. PROMOTED TO AIE-17 on 2026-09-16, re-verified at CEPR's own page. Every figure matched, and the re-read surfaced one sentence this item had not quoted which is the most useful in the abstract: 'AI users who maintain similar homework completion time as non-AI users experience small learning losses.' That turns the finding from 'AI lowers scores' into 'outsourcing lowers scores', which is actionable and is what the page now says. Registered TIER 3, deliberately below AIE-16: unreviewed AND difference-in-differences on self-selected adopters rather than randomisation. Still abstract-only -- the full paper is behind a GBP 6.00 charge.
eSafety's first national survey of children's AI use: 78% of Australian 10-to-17-year-olds have used an AI assistant, and of those, 70% have used one for writing or editing
What it found
“Drawing on data from an online survey conducted between 2 February and 4 March 2026 of 1,950 children aged 10 to 17 living in Australia”, the report finds: “Most children in Australia have used an AI assistant. Around 4 in 5 children aged 10 to 17 (78%) told us they had ever used one or more AI assistants. ChatGPT was the most popular AI assistant, with 70% having used it, followed by Gemini (20%) and Copilot (20%).” Among children who had used an assistant or companion, “most had done so because they wanted help with schoolwork, organisation or problem solving (68%)”, and “Most children who had used AI assistants or companions had used them for functional reasons, such as to find information, to support learning or translation (91%), or for writing or editing (70%), as well as for entertainment (68%).”
How to read it
The Australian counterpart to the US teen figures in AIE-02, AIE-03 and FR-07, and stronger than AIE-13 (a tutoring company's own survey) because the sample is nationally representative and the publisher is a statutory regulator. The base matters: 70% is of users, so roughly 55% of all Australian 10-to-17-year-olds report having used an AI assistant for writing or editing at least once. Read with FR-20 (ACARA's Year 6 and Year 10 monthly-use figures): two independent national instruments in the same year now put AI writing help among a majority of secondary students.
What would weaken it
Self-reported, online panel, 'ever used' rather than frequency; 'writing or editing' is not defined as schoolwork and includes messages and captions. Ages 10 to 17 pooled: no year-level breakdown was read for the writing figure. First cycle of this survey, so no trend. The report's companion-app findings, its main subject, are outside this reading.
A ten-year panel of 3.2 million maths-platform interactions shows study time on AI-susceptible problems falling after ChatGPT, most for high-schoolers and not at all for Grade 5, with proctored retention falling alongside it
What it found
A US quasi-experiment comparing problem types an AI can answer (text word problems) with ones it cannot (interactive graph-based), before and after ChatGPT's release: “Learning time on AI-susceptible problems declines 2.8% per quarter among college students after ChatGPT's release, cumulating to 26.9% over eleven quarters; high-schoolers show 31.3%, middle-schoolers 9.0%, and Grade 5 students no detectable change. Among college students, the post-ChatGPT divergence vanishes entirely under proctoring, ruling out broad efficiency gains as the likely explanation. Logistic fixed-effects models on randomly assigned proctored retention items yield a 25% cumulative decline in odds of correct response; the same estimator on non-proctored assessment produces a large opposite-signed increase.” The authors state the limit themselves: “The individual-level causal chain — a specific student used AI during learning and later failed a retention test — cannot be established with current data.”
How to read it
Mathematics, not writing, so this is a mechanism analogue for Argo's pages rather than a writing finding. What it adds to FR-23 is the age gradient: the decline scales with the autonomy students have over their own study, absent in Grade 5 and largest in high school, which is exactly the population Argo teaches. The proctored-versus-unproctored split is the cleanest signature of outsourcing in any study the register has read: scores rose where AI could be used and fell where it could not.
What would weaken it
Preprint, not peer-reviewed; US platform users, not a representative sample; AI use is inferred from behavioural signatures, never observed, and the authors say so. The 'cognitive surrender' framing is their theoretical vocabulary, not a measured construct. One of the four quoted passages was matched only in part against the extracted PDF text because of line-wrapping; the figures are as printed. ASSESSED AND NOT REGISTERED, 2026-09-16 -- a VALUE-GATE decision, not a verification one, recorded so it is not re-assessed cold. AIE-16 (randomised, a coaching tutor left unused) and AIE-17 (26,811 students, the penalty concentrated in the outsourcing pattern) were registered the same day and between them carry what this item would add. AIE-17 in particular already holds the supervised-versus-unsupervised divergence, from a larger sample, with AI adoption OBSERVED rather than inferred. This study infers AI use from behavioural signatures and never observes it -- its authors say so -- and it is mathematics rather than writing. The one thing it holds that the register does not is the AGE GRADIENT: no detectable change at Grade 5, 9.0% in middle school, 31.3% in high school. If a page ever needs to argue that the risk rises with a student's autonomy over their own study, this is the source to read in full first.
Queensland's 2025 English subject report names the same essay faults VCAA has named for a decade: retelling plot, treating characters as real people, theses that answer half the question
What it found
The report covers the 467 schools that offered General English in 2025 and states that 99.51% of students received a C or higher. On the external examination (an analytical response to a prescribed literary text), assessors write: “Less successful responses tended to retell plot events or treat Tommy as a real person rather than a constructed literary figure, limiting the sophistication of analysis”; “Weaker theses neglected to deal with the whole statement, and focused on what the text says about human nature generally”; “Weaker essays listed the various examples of the ways reputation was explored as a theme in the play through characters and events”; “Weaker responses revealed a pattern of repetitive long sentences that affected the fluency of the ideas or attempted complex phrasing without full control of grammatical relationships between sentence parts”; and “A lack of precision of language (e.g. not using the right words in the right places, or incorrect use of pronouns) or a lack of formality in register (e.g. in the use of colloquialism ‘Lady Macbeth goes crazy …’, ‘her guilty conscience eats her alive …’) detracts from the authority of the response.” Marker-reliability agreement on the examination is reported at 98.29% to 99.14% by criterion.
How to read it
The register's assessor-commentary evidence has been Victorian only (FR-02 and the VCE entries). This is the first Queensland report read, and its fault list is the VCAA list in different words: narrative retelling, character-as-person, partial thesis, listing evidence, uncontrolled sentences, informal register. The same faults recurring across two state systems with different exams is stronger evidence that they are properties of adolescent analytical writing than any one report can be. Worth a cross-state line on the essay-technique pages.
What would weaken it
Retrieved through a text proxy because qcaa.qld.edu.au returns 403 to direct fetches; the rendering carried page structure and all five quoted sentences were found in it. The report is written as guidance to schools, structured as effective practices and practices to strengthen, so the faults are illustrated rather than counted: no percentage of responses is attached to any of them. Queensland's General English, 2025 cohort, external examination only.
The 2025 VCE Literature examination report: passages analysed in isolation and readings the text cannot support are what separated low from high scores, with a Section A mean of 3.4 out of 6
What it found
On Section A (Developing interpretations): “More successful responses clearly articulated how and why the set passage was significant within the broader text, rather than treating it as a self-contained extract”; “Less effective responses relied heavily on narrative summary or paraphrase, with insufficient attention to how meaning was created through language or technique.” Question 1 mark distribution across 0 to 6: 1, 2, 15, 36, 31, 13, 2 per cent, average 3.4. On Section B (Close analysis): “High-scoring responses used detailed analysis of two or more set passages as a springboard to develop a sustained interpretation of the whole text, rather than treating each passage in isolation”; “Some responses offered readings that could not be supported by the passages provided, suggesting insufficient familiarity with the text”; and “Some responses were constrained by imprecise wording (including continued misuse of key analytical terms such as ‘juxtapose’), awkward sentence construction or unclear sequencing of ideas”. Section B average: 12.8 out of 20.
How to read it
Literature's version of the FR-02 pattern. The two examined skills, using a passage as evidence for a whole-text interpretation and analysing how language makes meaning, are the two that most responses do only partly, and the assessors say so in almost the same words as the English report. The 'continued misuse' of juxtapose is a small, teachable, recurring fault. This is the first VCE Literature reading the register has; LIT-13 now holds the enrolment base (about 4,100 students a year).
What would weaken it
Fetched directly as the .docx and extracted locally from the document XML, the most reliable route available; the mark-distribution rows were read against the marks header and the two averages (3.4 and 12.8) confirmed in the main session. Assessor commentary is qualitative and undated within the report; no proportion of responses is attached to any fault. Victoria, 2025 examination only.
NAPLAN 2026, Year 9 reading: the parental-education gap (4% needing additional support against 30%) and the remoteness gap (9% against 53%) dwarf the gender gap the register already holds
What it found
Read from ACARA's own published dataset (Calendar year 2026, Year 9, Reading, Australia), values rounded to one decimal. By parental education: students whose parents' highest education is a bachelor degree, 4.1% Needs additional support and 78.7% Strong or above (143,737 participants); parents' highest education Year 11, 30.0% Needs additional support and 29.2% Strong or above (16,890). By remoteness: major cities 8.9% Needs additional support and 66.1% Strong or above (225,252); very remote 52.8% and 21.5% (1,068 participants, 52.3% participation). By language background: LBOTE 65.1% Strong or above (87,291) against non-LBOTE 60.8% (198,140). The writing cells run the same way: Year 11 parental education 32.1% Needs additional support against 5.4% for bachelor; very remote 58.6% against 9.9% in major cities.
