Finding and Growing Academic Talent

What the evidence says, and what it means for India. An evidence review for Central Square Foundation, GenWise and Avanti Fellows. Version of 30 September 2026.

This review asks how a country should find children with high academic potential, especially children from poor and rural families, and what develops that potential once they are found. It draws on fifteen research memos, each of which downloaded and read its sources, and on Avanti Fellows’ linked administrative records. Every figure from those records was re-derived for this review; where it differs from an earlier internal Avanti estimate, the re-derived figure is used (Section 8.5).

Evidence labels.

Label	Design	What it can support
Causal	Randomised trial, regression discontinuity (RD), lottery, or difference-in-differences with a credible comparison	“X raised Y”, for the population and margin studied
Quasi-experimental	Matching or regression on rich observables, or a policy change compared with other districts	“X is associated with Y, holding observed differences fixed”
Predictive	Correlation between an early measure and a later outcome	Forecasting; nothing about what a programme did
Descriptive	Administrative counts, documents, cross-sections	Scale, structure, who is where
Expert	Advocacy reports, handbooks, frameworks	Considered opinion
Correlational / Observational	Association in observational data, without a credible design for causation	“X goes with Y”
Meta-analysis	A pooled estimate across studies	As much as the pooled studies’ designs support
Model / Simulation	Results from a structural model or from simulated data under stated assumptions	What follows if the assumptions hold

Terms. SD: standard deviation, the usual yardstick for effect sizes (0.1 SD is small, 0.3 SD large for an education programme). g: the gap between two groups’ average scores, in SDs. h: the same idea for a gap between two proportions. ρ: rank (Spearman) correlation. DIF: an item that behaves differently for two groups of equal ability. u: the ratio of applicant to admitted score spread, used to correct a correlation for selection.

Indian tests and programmes named in this review.

Term	What it is
JNV	Jawahar Navodaya Vidyalaya: free, residential, CBSE government schools (about 660), one or two per district, run by Navodaya Vidyalaya Samiti (NVS)
JNVST	JNV Selection Test: the Class 5 entrance test for Class 6 admission to JNVs, sat by about 19 lakh children for about 46,000 seats
NCST	Navodaya CoE Selection Test: a Class 10 test for JNV students competing for free two-year JEE/NEET coaching seats (Classes 11–12) at Dakshana Foundation, Ex-Navodaya Foundation and Avanti. Dakshana ran it for its own applicants in 2022–2025 (opt-in; about 10% of the JNV Class 10 cohort in 2022); in 2026 NVS administered it nationally, reaching most of the cohort
CoE	Centre of Excellence: intensive on-campus JEE/NEET coaching for Classes 11–12 inside a JNV (dedicated teachers, a structured JEE/NEET curriculum, regular full-length tests), run by Avanti, Dakshana or ENF, with admission on NCST
Nodal	Avanti’s lighter-touch JEE/NEET programme in other JNVs: less contact time and less intensive preparation
KV	Kendriya Vidyalaya: central government day schools, mainly for children of central government employees
NMMSS	National Means-cum-Merit Scholarship Scheme: a central scholarship for Classes 9 to 12, awarded on a state Class 8 test with an income ceiling (Section 5.7)
NTSE, KVPY, INSPIRE	National Talent Search Examination; Kishore Vaigyanik Protsahan Yojana; the Department of Science and Technology’s science-talent scheme (Section 7.3)
HBCSE	Homi Bhabha Centre for Science Education, the nodal centre for India’s science and maths Olympiads
NIAS	National Institute of Advanced Studies, Bengaluru, which ran a nominated gifted-education programme
AGC	CBSE’s Aryabhata Ganit Challenge, a school maths competition
EI, ASSET, ATS	Educational Initiatives, a private assessment company; ASSET, its diagnostic school test; ATS, its ASSET Talent Search
NTA	National Testing Agency, which runs JEE Main
NAS	National Achievement Survey, the Ministry of Education’s sample survey of learning
CMS-E	Comprehensive Modular Survey: Education, a MoSPI household survey of education spending (2025)
UDISE+	Unified District Information System for Education Plus, the national school census
KGBV, NPEGEL	Kasturba Gandhi Balika Vidyalaya, residential schools for girls from disadvantaged groups; National Programme for Education of Girls at Elementary Level
EMRS	Eklavya Model Residential Schools, for Scheduled Tribe students
SC, ST, OBC, EWS	Scheduled Castes, Scheduled Tribes, Other Backward Classes and Economically Weaker Sections, the reserved categories
APAAR, DPDP Act	Automated Permanent Academic Account Registry, a national student ID; Digital Personal Data Protection Act 2023
NITI Aayog DMEO	The Development Monitoring and Evaluation Office of NITI Aayog, the government’s policy think tank

---

1. Summary for decision-makers

Each item is one claim, the evidence behind it, and its evidence label (defined above).