How to read it
The frontier's NAPLAN 2026 items (FR-01, FR-08, FR-09, FR-11) are national and gender figures. These are the equity cells, and they order the same way RDG-15 found in 2024: the largest gaps are by parental education and remoteness, and LBOTE students sit above non-LBOTE, not below. For Argo, which serves metropolitan families who mostly hold degrees, the useful reading is that its typical student is in the 4%-needs-support group, and the site's language about 'falling behind' should be scaled to that.
What would weaken it
Participation differs sharply across cells (76.9% for Year 11 parental education, 52.3% for very remote, 94.9% for bachelor), and non-participation is not random, so the gaps are understated, not overstated. Parental education is parent-reported at enrolment and missing for some students. Very remote is 1,068 students nationally. Read from the dataset by the collecting agent and re-read cell by cell in the main session; the file was the current one on 2026-09-11.
Victoria's Year 9 sits above the nation in NAPLAN 2026: 66.4% Strong or above in both reading and writing, against 62.0% and 59.9% nationally
What it found
From the same ACARA dataset (2026, Year 9, all students). Reading: Victoria 48.0% Strong, 18.4% Exceeding, 66.4% Strong or above, 7.9% Needs additional support, 76,801 participants of 82,955 enrolled; Australia 45.7%, 16.3%, 62.0%, 11.1%, 297,304 participants. Writing: Victoria 41.0% Strong, 25.4% Exceeding, 66.4% Strong or above, 8.8% Needs additional support, 77,265 participants; Australia 38.2%, 21.6%, 59.9%, 12.8%, 299,126 participants. The national writing Strong figure (38.2%) is the same cell FR-01 reported from ACARA's summary, which is the cross-check that the file and the reading of it agree.
How to read it
The first Victorian cells on the frontier. Victoria's advantage is larger in writing (6.5 points on Strong or above) than reading (4.4), and its Needs-additional-support share in writing is roughly two thirds of the national one. That is the honest frame for Argo's Victorian pages: Year 9 writing is one of the nation's two weakest cells (FR-01, TST-14) and Victoria is the least weak part of it. The identical 66.4% in both domains is coincidence; the components differ and each re-derives from Strong plus Exceeding.
What would weaken it
Victoria's participation (92.6% reading, 93.1% writing) is above the national rate (89.8%, 90.3%), which slightly flatters the comparison since non-participants score lower on average. Proficiency shares, not scale scores; ACARA's own caveat on the 2026 writing results after the platform outage (FR-11) applies to both rows. Values are one-decimal roundings of the dataset cells.
A definitive trial of reciprocal reading in 20 English high schools found no effect on comprehension, after two smaller trials of the same programme had
What it found
“The trial was conducted with students in 20 English high schools. Process data found the technique implementable by schools. Test data (n=733) found no significant differences between treatment and control groups, although the observed changes in treatment group students was of similar magnitude to those detected in a previous study.” Students aged 11 to 12 were individually randomised: “In total 782 students from 20 recruited schools were randomised, 388 students to the treatment group and 394 students to the control group”, with 369 and 364 analysed on the primary outcome. “ANCOVA indicated no significant differences in gains between control and intervention groups (F(1730)=0.12, p=0.73. The effect size intervention to control was -0.05”. The paper contrasts this with the programme's earlier record: an efficacy trial in 98 primary schools with effects of +0.14 to +0.18, and a Phase 2 trial in 14 high schools with +0.13 to +0.25.
How to read it
A well-run null on a structured small-group comprehension intervention for the first year of secondary school, the kind of programme TUT-16 (small-group tuition) rates positively on average. The sequence is the instructive part: two smaller trials positive, the larger definitive one null, which is the pattern replication work keeps finding (compare TUT-06 on Bloom). For Argo the reading is that 'evidence-based' comprehension routines can carry positive labels from early trials that the decisive trial does not confirm.
What would weaken it
The trial ran through COVID-19 disruption, which the authors say may have produced atypical findings; a null under those conditions is weaker evidence than a null in a normal year. England, Year 7, one programme (FFT Literacy's Reciprocal Reading); the register holds no reciprocal-reading entry to weigh it against. Open-access author PDF read in full; one of four quoted passages was matched in part against the extracted text because of a table break. PROMOTED TO RDG-24 on 2026-09-16. This item's figures were CONFIRMED accurate against the repository PDF, and the confirmation is worth recording because verifying it nearly produced a false correction: the paper reports at least eight ANCOVA results, one per outcome measure, each phrased in nearly identical words, so checking a quoted figure against the first matching sentence returns a different F statistic and effect size and looks exactly like an error here. It was not. Where a paper repeats a results sentence per measure, match on the measure, not on the sentence pattern. Reading all eight also made the finding stronger than the single result quoted above: effects run from -0.09 to +0.13 across two test batteries with none significant, and the on-treatment sensitivity analysis is null as well.
Victoria's 'nearly 60 per cent' Year 1 phonics figure now traces to a Premier's press release, not a statistical report, and no denominator has been published
What it found
The release states that “94 per cent of Victorian government schools” chose to use the check ahead of its mandatory rollout in 2026, that it “has already assessed the literacy skills of nine in 10 students in this year level since it was made optional in Term Three”, and that “Victorian students have also showed strong results, with nearly 60 per cent of students achieving the expected knowledge level two years ahead of the full rollout of phonics education in 2027”. It also reports an average completion time of 6.5 minutes. No student count, no cut score and no results table are given.
How to read it
This resolves the provenance question the coverage note left open on 2026-09-05: the circulating ~60% figure is a government figure, but a press-release one. It is now usable with that label. Against NSW's published series (ECD-14: 64% in 2025 on a stated 28-of-40 cut score, 61,964 students), Victoria's 2025 pilot looks similar, but the comparison is soft on both the cut score and the population, since schools opted in. The 2026 mandatory round is the one to wait for.
What would weaken it
A media release, tier 3 for the number even though the publisher is government: no denominator, no threshold, no methodology, and 'nearly 60 per cent' is not a figure. Opt-in pilot, so the 94% of schools is not 94% of students and the assessed cohort is not the state cohort. No VCAA or Department statistical publication of the pilot results was located; if one appears, it supersedes this.
NSW's new English Advanced syllabus, first examined in 2027, drops the mandated multimodal task, drops student-chosen related texts, and sets four prescribed texts including Shakespeare
What it found
The department's implementation page states that the syllabus “will be taught from 2026 and will be assessed in the 2027 HSC”, and lists the changes: “a reduction in the number of outcomes from 9 to 6 in Years 11 and 12”; “focus areas which replace modules and focus area content which replaces module descriptions”; “the removal of a mandated multimodal presentation as part of the school-based Year 11 and 12 – English Advanced assessment requirements”; “the removal of the requirement to integrate student selected related material into the ‘Texts and human experiences’ assessment task”; and “a change to text requirements in Year 12 with students now required to study 4 prescribed texts, with ONE drawn from each of the following categories” (prose fiction, poetry, drama or nonfiction or film or media), with “at least one of the texts selected must be authored by Shakespeare”. Text requirements across Stage 6 now mandate Australian authors including Aboriginal and Torres Strait Islander authors.
How to read it
The structural companion to FR-16 (NESA's cap on take-home assessment). Together they move NSW senior English toward fewer, more prescribed, more examined tasks: no self-chosen related text, no multimodal, four set texts. For Argo's NSW-facing material the concrete change is that the 2027 cohort, now in Year 10, prepares for a course with less student choice and more close study, which is the VCE shape. The syllabus itself is the primary source; this page is the department's summary of it.
What would weaken it
Read from the NSW Department of Education's syllabus-information page via a text proxy (the NESA curriculum site is a JavaScript application that serves navigation only to automated fetches), so this is a departmental summary of the syllabus rather than the syllabus text. Assessment weightings were not found and are not stated here. Year 11 and 12 English Advanced only; English Standard's parallel changes were not read.
NSW publishes same-cohort NAPLAN gains in grammar and spelling — and names writing and secondary grammar as the domains that did not improve
What it found
NSW reports gains tracked against the SAME COHORT's earlier result rather than against a different cohort at the same year level, which is the comparison NAPLAN reporting almost never makes public. In the release's own words: "Year 5 grammar and punctuation is up 8.9 percentage points in strong and exceeding compared to the same cohort's Year 3 results in 2024"; "Year 5 spelling is up 8.6 percentage points in strong and exceeding compared to the same cohort's Year 3 results in 2024"; "Year 5 numeracy is up 6.8 percentage points in strong and exceeding compared to the same cohort's Year 3 results in 2024"; and "Year 7 spelling is up 4.9 percentage points in strong and exceeding compared to the same cohort's Year 5 results in 2024". The same release names where nothing improved: "This includes writing for Years 5, 7 and 9 and grammar and punctuation for Years 7 and 9." It also commits to "mandatory writing professional development for all NSW public schools in 2027".
How to read it
The shape of the result matters more than its size. Every gain NSW can point to is in primary or early secondary and in the CODIFIABLE strands — spelling, grammar and punctuation, numeracy — which are exactly what explicit teaching of rules is good at. The domains it lists as still needing work are writing at Years 5, 7 and 9, and grammar and punctuation at Years 7 and 9: composition, and the conventions once they have to serve composition. That is the same fault line the register already holds from other sources — the gap that widens with age is in writing, not in the mechanics measured on their own — and here it appears inside a government's account of its own successful reform. A jurisdiction crediting explicit teaching for primary gains, while separately funding mandatory writing PD because writing did not move, is evidence that the two are different problems rather than one.
What would weaken it
NSW only: not Victoria, where this practice operates, and not national. A ministerial release, not a technical report — no methodology, no confidence intervals, no underlying tables, and the publisher is the interested party reporting on its own reform. The figures are movements in the proportion at 'strong or exceeding', not mean scores, so they are sensitive to where the proficiency cut falls; a small shift in a distribution bunched near the cut moves this number more than it moves achievement. Nothing here establishes that explicit teaching CAUSED the gains — no comparison group is offered. And the 2026 writing results carry ACARA's own disruption caution (FR-11), which is a further reason not to read the flat writing result as a clean measurement.
ACARA's own commentary says NAPLAN has not moved nationally in three years — 163 of 165 subgroup comparisons negligible since 2023
What it found
ACARA reports the same conclusion separately for each year level, in each case as a national and state-level statement: “Nationally and across all reported states and territories, in all domains, the differences in average scores since 2025 were negligible”, and the same sentence again for “since 2023”. The subgroup counts differ by year level and are given here rather than generalised. For 2025 to 2026: Year 3 “164 of 165 comparisons ... showed a negligible difference”, and Years 5, 7 and 9 each “every one of the 165 comparisons”. For 2023 to 2026: Year 3 163 of 165, Year 5 162 of 165, Year 7 160 of 165 and Year 9 164 of 165. ACARA discounts every exception in the same terms — they “were in subgroups with very few students, and should not be considered as showing a meaningful pattern of change” — including Year 7's five, of which “4 were below and significantly different, and one was above and significantly different”. Separately, participation is falling and is severely uneven, and this too is reported per year level. Year 3 participation in at least one test fell from 96.4% to 95.4%, and “was 58.2% in Tasmania, 82.7% in the Northern Territory, and ranged from 94.2% to 97.3% in the remaining states and territories”. Tasmania's rate differs by year level: 58.2% at Year 3, 58.7% at Year 5, 69.9% at Year 7 and 67.7% at Year 9. ACARA does not report state achievement comparisons for Tasmania at all, “due to low participation”.
How to read it
This is the measuring authority saying, in its own commentary and at every year level separately, that the national numbers have not meaningfully moved since 2023. It matters most for what it forbids: any page on this site that reads an annual movement as improvement or decline is reading noise, and TST-14's framing of 2026 as a distinct picture needs to sit inside a three-year flatness the authority itself asserts. It also complicates the tutoring market's standard pitch in both directions — there is no national crisis worsening year on year, and equally no evidence that anything done nationally since 2023 has shifted the average. The participation figures are the sharper find and were not what the sweep was looking for. Tasmania at 58.2% means roughly two in five Tasmanian YEAR 3 students sat no NAPLAN test at all, which is why ACARA withholds Tasmanian comparisons — a fact that should be stated wherever this site presents a state-by-state table, because a reader will otherwise assume the absence is an oversight.
What would weaken it
“Negligible” is ACARA's own effect-size classification, not a claim that nothing changed anywhere: it is a statement about AVERAGE scores at national and state level, and averages can hold steady while the distribution beneath them spreads — which is exactly what this site's writing-divergence analysis documents. Flat averages and a widening spread are not in tension. ⚠️ The figures here are PER YEAR LEVEL and must not be quoted as a single report-wide number. An earlier version of this item did exactly that, presenting Year 3's counts and Year 3's 58.2% Tasmanian participation as though they were the report's general finding; the counts differ at every year level and Tasmanian participation ranges from 58.2% at Year 3 to 69.9% at Year 7. Corrected 2026-09-11 after an adversarial re-read of the source. The 2026 writing results at Years 5, 7 and 9 separately carry ACARA's platform-outage caution (FR-11), which this item does not supersede.
Corrected
Corrected the same day it was written. The item quoted Year 3's subgroup-comparison counts (164 of 165 since 2025, 163 of 165 since 2023) and Year 3's Tasmanian participation (58.2%) accurately, but presented them as the report's general finding. ACARA reports all of these separately for each year level: the 2023-to-2026 counts are 163, 162, 160 and 164 of 165 at Years 3, 5, 7 and 9, and Tasmanian participation ranges from 58.2% at Year 3 to 69.9% at Year 7. The headline conclusion is unchanged and is if anything stronger — every year level carries the same 'negligible' statement, and Years 5, 7 and 9 are 165 of 165 for 2025 to 2026 — but a figure scoped to one year level was being published as a national one, which is the exact defect this register's protocol exists to prevent.
AERO's own pilot of its School Writing Instruction Framework reports no figures — three schools, one year, no control
What it found
AERO ran a year-long pilot of its School Writing Instruction Framework (SWIF) with a Year 8 cohort in three secondary schools in Western Australia, South Australia and New South Wales, spanning ICSEA 941 to 1,168. On outcomes it reports, in full: “When we examined the NAPLAN writing data from all 3 schools, there was an increase in the percentages of students improving in the writing criteria of sentence structure and vocabulary from Year 7 to Year 9. Two out of the 3 schools also improved in the criterion of cohesion.” No effect size, no percentage, no comparison group, and no figure of any kind appears anywhere in the nine-page document. AERO states its own limit: “These key actions were identified through working with 3 schools. Future research with a larger number of schools in different jurisdictions could help strengthen and refine the findings.” It also concedes the wider position that motivates the framework: “while guidance for teachers on reading instruction is readily available, guidance on writing instruction is limited, and the quality of research on effective ways to teach writing is variable.”
How to read it
The value here is the pairing, not the pilot. LRX-05 already holds a July 2025 AARE critique arguing SWIF rests on a narrow subset of evidence — applied linguistics and genre theory — while disregarding AERO's own 2022 position that ‘no single pedagogy adequately addresses all aspects’ of writing. This document is AERO's own evaluation of that same framework, and it is the natural place to answer that criticism with results. It reports none. Three schools, no control, no numbers, and an explicit note that generalisation needs more schools. That does not make SWIF wrong. It makes the current Australian position on whole-school writing instruction exactly what AERO itself says in its opening paragraph: guidance is limited and research quality is variable. For this site the use is defensive rather than promotional — it is evidence about how thin the evidence is, which is the honest answer when a parent or a school asks whether the national framework is proven. It also converges with LAN-12 (Andrews et al.: ‘insufficient quality of research to prove the case with either approach’) from a completely different direction: the UK systematic-review literature and the Australian national body arrive at the same shrug.
What would weaken it
This is a practice-guide case study, not a study: AERO is evaluating its own framework, there is no control group, no pre-registration and no statistical test, and the sole outcome sentence is directional prose. It should never be cited as evidence that SWIF does NOT work — absence of reported figures is not a null result, and a pilot designed to refine a framework is not designed to measure it. The NAPLAN comparison it does mention is Year 7 to Year 9 within the pilot cohort, which is a period over which writing scores move for many reasons. Registered as a reading about the state of the evidence, not about the framework's effect.
Australian teachers estimate losing about a third of lesson time to misbehaviour — while most still rate behaviour as good
What it found
A survey of a representative sample of 377 Australian school teachers, adapted from England's National Behaviour Survey. In the report's own words: "Teachers reported losing an average of approximately nine minutes per 30 minutes of teaching time — almost one-third of lesson time — due to student misbehaviour. This is equivalent to 12 weeks of lost learning per year." On frequency: "Four in 10 teachers (40%) said misbehaviour interrupted teaching in most or all lessons during the previous week, while a further 39% said this occurred in some lessons." On type: "57% of teachers reported students talking when they were not supposed to in most or all lessons, while 45% reported students calling out in most or all lessons." Only 3% reported losing no teaching time to misbehaviour.
How to read it
The number worth carrying is not the nine minutes but the contradiction the report itself names: the same teachers who estimate losing a third of the lesson also rate behaviour positively — "56% rated student behaviour as good or very good". Both cannot be a complete description of the same classroom, and the report presents them together rather than resolving them. The honest reading is that low-level disruption has been normalised to the point where it is not experienced as a behaviour problem, which is a different claim from 'classrooms are in crisis' and a more defensible one. For a one-to-one practice this bears on why a student who is compliant and attentive in class can still arrive without the instruction the lesson nominally delivered: attendance is not the same as instructional time received.
What would weaken it
Self-reported teacher estimates of time lost, not observation or timing — people are poor at estimating elapsed time, and the direction of that error is unknown. n=377, and the instrument is adapted from an English survey rather than built for Australian conditions. It measures perceived disruption only: the report establishes no link to any literacy, writing or NAPLAN outcome, and none should be asserted from it. The publisher is a think tank with a stated policy orientation, not a peer-reviewed or government source, and the paper's own framing ("crisis") is stronger than the internal contradiction it reports supports.