Why find and nurture talent

• Able children from poor families are lost, and the loss is about environment, not ability. In US data on 1.2 million inventors, children from top-1% families are ten times as likely to become inventors as those from below-median families, and the gaps “persist even among children with similar math test scores in early childhood”. Children who move to high-innovation areas young are more likely to invent, a causal exposure effect (Bell, Chetty, Jaravel, Petkova & Van Reenen 2019). Causal (exposure); Correlational (gaps).
• The same holds across countries at the very top. Maths Olympiad participants from low-income countries produce 34% fewer publications than equally scoring peers from rich countries, which the authors call a lower bound (Agarwal & Gaulé 2020). Correlational.
• Without support, students in weak schools stall. In Avanti’s Haryana cluster RCT, control-group students in government senior secondary schools showed no learning gain over two years in maths or science, at any skill level: “we can confidently rule out even small learning gains” (de Barros & Kumar 2024, J-PAL/Avanti Haryana RCT endline report). This covers all students, not only the most able. The clearer gap is in who converts ability into outcomes: among JNV students with 90%+ Class 10 marks, 9% of girls against 21% of boys qualify for JEE Main (Section 4.5). Causal design; descriptive for the control group.
• The size of the top tier matters for growth. Across countries, a 10 pp larger share of top performers goes with 1.3 pp faster annual growth, against 0.3 pp for basic literacy; the two are complementary and the top-performer effect is larger in poorer countries (Hanushek & Woessmann 2008, 2012). Growth models tie long-run growth to the number of people producing ideas (Bloom, Jones, Van Reenen & Webb 2020). Correlational, cross-country; growth accounting.
• Emigration does not cancel the gain from a bigger pool. 36% of the top 1,000 JEE scorers and 62% of the top 100 have migrated, and attending a top-5 IIT itself adds 4 to 5 pp to emigration at equal scores, through signalling and alumni networks (Choudhury, Ganguli & Gaulé 2023). Leaving is concentrated at the very top and runs through elite institutions, so a broader pipeline adds mostly to the part of the pool that stays; that last step is our inference. Migration prospects can also grow the stock at home: each additional Filipino nurse abroad came with 9 more nursing graduates passing the licensure exam (Abarcar & Theoharides 2024, causal), and US H-1B demand helped build India’s IT sector (Khanna & Morales 2017, structural model). Quasi-experimental; Causal; Model.
• What we do not know. No study we found shows that rural, low-income or reserved-category top talent emigrates less; the IIT paper covers the general category only. Losing top inventors is a real cost to India’s access to new knowledge (Agrawal, Kapur & McHale 2011), and elsewhere wealth barely predicts who leaves (Gibson & McKenzie 2011).

A. Finding talent

1. Test every child instead of waiting for referrals.
   • Broward County, Florida, moved from testing only referred children to testing every second grader, with the same pass mark. Low-income and English-learner children identified as gifted rose from 1.4% to 4.3%, and fell back when universal testing was cut.
   • The children found this way had the same IQs as referred children. Teachers had missed them because their classwork looked weaker.
   • India: about three quarters of JNV students at the JEE Main 95th percentile in 2024 had never sat NCST 2022, which was opt-in and reached about 10% of the Class 10 cohort. This shows what a voluntary test misses, not what a universal one would add.
   • Quasi-experimental, one district (Broward); Descriptive (Avanti).
2. Children near the top move in and out, so one early test should never be a final gate.
   • Even on a test with reliability .98, about 40% of a Grade-3 top-3% group has left it a year later, and others have joined. Part of this is measurement noise; part is real differences in development and schooling.
   • So: keep re-entry points at transitions (Class 5, 8 and 10); select a wider band than the target (for example the top 10% to find the top 3%) and refine it; combine a test with marks.
   • Avanti’s data fit this. JNVST at Class 6 correlates r = 0.07 with JEE Main (0.14 within school; plausibly 0.2 to 0.4 allowing for JNV admitting about 2% of applicants, unchecked without applicant data); NCST at Class 10 correlates r = 0.585. The evidence supports repeated identification, not moving the first gate earlier than Class 5.
   • Predictive.
3. The best way to pick the strongest students is a hard test taken close to the decision. Giving young children tests meant for older students (“above-level testing”) helps a little, but is not proven to do better.
   • Above-level testing gives a 12-year-old a test built for 17-year-olds, so the very top students are not all bunched at full marks. In the US SMPY study, among children already in the top 1% at age 13, the top quarter on such a test went on to earn more doctorates and patents than the bottom quarter, but the differences are modest (effect sizes h = .28 and .18; 0.2 is small). No study has compared it with a hard test pitched at the child’s own grade.
   • What predicts best is recent, demonstrated achievement on a demanding test close to the outcome: NCST 2022 correlates 0.585 with JEE Main 2024 and board maths 0.595 (n = 1,371), and NCST separates better at the top.
   • Combining measures adds a little: NCST lifts explained variance from 0.41 to 0.44 over board marks, category and gender. Spatial ability adds its own signal for STEM (Wai et al. 2009).
   • For adult outcomes, what happens after selection matters more: third-grade maths scores explain under a third of the gap in who becomes an inventor between top-quintile families and the rest (Bell et al. 2019).
   • Predictive.
4. Girls fall behind at the top of NCST and JEE, not on boards.
   • JNV CBSE Class 10 Maths Standard (2022-2026): girls are about 40% of takers and 33 to 39% of the top decile.
   • NCST 2026 (engineering): girls are 34.6% of takers and 11.0% of the top 1%. NCST predicts JEE equally for girls and boys.
   • Descriptive.