Umbrella review of gifted education: enrichment's academic effect more than halves under publication-bias correction, and no ability-grouping meta-analysis cleared the evidence bar
What it found
An umbrella review of 17 meta-analyses covering 397 unique primary studies, in the journal's own words. On enrichment, effects are reported as "Large (g = 1.14)" for cognitive outcomes, "Moderate (g = 0.68; 0.32 after trim-and-fill)" for academic outcomes, and "Small–moderate (g = 0.44–0.51)" for affective outcomes. Acceleration shows "positive academic effects together with neutral to slightly positive socio-emotional outcomes". On the third approach the review states plainly: "No ability-grouping meta-analysis met the predefined criteria for best-evidence prioritization." Quality of the underlying evidence, per AMSTAR 2: "One meta-analysis was classified as high quality, eight were classified as moderate quality, six as low quality, and two as critically low quality."
How to read it
Two things here cut against how selective and scholarship coaching is usually sold. First, enrichment's academic effect falls from g = 0.68 to 0.32 after trim-and-fill — a correction for publication bias more than halves it, so any provider quoting the headline figure is quoting the number the correction exists to discount. Second, ability grouping — streaming, the structural logic behind selective classes — has no meta-analysis that cleared the review's own best-evidence bar. That is not a finding that grouping fails; it is a finding that the evidence to support it is not there at the level this review would accept, which is a different and more awkward thing to say. Read alongside TST-02 and TST-03, the picture is consistent: the strongest claims in this market rest on the weakest evidence, and the defensible case for preparation is what the preparation itself teaches.
What would weaken it
An umbrella review inherits the quality of what it reviews, and by its own AMSTAR 2 assessment only one of the underlying meta-analyses is high quality while eight are low or critically low. The evidence base is international, not Australian, and nothing here speaks to Victorian selective entry specifically. 'Gifted education' as studied here means school-level programmes — acceleration, enrichment curricula, streamed classes — not private tutoring, so no inference about one-to-one preparation should be drawn from it in either direction. The absence of qualifying ability-grouping meta-analyses is an absence of admissible evidence, not evidence of absence of effect.
PISA 2025 retires the “stable since 2015” reading picture: Australia’s reading fell a statistically significant 7 points since 2022, and its decade-long trend is declining
What it found
PISA assesses nationally representative samples of 15-year-olds; all figures below are for that population. Table I.1 (“Snapshot of performance in science, reading, mathematics and computational problem-solving”) gives Australia’s PISA 2025 means as science 509, reading 491, mathematics 478 and computational problem solving 532, and its short-term changes from PISA 2022 as science +2, reading −7 and mathematics −10. The table’s note states that “values that are statistically significant are indicated in bold”: Australia’s reading and mathematics changes are bold, its science change is not. The OECD average is science 482, reading 461, mathematics 463 and computational problem solving 500, with changes of −3 (not bold), −14 (bold) and −9 (bold). Australia’s share of top performers — Level 5 or 6 in science, reading OR mathematics — is 18.2%, and its share of low performers — below Level 2 in science, reading AND mathematics — is 14.0%, against OECD averages of 11.9% and 19.7%. On the decade, Table I.2.9 reports an “average decennial trend”, defined in its note as “the average change, per 10-year period, between PISA 2015 and PISA 2025, calculated by a linear regression”, of +1 in science, −12 in reading and −16 in mathematics for Australia. The report’s own text places Australia among “countries/economies with declining decade-long trends” in both reading and mathematics, and says that in these systems the drop “was driven by low-achieving students (i.e. changes were statistically significant only for low-achieving students)”, naming Australia in both lists, with the result “a widening of the performance gap”. On gender, Table I.3 gives Australian girls 507 and boys 475 in reading, a difference of −31 shown in bold; mathematics girls 469 and boys 486, +16 bold; and science 509 and 509, a difference of 1, not bold. The OECD-average reading gender difference is −30. Australia carries no asterisk in these tables; the asterisk marks systems where “one or more PISA sampling standards were not met” and is applied to Canada, New Zealand, the Netherlands and the United States, not to Australia. Two findings quoted here are OECD-wide and not Australian: reading fell “across the OECD by 25 score points” between 2018 and 2025, “more than a year of learning”, in “three-quarters of education systems with comparable data”; and, on the same OECD average, the foreword reports a rise in “hasty readers” — defined there as “students who rush through texts and give quick but incorrect answers” — which Chapter 2 quantifies separately as an increase of “about five percentage points, on average, from almost 7% in 2018 to 11% in 2025”.
How to read it
This is the cycle that retires the closing clause of RDG-12. That entry is correctly attributed to PISA 2022, and its quoted source says “While the most recent trend is stable, at least since 2015” — which was true of the 2022 data. PISA 2025 is the first cycle able to test that clause, and it does not survive: the 2022→2025 reading fall is statistically significant, and the report classifies Australia’s decade-long reading trend as declining. RDG-12 is therefore not wrong, it is dated, and it is specifically the stability half that the new cycle removes; the decline half — the long-run fall against the early 2000s — is corroborated and extended rather than weakened. The significance structure is the part most likely to be overstated, so read it precisely: over the decade the decline was statistically significant ONLY among low-achieving students, high achievers did not change significantly, and the gap widened. That is a distributional finding, not a uniform national slide, and it is the more useful one here — the students falling further behind are the ones already behind. The result is also less dramatic than the headline framing once Australia is separated from the OECD. The “sharp decline” and “record low” framing in circulation is an OECD-wide statement. Australia’s reading fell 7 points while the OECD average fell 14, so Australia’s margin over the OECD average widened rather than narrowed — from 22 points in 2022 (WRI-16 records 498 against 476) to 30 in 2025 (491 against 461). Both are true simultaneously, and citing either alone misleads. One citation trap is worth naming, because the numbers look invitingly comparable and are not. Table I.1’s low-performer column counts students below Level 2 in science, reading AND mathematics — low in all three — and its top-performer column counts Level 5 or 6 in science, reading OR mathematics — top in at least one. Neither is commensurable with RDG-13’s reading-only figures (79% at Level 2 or above, 12% at Level 5 or higher in 2022), so Australia’s 14.0% and 18.2% must not be set against those as a later reading of the same measure. Finally, what this volume does NOT do: it does not re-report reading for enjoyment. The phrase appears once, in the foreword, inside an explicit hedge — “We cannot say for sure, but declining levels of reading for enjoyment and the shift towards digital reading … may be compromising the ability of students to engage with complex texts”. RDG-16, RDG-17, RDG-19 and LIT-05, which rest on reading-engagement data, are untouched by this volume and are deliberately not cited here.
What would weaken it
The significance marks in Tables I.1 and I.3 were recovered from the PDF’s embedded font data rather than read visually: the report says significant values are shown in bold, and boldness was detected programmatically. The method was checked against controls before being relied on — Türkiye’s POSITIVE changes (+18, +16, +8) are bold, so bold tracks significance and not sign, and Chinese Taipei’s −7 in reading is NOT bold while Australia’s identical −7 is, so it is not a magnitude threshold either. Table I.2.9 uses no bold at all — it flags non-significant values with a legend symbol that text extraction does not preserve — so the −12 and −16 decennial figures are given here as the report’s point estimates, and the claim that Australia’s decade trend is declining rests on the report’s prose naming Australia, not on those cells. This is Volume I only, read at the OECD international level: ACER’s Australian national report (De Bortoli, Underwood & Bertling, published the same day at https://research.acer.edu.au/ozpisa/67/) sits behind a Cloudflare challenge and was NOT read, and Australia’s student and school sample counts live in a StatLink spreadsheet outside the PDF and were not seen, so no sample size is stated here — the counts circulating in search results remain unverified. Comparisons across cycles carry link error, which the report describes as “uncertainty around scale values” independent of sample size (2.9 score points for science between 2022 and 2025); and the scales have “means around 500 score points, standard deviations around 100 score points”, so a 7-point move is statistically significant but small. The report’s own one-year-of-learning benchmark of about 20 score points is offered with the warning that it “should be interpreted cautiously and not applied mechanically”. Computational problem solving is new in 2025, measured as part of the innovative domain “learning in the digital world”, so Australia’s 532 has no earlier cycle to be compared against.