B. What programmes work, and why

5. What works: teach able students more, faster and with extra help, and support them through transitions. In plain terms, the gains come from what happens in the classroom and around it, not from the label of being selected.

   • Acceleration is the best-supported practice. Accelerated students beat same-age peers by g = 0.70 and match the older peers they join (g = 0.09). g is the gap between two groups’ average scores in standard deviations: 0.70 moves an average student from the 50th to about the 76th percentile, and 0.09 is close to no difference. So acceleration works by teaching more content sooner; students keep up with the older class. Quasi-experimental (meta-analysis of mostly pre-1990 matched studies).
   • Grouping high achievers helps only when the curriculum changes with it: special grouping for the gifted g = 0.37, against 0.04 to 0.06 for tracking that keeps the same course of study (Section 5.5). Quasi-experimental.
   • A changed classroom for selected high achievers: low-income and minority high achievers in Florida gifted classes (+0.4 to 0.5 SD); Boston’s Advanced Work Class (+15 pp college enrolment, marginally significant); a tracked high-achieving class in one Chinese high school (+0.23 SD maths). The authors credit teachers (China) and expectations and peer pressure (Florida, their interpretation). Causal, mostly high-income.
   • Tutoring and frequent testing: daily small-group tutoring has the cleanest trial support of any school practice (+0.16 and +0.37 SD in two trials, effect on those tutored). Frequent assessment used to adjust teaching, and extra instructional time, also hold up. Causal (tutoring); Correlational (other practices).
   • Residential schools that also change the school: SEED in Washington DC (about +0.2 SD a year in reading, +0.23 SD in maths, almost all among girls) and France’s internats d’excellence (over +0.2 SD a year in maths from year two, about +0.57 SD for the top third), at about twice the cost of a day school. Malawi’s National boarding secondary schools, which admit roughly the top 5% of candidates at the end of Grade 8, raised Grade 10 exam scores by 0.40 SD at the admission cut-off (lower bound 0.23 to 0.28 SD after attrition), more for girls and for students from districts with weak primary schools. In the same data, ordinary district boarding schools add only about 0.13 SD over community day schools, against about 0.57 SD for the resource-rich National schools, so the school, not the dormitory, carries most of the gain. Boarding works as an amplifier of the above (supervised study time, access for children who would otherwise drop out), not as a cause on its own. Causal (lotteries, RD) for the bundles; value-added for the boarding comparison.
   • India: no Indian residential school has been evaluated for learning or exam results against a comparison group. The causal Indian estimates concern enrolment, and run both ways: the NPEGEL and KGBV girls’ bundle raised girls’ upper-primary enrolment by 6 to 7 pp, mostly through day schools (Meller & Litschig 2015), while an EMRS opening cut ST women’s schooling by about 1.25 years by crowding out day schools (Tripathi 2025, unrefereed). Programmes that report strong results select their students on achievement, so their pass rates are not effects. Causal (enrolment); Descriptive (results).
   • Support at transition points (scholarships and fee removal): merit awards in Kenya, fee removal in Ghana, and means-and-merit aid in Colombia, which raised earnings by 18 log points. India’s NMMSS targets similarly, is far smaller in value, excludes JNV students and has never been evaluated. Causal abroad; none in India.
   • For Avanti: the CoE combines dedicated teachers, a structured JEE/NEET curriculum taught ahead of the school syllabus and regular full-length tests inside a residential JNV. CoE students score about 14 to 17 JEE percentile points above Nodal students at equal Class 10 marks, and Nodal students look no different from JNV students with no Avanti programme. Observational; CoE selection on NCST and motivation is not controlled.