ACARA’s national sample assessment asked Australian students about generative AI for the first time: 60% of Year 10 and 31% of Year 6 use it to generate written content for school at least monthly
What it found
NAP–ICT Literacy is one of three national sample assessments ACARA runs for Education Ministers; the 2025 cycle is its seventh since 2005 and was conducted over four weeks in Term 2, 2025 on nationally representative samples of Year 6 and Year 10 students — those two year levels only. The achieved sample was 5,498 Year 6 students from 328 schools and 4,753 Year 10 students from 313 schools, 641 schools in all, with national response rates of “90% for Year 6 and 82% for Year 10, after accounting for school replacements and applying weighting adjustments”. Students completed a questionnaire after the assessment, and for the first time it asked about generative AI: the report’s own glossary defines AI tools as “generative AI (e.g. ChatGPT, Microsoft Copilot) used by students for study or content creation, which is a new focus in the 2025 cycle”. Table 7.1 asks how often students use each tool “for school-related purposes”. For “Artificial intelligence tools for generating written content (e.g. ChatGPT, Microsoft Copilot, Gemini, Claude)”, Year 10 students report 34 (±2.3) at least once a week, 26 (±1.8) at least once a month but not every week, 19 (±1.5) less than once a month and 20 (±2.0) never — 60 (±2.5) at least monthly. The Year 6 figures are 15 (±1.3), 17 (±1.5), 19 (±1.6) and 50 (±2.4) never — 31 (±2.2) at least monthly. AI image generation is used at least monthly by 31 (±2.1) of Year 6 and 39 (±2.3) of Year 10 students. Table 6.1 asks separately about “Use artificial intelligence tools (e.g. ChatGPT, Microsoft Copilot) to assist with your school work”, banded as rarely (“less than once a month or never”), occasionally (“between a few times each week and once a month”) and frequently (“almost every day or more”): at school, Year 6 splits 72/19/9 and Year 10 splits 38/36/26; outside school, Year 6 splits 69/20/11 and Year 10 splits 43/34/22. The report states it as “1 in 4 Year 10 students (26%) reported frequent use of artificial intelligence (AI) tools at school, compared to just 1 in 10 Year 6 students (9%). Outside of school, 22% of Year 10 students and 11% of Year 6 students reported frequent use of AI tools.” Two further items, also new in 2025, ask what students have learnt at school: 50 (±2.1) of Year 6 and 66 (±2.0) of Year 10 students say they have learnt “how to judge whether information produced with artificial intelligence tools … is accurate”, and 65 (±2.0) and 78 (±1.9) say they have learnt “understanding when it is appropriate to use artificial intelligence generated content for schoolwork”. On the assessment itself, 50% (±2.1) of Year 6 students met the proficient standard of 409 scale points, down from 55% (±2.4) in 2022 though “not significantly different from the 2017 and 2014 cycles”, with a mean of 398 (±4.8) against 414 (±6.0) in 2022; 37% (±2.8) of Year 10 students met the Year 10 proficient standard of 529, down from 46% (±3.2) and “the lowest proportion of students achieving the proficient standard since the assessment began in 2005”, with a mean of 482 (±7.7) against 503 (±7.2).
How to read it
This is a nationally representative, government-run measurement of Australian school students’ generative-AI use — the first cycle of this assessment to ask — and its value here is a tier change rather than a new number. AIE-13 is currently the register’s Australian AI-use claim and it is a tier-3 industry survey whose own entry notes record that fieldwork dates and sampling method were never confirmed; NAP–ICTL is a national sample assessment with a documented two-stage stratified design, published confidence intervals and response rates above the OECD and IEA thresholds. AIE-02 and AIE-03 are Pew surveys of US teens that the register itself files as international comparators of low Australian relevance. This report is the Australian datapoint those three have been standing in for. It does not correct any of them, and it must not be presented as reconciling with them: the instruments ask different questions of different populations. Elevate asked high-school students whether they use AI “at least a few times per week” and got three-quarters; ACARA asked Year 6 and Year 10 students how often they use AI for school-related purposes and got near-daily use at school of 9% and 26%. No arithmetic connects those, and neither refutes the other. The finding that matters most for a writing-focused site is the narrowest one: the Table 7.1 item is specifically about generating WRITTEN content for school purposes, and on it a third of Year 10 students (34%) report at least weekly use while a fifth (20%) report never using it at all. That is a level, not a rise — the item did not exist in 2022 or 2017, so the report supplies no trend and nothing here licenses “AI use is growing”. Two further limits are worth stating because the source is tempting to over-read. First, the report measures reported use, not misuse: it asks nothing about cheating, disclosure or whether the work was passed off as the student’s own, and it reports no achievement association for the AI items specifically. That absence needs care rather than relief: the AI question is one component of a broader “use of study utilities” index, and the index does carry achievement relationships — at Year 6, students with lower ICT literacy report more frequent use of study utilities at school — but that is a property of the index, not of AI use, and must not be reported as one. The achievement associations the report does state at item level are for productivity applications (positive) and specialist applications such as concept mapping and simulation (negative). Second, the ICT literacy decline in the same report is not evidence of an AI effect and the report never suggests it is — the AI items are new, so no before-and-after exists, and the fall is confounded by the timing change described in the caveat. One citation warning: use the report, not ACARA’s media release of 27 May 2026, which states response rates of 87% and 78% against the report’s 90% and 82%.
What would weaken it
The largest qualification is ACARA’s own and is easy to lose. The 2025 cycle was the first administered in May rather than October, which produced a younger modal age, and every trend table in the report carries the footnote “Comparison between 2025 and previous cycles should be interpreted with caution due to the earlier testing window.” Any statement that ICT literacy fell since 2022 inherits that caution, and the achievement figures here are quoted only as context for the questionnaire, not as a trend claim. The AI figures themselves are self-reported by students immediately after sitting a test, not observed behaviour, and self-reported technology use is the weakest part of an otherwise well-controlled instrument. The population is Year 6 and Year 10 only: nothing here covers Years 11 and 12, so it says nothing directly about VCE or senior English cohorts. Because the AI items are new in 2025 there is no baseline, so the report can establish prevalence and nothing about change. The document read was Version 3.1: its PDF metadata records “Initial Version 3.0 published May 2026” and “Updated Version 3.1 published September 2026” with a modification date of 7 September 2026, while its filename is dated 17 July 2026 and the document itself carries no visible amendment note, so what changed between versions is not stated anywhere in it and the figures above are from the file fetched on 2026-09-09. The NAP–ICT Literacy 2025 Technical Report, which holds the response-rate and weighting detail this public report defers to, was not read.
VCAA is consulting on extending the VCE English/EAL and Literature text cycle — feedback closes 11 September 2026
What it found
VCAA issued a consultation survey on 20 August 2026 proposing to extend the text cycle for two lists: the “VCE English and EAL text list: List 1 only” and the “VCE Literature text list (complete list)”. Responses are invited from “Principals/Directors, VCE Coordinators, and teachers of VCE English, EAL and Literature”, including retired teachers, and the notice states “Friday 11 September” as the last day to submit feedback. The notice does not state the current cycle length or the proposed new one — only that an extension is proposed. Queries go to vcaa.vce.curriculum@education.vic.gov.au.
How to read it
This is the missing context for a question the register has been circling. VCE-35 through VCE-40 document a text-tenure policy VCAA has already rewritten twice, and the 2028 List 1 has still not appeared on the schedule the 2027 list set. The consultation explains why: the rotation cadence itself is under review, so the absence of a 2028 list is a live policy question rather than a delay. Two registered claims are affected in their timing rather than their truth — VCE-37 predicts most of the 16 List 2 texts at tenure 4 leave at once, and VCE-38 says the tenure paragraph keeps being rewritten. A longer cycle would push the first prediction's test date out, and would itself be a third rewrite, confirming the second.
What would weaken it
A consultation is not a decision. Nothing has changed yet, the notice gives neither the current nor the proposed cycle length, and VCAA may not proceed at all. Any advice to students or schools about future text rotation must say this is a proposal under consultation, not a published change. The outcome should be re-checked after 11 September 2026. RE-CHECKED 2026-09-16, which is the check this item itself asked for ('The outcome should be re-checked after 11 September 2026'). THE CONSULTATION HAS CLOSED AND NO OUTCOME IS PUBLISHED. The full 2026 Notices to Schools index was enumerated rather than searched: the most recent notice is 94 (3 September 2026), and nothing after Notice 83 concerns the text cycle. So five days past the closing date VCAA has announced no decision, and the proposal's terms remain unstated -- the notice never gave the current cycle length or the proposed one, and still does not. SUPERSEDED FOR REGISTRATION by VCE-51, which holds the notice itself as a register entry; this item remains useful only as the standing re-check, and the next one should start from 2026-09-16 and from Notice 94 rather than re-enumerating the year. Also confirmed already known: Notice 80 (6 August) is an ISBN edition swap on the text list, not a cycle change.
ACARA attaches a caution to the 2026 NAPLAN writing results after a platform outage hit the Year 5, 7 and 9 writing tests
What it found
ACARA reports that “a technical issue affecting the assessment platform impacted many students in Years 5, 7 and 9 who sat their writing test on 11 March 2026”. It judges that the national writing results were minimally impacted and can validly be reported, but adds that “ACARA acknowledges that the effects of disruptions will have varied among students, schools and subgroups” and that “caution should be applied in interpreting writing results in 2026”. The affected year levels are exactly the three that sit the online writing test.
How to read it
This qualifies the frontier's own 2026 writing findings rather than adding a new one, which is why it matters more than its size suggests. FR-01 reports Year 9 writing at 38.2% Strong and FR-09 reports a writing gender gap of roughly sixteen points at Year 9 — both drawn from the 2026 collection, and both previously carrying caveats about proficiency-level cut points and participation, but none about the disruption. The uneven part is the problem: ACARA says the effect varied by student, school and subgroup, so a national average holding steady does not license a subgroup comparison. A gender gap is precisely a subgroup comparison. The caveats on FR-01 and FR-09 have been amended to carry this.
What would weaken it
ACARA's own judgement is that the national writing results remain valid to report, so this is a caution about interpretation, not a retraction — quoting the 2026 writing figures is legitimate provided the disruption is stated alongside. What is NOT established is the size or direction of any bias: ACARA does not quantify the variation by subgroup, so this cannot be used to argue a particular figure is overstated or understated. Reading and numeracy are unaffected.
Australian Curriculum, Assessment and Reporting Authority (ACARA), 2026. NAPLAN national results
AI detectors flag honest light editing more often than they catch deliberately disguised AI text
What it found
A controlled study of published English abstracts across four domains, comparing 2013-2015 with 2023-2025 text, put commercial AI detectors through three editing conditions. The authors report that “light ‘refine abstract only’ edits, a proxy for guideline-compliant AI assistance, are flagged at 38 to 80%”, while “unmodified 2023 to 2025 originals are flagged at 9 to 15%”. Against text passed through commercial humanising software the detectors collapse: “after Undetectable AI humanization, evasion is near-total: fewer than 4% of AI-labeled rewrites remain flagged (post-humanization detection rate <4%; FNR >96%)”. Their framing of the policy consequence is that detectors “cannot distinguish AI editing from full LLM drafts and may treat both as misconduct”.
How to read it
The asymmetry is the finding, and it inverts what a detector is bought to do. A student who used AI the way most school guidelines now permit — to tidy prose they wrote — is flagged at 38-80%. A student who used AI to write the thing and then ran it through a humaniser is flagged at under 4%. So the tool's errors do not fall randomly: they fall on the compliant student and spare the one it exists to catch. Roughly one in ten unmodified human abstracts was flagged too, against vendor claims of under 1%. For a tutor this is directly actionable — a detector score is not evidence, and a student accused on one has a substantive answer. It also puts a number against AIE-06: Turnitin's own guidance claims a false-positive rate below 1%, and this study measures 9-15% on unmodified human text across the detectors it tested, which is the same order of gap AIE-05 found for non-native writers.
What would weaken it
The corpus is published academic abstracts, not school-age student essays; the register and vocabulary of a Year 10 response differ from a chemistry abstract, and the false-positive rate on student writing is not established by this study. It is an arXiv preprint, not yet peer-reviewed. Detector vendors update models frequently, so the specific percentages will age quickly even if the asymmetry persists — cite the shape of the finding rather than the numbers if it is used after a few months.
The “handwriting builds brain connectivity” EEG study has a published methodological critique — and it measured no learning at all
What it found
A peer-reviewed commentary on the van der Weel & van der Meer (2024) EEG study that HND-04 records. Its central objection is that the study measured no learning: “the protocol did not include any learning. Participants repeatedly wrote well-known words without any requirement for memory encoding.” It also reports that “participants were instructed to type using only their right index finger, making the typing condition quite different from typical typing”, that there was an “absence of visual feedback on the screen during typing” but not during handwriting, that “participants' typing skills were neither assessed nor controlled”, and that “behavioral measures are not reported. Establishing a solid behavioral pattern is usually a requirement before interpreting neurophysiological measures.”
How to read it
This is the highest-salience claim in the handwriting domain — the study behind the “handwriting is better for your brain” coverage that ran in Scientific American and NPR through 2024, and the one a parent is most likely to raise. HND-04 states its finding accurately and carried no caveat at all. The critique does not allege the connectivity result is false; it argues the design cannot license the conclusion everyone drew from it, because the comparison condition was one-finger typing without screen feedback and nobody learned anything or was tested on anything. The register's own framing should therefore be that this is an interesting neural signal, not evidence of a learning benefit — which is a narrower claim than the media one, and the defensible one. HND-04's notes have been amended to carry this.
What would weaken it
A commentary is not data: it is a methodological argument, and no independent lab has attempted to replicate the EEG paradigm and failed. The connectivity differences as reported may well be real. What is contested is only the inference from them to classroom learning — so this should not be cited as evidence that handwriting does NOT help learning, which it does not show. The separate evidence for handwriting in early letter learning (HND-05, HND-06) rests on different studies and is untouched by this.
In a real primary classroom, retrieval practice worked on connected prose and spacing on its own did nothing
What it found
Thirty-three Italian children aged 10-11, in two intact 5th-grade classes, studied connected history prose (“The Ancient Greece”, “The foundation of Rome”) under testing versus re-reading, at one-day versus four-day intervals. Testing beat re-reading substantially — 67.0% against 48.6% accuracy, F(1, 25) = 41.42, p < 0.001 — and held when strategy was isolated, 67.7% against 41.3%, F(1, 24) = 8.2, p < 0.01. The phase that varied the interval WITHOUT retrieval found nothing at all: “the result was not significant, F < 1”, with 45.9% against 44.6%. The authors state that “most of the questions of the four final tests were literal, with only a few inferential ones”.
How to read it
Two things the `practice` domain needed, and it is the domain whose newest source was 2020 while holding eleven tier-1 entries. First, the effect is demonstrated on connected school prose in a real classroom rather than on word lists in a lab, which is the generalisation the register's Dunlosky-era entries (PRC-01, PRC-02) invite but do not themselves establish. Second, and more usefully, it is a boundary condition: spacing carried no weight on its own here — the benefit was retrieval's. PRC-02 rates distributed practice as high-utility on the same tier as practice testing, and this is one classroom study, not a refutation, but it is a reason to describe spacing as a way of arranging retrieval rather than as an independent lever.
What would weaken it
Thirty-three children in one Italian school, and the ceiling that matters most for this site: the tests were mostly literal recall. Nothing here shows retrieval practice improves inference, analysis or essay writing, which is the work Argo actually teaches. The gap is real and not just unsearched — a deliberate hunt for retrieval-practice studies using essay or literary-analysis outcomes found none. Treat this as evidence for how students should learn content, not for how they should learn to write about it. PROMOTED TO PRC-16 on 2026-09-16, and re-verified at source rather than carried across from this summary: every percentage was checked against the article's Table 1 and every F statistic against the Results text. Reading the full article added three limits this item did not record, all of them the authors' own, and all of them needed before the null is quoted. Both Phase 4 conditions used testing, so they say the testing itself 'may have masked' any additional benefit of the longer gap. Their 'massed' condition was a one-day gap rather than the zero gap most spacing studies use, so this compares a short delay with a longer one rather than spacing against cramming. And Phase 4 scored lower across both conditions because, on re-reading, the authors found their fourth text more expository than the rest, which 'may have obscured potential spacing effects'.
There is no evidence either way on whether studying whole books beats studying extracts
What it found
Reviewing the question for teachers, Shanahan states as one of five numbered conclusions that “there is no research showing that books are superior to excerpts or that excerpts are superior to books”, and adds that nor is there evidence that either does more to encourage independent reading. His own bottom line is not that extracts are adequate: it is that “there is no reason why schools cannot combine both excerpts and whole books in their English Language Arts instruction — fostering both depth and breadth”. A commenter on the post objects that the claim is “technically accurate and also incomplete and misleading” because no assessment exists that measures the ability to analyse a long work in the first place — which is a statement about the absence of the instrument, not a counter-finding.
How to read it
This is a finding about an absence, and it is the useful kind. Studying whole texts versus extracts is a live decision in Australian senior English — VCE English/EAL List 1 sets whole texts while List 2 works from mentor texts — and it is the sort of question where a tutoring site is tempted to assert what it believes. The evidence to assert it does not exist, in either direction. Registered so that the site's position on whole-text study rests on pedagogical reasoning it states openly, rather than on an implied research base that is not there. Note the asymmetry the commenter identifies: the absence is partly an artefact of measurement, because reading assessments use short passages, so the thing whole-book study would plausibly improve is not something the field measures.
What would weaken it
Expert commentary on a personal blog, not a systematic review — Shanahan is an eminent reading researcher and former IRA president, but this is his reading of the literature rather than a search anyone can audit. A null of this kind is also always provisional: it says nobody has run the study, not that no difference exists. It speaks to reading ability and independent-reading motivation, and says nothing about whether whole-text study serves other goals a literature course legitimately has.
NSW caps HSC take-home assessment at one task worth 15%, from Term 4 2026, and tells schools not to rely on AI detectors
What it found
NESA is changing the ACE Rules so that a Stage 6 course may set “no more than one take-home assessment task, worth a maximum weighting of 15%”. It takes effect in Term 4 2026 for HSC courses and Term 1 2027 for Preliminary courses, reaching students currently in Years 10 to 12. The stated principle is that “marks reflect the student's own knowledge, skills and understanding. Marks must not reflect a student's use of generative AI.” NESA also instructs that “schools should not rely on AI detection tools as the main safeguard against malpractice.”
How to read it
Two things, and the second is the one nobody will notice. The first is practical: about half an HSC result comes from school-based assessment, and from Term 4 the unsupervised essay is capped at a single task at 15% of that. A NSW student's English marks now rest almost entirely on work produced under supervision, which changes what practice is worth doing — drafting under time, in the room, rather than polishing at home. The second is that an assessment authority has told its schools in writing not to lean on AI detectors. That is FR-12's finding arriving as policy from an independent direction: FR-12 measures detectors flagging compliant editing at 38-80% while humanised text escapes at over 96%, and NESA reaches the same operational conclusion without citing any of it.
What would weaken it
NSW only — it binds the HSC and has no effect on VCE, where most of this site's students sit. It is also an interim measure: NESA says its Board will consider subject-specific advice for the 2028 HSC and beyond, so the settings may move again. The rule caps take-home tasks; it does not ban AI use, and NESA's separate guidance still expects students to acknowledge AI where used. PROMOTED TO AIE-18 on 2026-09-16, re-verified at nsw.gov.au. The re-read added four things this item lacked: the reference number NESA 29/26, the publication date of 31 August 2026, the exempt-course list (which includes English Extension 2 and every major-project or practical-examination course, so the cap is NOT uniform across English), and a fourth instruction stronger than the one quoted here -- schools are directed to 'Check your malpractice policy does not rely on the use of AI tools to detect malpractice.' REGISTERED FOR THE DETECTOR INSTRUCTION, not the take-home cap: the cap is NSW-only and interim, while the detector directive is the first assessment authority in the register to say in writing what /ai-in-schools-australia already argued from research and vendor behaviour.
The 2024 AEDC is out — 288,000 children, and the language and cognitive domain is one of its five measures
What it found
The sixth AEDC collection was released on 13 June 2025. The department states that “data was collected on over 288,000 children in their first year of full-time school, from 7,323 schools”, which “represents over 94% of eligible children nationally”, with teachers responding to “around 100 questions” across five domains — physical health and wellbeing, social competence, emotional maturity, “language and cognitive skills (school based)”, and communication skills and general knowledge. The AEDC “is conducted every 3 years to examine how children have developed by the time they start school”.
How to read it
Registered as a pointer rather than a finding, because that is what has actually been read. The AEDC is the largest early-childhood dataset in the country and its language-and-cognitive domain is directly relevant to the register's twelve `early-childhood` entries, none of which the frontier had ever examined. What matters for this site is the domain-level breakdown — how many children arrive at school developmentally vulnerable on language and cognition specifically — and that number is in the National Report, not in this announcement. Secondary sources report national vulnerability on two or more domains rising from 11.4% in 2021 to 12.5% in 2024, but I did not read that in a primary document and it is an all-domain aggregate that could be driven by physical health or emotional maturity rather than language.
What would weaken it
This item records the collection's existence, scale and scope — nothing about its results. The aedc.gov.au report page returns HTTP 403 to automated fetching, so the National Report was not opened. Do NOT cite a vulnerability percentage on the strength of this item. The AEDC is also a teacher-completed instrument administered in the first year of school, so it measures development at school entry, not the effect of anything.
A 155-school trial of grammar taught inside writing found no benefit to writing, and a small negative effect on the grammar test
What it found
A cluster-randomised trial in 155 English primary schools, 7,239 Year 6 pupils, of grammar taught explicitly but “embedded in the context of teaching about writing genres” rather than “teaching grammatical rules in the abstract” — six weeks of daily hour-long lessons replacing existing writing lessons, after three days of teacher training. Its two headline conclusions: “the project found no evidence that Grammar for Writing improves writing attainment for children in Year 6, as measured by the bespoke test”, and “the project found no evidence that Grammar for Writing improves reading, writing or grammar, punctuation and spelling (GPS) as measured by KS2 SATS. Indeed, it found a small, negative effect size (equivalent to one month less progress) for the GPS outcome.” Writing effect size -0.02 (95% CI -0.08 to 0.03); GPS -0.06, with the report stating that “pupils in schools that received the intervention did slightly worse in the KS2 GPS assessment than pupils in the control schools”. Pupils ever eligible for free school meals gained 0.05 (-0.03 to 0.13), which the report says “is not statistically significant”. Three padlocks out of five.
How to read it
This is the best evidence available on the question a writing tutor actually faces, and it is a null. Note precisely what was tested: not decontextualised grammar drills, which few defend, but the contextualised approach usually offered as the sophisticated alternative — grammar taught inside genre writing. Six weeks of it moved writing not at all. The authors' own reading of the negative GPS result is worth carrying rather than dropping: they suggest “this could be a result of the GPS assessment being a decontextualized assessment whereas” the teaching was contextualised, which is a fair explanation rather than an excuse, and it means the negative sign should not be read as grammar teaching harming grammar. It corroborates rather than corrects what the register already holds: WRI-14 records grammar instruction's small significant NEGATIVE effect on writing quality in Writing Next, WRI-12 records sentence-combining at 0.50, and both the writing and language-conventions synthesis essays already draw the distinction this trial illustrates — the same content taught as a rule to memorise versus a tool to build sentences with produces opposite results. What FR-18 adds is that the contextualised approach, the one usually offered as the answer to that critique, was itself tested at scale and did not work either.
What would weaken it
England, Year 6, one programme, six weeks — not Australian, not secondary, and not a test of every way grammar might be taught. The security rating is three padlocks of five because more than 28% of pupils who started were missing from the final analysis after schools withdrew and the outcome measure changed mid-trial, which is a real limitation on the null. A null result also cannot show that grammar teaching does nothing; it shows this programme at this dose did not register on these measures. Sentence-combining, which older syntheses treat more favourably, was not tested here.
A stricter re-analysis of the adolescent writing evidence: teaching writing still works, but the method matters more than the fact of teaching
What it found
Re-analysing an earlier meta-analysis after discarding every methodologically weaker study — single comparisons, attrition over 20%, no control for teacher effects, no reliable writing measure — the authors were left with 148 independent treatment/control comparisons, 1,076 effect sizes and 22,838 students in Grades 6 to 12. Teaching writing still produced a positive, statistically detectable effect on writing quality (ES = 0.38). Self-Regulated Strategy Development returned 0.84, the process approach 0.75, strategy instruction and peer assistance 0.59 each, transcription instruction 0.54, pre-writing 0.32 and feedback 0.30.
How to read it
The value here is in what was thrown away. This is the same evidence base as the 2023 review with the weaker studies removed, and the headline effect survives that pruning — which is the opposite of the usual pattern, where effects shrink once methodological standards are raised. The more useful figure for a teacher or a parent is not the 0.38 average but the spread beneath it: the strongest approach in the set is nearly three times the weakest. 'Teaching writing' is not one intervention, and an average across ten of them describes no actual classroom. Self-Regulated Strategy Development — explicitly teaching a strategy alongside the self-management to deploy it — is the standout, and it is a teachable procedure rather than a disposition.
What would weaken it
Effect sizes are pooled across very different outcome measures and the studies are overwhelmingly US-based, so transfer to Australian secondary classrooms and to NAPLAN-style assessed writing is untested here. Six of ten treatments were tested in four or more independent comparisons; the rest rest on thinner evidence than the tidy numbers suggest. An effect size on a writing-quality measure is not the same as a durable gain in a student's ordinary written work.
The first large AI-tutoring trial finds gains that match the same platform without the AI
What it found
A two-year cluster-randomised trial across 18 Tennessee middle schools gave randomly assigned students Khan Academy with its AI tutor Khanmigo, configured to coach rather than supply answers, inside existing daily remedial maths sessions. Assignment raised maths achievement by 1.3 national percentile ranks per term, about 0.06 to 0.08 standard deviations over a school year. The authors state plainly that 'these gains resemble those from Khan Academy practice without AI assistance'. Engagement was the constraint: 96% of students tried Khanmigo at least once, but the median student messaged it on only a third of the days they practised, and in only 17% of exercise sessions in which they made a mistake.
How to read it
This is the study to hold against the claim that a chatbot is a personal tutor. It is a genuine randomised trial at real scale, it was run with the tutor deliberately configured to coach, and the effect it found is indistinguishable from what the same practice platform delivers with no AI at all. The finding is not that AI tutoring failed; it is that access is not use. Students overwhelmingly opened it once and then, at exactly the moment a tutor earns its keep — having just made a mistake — did not turn to it in five sessions out of six. Any claim that an AI tutor will lift a struggling student has to clear this bar first: it must get used at the point of difficulty.
What would weaken it
An unreviewed NBER working paper, not yet peer-reviewed. It is a single subject (remedial mathematics), a single US district, and one product in one configuration, so it does not settle what a differently designed tool or a supervised deployment would do. The 0.14 SD figure for 'a full year of active participation' is an implied extrapolation from observed usage, not a randomised comparison, and should not be quoted as a measured effect. PROMOTED TO AIE-16 on 2026-09-16, re-verified at NBER's own abstract page rather than carried across from this summary. Every figure matched, and the reading added one line this item had not quoted -- the authors' own conclusion: 'realizing the promise of AI tutoring will require getting students to use it, not just giving them access.' Also confirmed from source: DOI 10.3386/w35620, issue date August 2026, Oreopoulos and Low. Registered TIER 2 because it is an unreviewed working paper, not because of the design, which is the strongest in the domain -- revisit the tier if it is published.
A randomised trial finds ChatGPT cuts teacher lesson-preparation time by about a third, with no detected drop in resource quality
What it found
68 English secondary schools and 259 teachers were randomised to use ChatGPT with a supporting guide, or to prepare Year 7 and Year 8 science lessons without generative AI. In weeks six to ten, the ChatGPT group spent 69% of the preparation time the comparison group did: 56.2 minutes per week against 81.5, a saving of 25.3 minutes per week. An expert panel that did not know how resources had been produced found no evidence that quality differed between the groups.
How to read it
The honest reading is narrow and useful: this measures teacher workload, not student learning, and it does not claim the lessons were better. What makes it worth watching is the blinded quality panel — the common objection to time saved with AI is that the saving is paid for in quality, and this trial looked for that and did not find it. Note also that teachers used the tool and the guide less as the trial went on, which the report attributes to moving out of a learning phase; the saving was measured in weeks six to ten, once that settling had happened, which is the more honest window than the first flush of novelty.
What would weaken it
Teacher attrition was substantial — 259 randomised, 211 analysed, 18.5% overall and higher in the ChatGPT group (21.7%) than the comparison group (15.4%), which is the kind of differential attrition that can bias a result. The trial covers one subject (science) at Years 7 and 8 in English schools, and measures preparation time and resource quality only; no pupil-attainment outcome was measured. 'No evidence of a difference' in quality is not the same as evidence of no difference.
Pew: just over half of US teens now use chatbots for schoolwork, and a majority believe AI cheating is routine at their school
What it found
Surveying 1,458 US teens aged 13 to 17 between 25 September and 9 October 2025, Pew found 54% say they have used chatbots to get help with schoolwork and 57% to search for information. Separately, 59% — a majority — think using AI to cheat is a regular occurrence at their school, happening at least somewhat often, including about a third who say it happens extremely or very often.
How to read it
Two different kinds of number sit in this report and they should not be blurred. The 54% is a self-report of one's own behaviour, which is reasonably reliable and unsurprising. The 59% is a perception of other people's conduct, which is not a measurement of cheating at all — it is a measurement of what teenagers believe is normal around them. That belief matters in its own right, because a student who thinks everyone else is doing it faces a different decision than one who does not, but it should never be reported as a cheating rate. Much of the press coverage of surveys like this makes exactly that substitution.
What would weaken it
US-only and self-reported throughout, with a margin of error of plus or minus 3.3 percentage points. Australian rates cannot be inferred from it. The cheating figure is a perception of peers, not an observed or verified incidence, and 'regular occurrence' is a subjective threshold each respondent sets for themselves.
NAPLAN 2026: reading slipped nationally, and ACARA has placed its own caveat on this year's writing results
What it found
In the 2026 national results, the average proportion of students at Exceeding or Strong in reading was 66.3%, down from 68.2% in 2025, and the proportion needing additional support rose from 9.3% to 10.2%. Reading proficiency rises from Year 3 (65.6%) to Year 5 (70.9%), then falls in Year 7 (66.8%) and Year 9 (62.0%). ACARA also reports a technical failure of the national NAPLAN platform on the first morning of testing that affected only the writing test and only Years 5, 7 and 9, largely on the east coast; national writing averages held steady, but because the disruption's impact varied between students and schools, ACARA has attached a caveat to the 2026 writing results.
How to read it
The unusual thing about this release is that the body publishing the data is the one telling you not to lean on part of it. That caveat should be honoured rather than skipped: the 2026 writing figures are not a clean year-on-year point, and anyone citing a writing trend across 2025 to 2026 is citing a number its own custodian has qualified. The reading movement is the more usable finding, and the shape within it matters more than the national average — proficiency peaks at Year 5 and then declines through secondary school, with Year 9 the weakest cohort. A one-year change of this size is not yet a trend, but it runs the wrong way against a decade of policy attention to early reading.
What would weaken it
A media release summarising national results, not the full national report, so these are headline figures without the underlying tables. Proficiency standards were rebuilt in 2023, so nothing here can be read as a trend against pre-2023 bands. National participation fell slightly, and Tasmanian state-level results were not reported at all because participation during industrial action fell below ACARA's technical standard — with localised flooding also affecting the Northern Territory. The writing results carry ACARA's own caveat.
The NAPLAN writing gender gap is now roughly sixteen points wide at Year 9
What it found
ACARA reports that female students outperformed male students in writing in every year group in 2026. At Year 7, 69.6% of female students reached Strong or Exceeding against 55.6% of male students; at Year 9, 68.1% against 52.0%. Male students generally outperformed female students in numeracy, where the reported gap is far smaller.
How to read it
A sixteen-percentage-point gap at Year 9 means that just over half of boys reach the standard while just over two-thirds of girls do — and the gap widens from Year 7 to Year 9, the years in which subject choice and senior-English confidence are being set. It is worth naming what this is not: a claim about ability. NAPLAN writing is a single timed task assessed on specified criteria, and a gap on that task is evidence about performance under those conditions. The asymmetry with numeracy, where the gap runs the other way but is much narrower, is the part that resists an easy explanation and is the reason this is worth watching rather than concluding.
What would weaken it
These are proportions at a proficiency level, not mean scores, so they are sensitive to where the cut sits — a cluster of students just below the Strong boundary exaggerates a gap that a mean-score comparison would show as smaller. The 2026 writing results carry ACARA's own caveat following the first-day platform disruption affecting Years 5, 7 and 9. The release reports the gap without decomposing it by socio-educational background or language background, so this cannot be read independently of those factors. ACARA states that “a technical issue affecting the assessment platform impacted many students in Years 5, 7 and 9 who sat their writing test on 11 March 2026”, and that “caution should be applied in interpreting writing results in 2026” — see FR-11. The writing figures above are 2026 figures and inherit that caution.
NAPLAN 2026: Year 9 writing and grammar remain the two weakest cells nationally
What it found
Across the twenty national cells (four year levels x five domains), the two lowest proportions of students at Strong both sit at Year 9: grammar and punctuation at 38.1% and writing at 38.2%. No other cell falls below 40% at that level. Counting Exceeding as well, the same two cells sit at 54.0% and 59.8%.
How to read it
The interesting figure is not that Year 9 is lowest — later year levels are assessed against higher expectations, so some decline is built into the design. It is that writing and grammar are the only two domains where fewer than two students in five reach Strong, while reading and numeracy at the same year level sit substantially higher. The gap is domain-specific, not a general Year 9 effect.
What would weaken it
NAPLAN proficiency levels are expectation-relative and were rebuilt in 2023, so this cannot be read as a trend against pre-2023 bands. Participation also fell from 94% to 89% across the period, which affects cohort comparability. ACARA states that “a technical issue affecting the assessment platform impacted many students in Years 5, 7 and 9 who sat their writing test on 11 March 2026”, and that “caution should be applied in interpreting writing results in 2026” — see FR-11. The writing figures above are 2026 figures and inherit that caution.
Corrected
As first published this reading said "Strong or above" where it meant "Strong". The two figures are Strong-level proportions: adding Exceeding puts the same cells at 54.0% and 59.8%, and no cell in the grid falls below 40% at Strong-or-above (the lowest is Year 3 grammar and punctuation at 53.9%). The finding is unchanged under the metric actually meant, and the corrected text now states both.
AERO: EAL students need at least six years of schooling to participate in the curriculum on equal terms
What it found
Drawing on NAPLAN and EAL/D survey data for over 110,000 EAL students in New South Wales public schools from 2014 to 2022, the report finds that students who start school at Beginning levels of English can on average expect to participate equitably in curriculum learning only after at least six years of schooling — consistent with international research putting the figure at roughly five to seven years. Each successive phase of the EAL/D Learning Progression took longer than the one before: a median of 1 year 1 month from Beginning to Emerging, about 1 year 8 months from Emerging to Developing, and about 2 years 7 months from Developing to Consolidating or higher. Socio-educational disadvantage (22% slower), refugee experience (14% slower), being male (6% longer) and starting in Kindergarten rather than a higher primary year (9% slower on average) were each significant independent predictors.
How to read it
The number that matters for schools is the shape rather than the headline: progress decelerates as proficiency rises, so the phase where a student looks closest to fluent is the phase that takes longest. That is precisely the point at which support is most likely to be withdrawn, because conversational English arrives well before the academic English the curriculum runs on. For a family, the useful reframing is that a child who sounds fluent in the playground may still be years from reading a set text on equal terms, and that gap is normal rather than a sign of not trying.
What would weaken it
New South Wales public schools only, so it describes one system rather than Australia. First Nations students learning English as an additional dialect were deliberately excluded, for data-completeness and Indigenous Data Sovereignty reasons the report states openly — the findings do not speak to them. The demographic effects are independent predictors from a model, not discrete causes, and the report says so. The headline is also an average across a wide spread: the reported ranges for the fastest 25% and 75% of students differ by more than a year at some phases.
VCE English examination report 2025: three standing criticisms hold for an eleventh year, the fourth does not
What it found
The 2025 report repeats three of the four criticisms that appear in every examination report from 2015 onward: topics not answered, identification rather than analysis, and plot retelling. The fourth, pre-prepared or formulaic responses, appears in nine of the eleven years -- absent from the 2022 and 2023 reports -- and returns in 2025. Technique-listing, the only criticism that had shown improvement, relapses in 2025.
How to read it
Ten or eleven consecutive years of the same complaints is not a series of coincidences about particular cohorts -- it is a statement about what the assessment reliably surfaces. The one criticism with a gap is instructive rather than reassuring: pre-prepared responses go unmentioned in 2022 and 2023 and are back by 2025, so the pattern is intermittent reporting of a persistent problem, not its resolution. For a student, the practical reading is that the highest-yield preparation is not more content coverage but the specific discipline of answering the topic as worded.
What would weaken it
Examination reports are assessor commentary, not measurement. They describe what assessors noticed, which is weighted toward what went wrong; they are not a representative sample of scripts.
Corrected
As first published this reading said all four criticisms appear in every report from 2015 onward. A full re-read of all eleven reports (registered at VCE-17) found the fourth -- pre-prepared or formulaic responses -- absent from the 2022 and 2023 reports: 42 of 44 theme-year cells, not 44 of 44. The title, finding and reading are corrected here to match the register entry they cite.
Once a frontier reading has been verified to the register’s full standard, it is registered as a numbered entry and cited from here — the frontier item stays as the dated record of when the research was first read. The verification protocol lives on the register’s main page.