6. The gains are largest for high achievers whose alternative is weak.

   • Select on demonstrated achievement: Florida students selected on past scores gained 0.4 to 0.5 SD, while IQ-selected students in the same gifted classes gained nothing.
   • The weak alternative is the strongest moderator in most studies. North Carolina’s residential STEM school helped rural, minority and lower-achieving admits most, and the same charter model raised scores in cities and lowered them outside cities. Kenya is the exception: national-school admission had no detectable test-score effect even for admits with the weakest alternative schools (Lucas & Mbiti 2014).
   • The exceptions to the selective-school nulls fit this: large exam gains in Trinidad and Tobago (about twice as large for girls), +0.25 SD at one Chilean school, and long-run attainment gains in Barbados (more for women).
   • For JNVs: a JNV admit’s alternative is often a weak rural school, so JNV’s effect could be larger than the nulls abroad, though Kenya’s null for its weakest-alternative admits makes this a hypothesis. It has never been measured.
   • Causal, mostly high-income; the moderator comparisons are across studies.

7. Common misconceptions: things that sound like they work but do not on their own.

   • Being selected. At Boston and New York exam schools, gifted programmes in Florida and a large Southwestern US district, Kenya’s national schools and top schools in Romania (+0.05 SD) and Barbados, students just above the cutoff end up with similar test scores to those just below. These schools look strong mainly because of whom they admit. Causal, local to the cutoff.
   • Strong classmates. Classmates are about a tenth of school value-added in Trinidad and Tobago; putting bright children together is not the active ingredient. Causal.
   • Boarding. Designs that hold the programme fixed and vary only residence or dorm life (online against residential STEM programmes; randomised dorm-mates in Peru) find little academic effect of residence itself. Causal.
   • High standards and expectations. “High expectations” add little once tutoring, feedback, data use and time are held fixed, and curricular rigour measured directly does not predict school effectiveness. Correlational.

8. Things to be careful of when selecting and grouping able students.

   • Self-concept falls when students join stronger peers (the big-fish-little-pond effect, β = −0.20 to −0.28), more so in Asia and in high school. Track rank, confidence and dropout, not only scores. Correlational, many countries.
   • Marginal admits can drop out. In Mexico City, admission to a school much harder than the alternative raised dropout by 9.4 pp, while raising end-of-school maths scores for the marginal admit who stayed. Causal (RD).
   • Gains take time and can skip weaker admits. France’s internats showed no maths gain in year one, and only the top third gained significantly. Causal (lottery).
   • Peer effects can hurt girls. In Peru’s selective boarding schools, a girl’s reading fell 0.12 SD for each 1 SD stronger dorm neighbour. Causal (field experiment).
   • Pace has to match readiness. Charlotte’s 7th-grade algebra for the top two quintiles gave mixed results. Causal.
   • The gender gap widens at the top: girls are 34.6% of NCST 2026 engineering takers and 11.0% of its top 1%. Design selection and support for girls explicitly. Descriptive.

C. What India and Avanti’s data show

9. India selects at scale but has never evaluated a school-age talent programme.
   • JNV selects about 46,000 a year from about 19 lakh who sit the test. NMMSS and an Olympiad ladder also exist.
   • NEP 2020 commits NCERT/NCTE to gifted-education guidelines and permits a B.Ed. specialisation; neither was found.
   • No causal learning evaluation of JNV, state residential societies or ashram schools was found in this search.
   • Descriptive.
10. Avanti’s linked records are the only Indian evidence connecting early tests and programmes to JEE outcomes.
    • JNVST → JEE, NCST → JEE and board marks → JEE (items 2 to 4; Section 8), and the CoE, Nodal and online comparisons inside JNVs (item 5; Section 5.3).
    • The data cannot answer two questions: what residence adds, since no comparable non-residential group has complete JEE records on the same board, and what JNV adds over the school a child would otherwise attend, since non-admits are not linked.
    • What Avanti’s data show, re-derived and confirmed:
      • A Class 10 test predicts JEE well. NCST 2022 correlates r = 0.585 with JEE Main 2024 (n = 1,371). Of the NCST top 10%, 51.7% reached the JEE 95th percentile, against 33.6% of students with 99+ in board maths.
      • Class 10 marks beat the Class 6 entry test. Among JNVST 2018 admits (n = 4,015), Class 10 maths correlates r = 0.367 with JEE Main 2025, JNVST r = 0.072.
      • The intensive CoE model is linked to a large JEE edge: about 14 to 17 percentile points over Nodal at equal Class 10 marks (2024-2026).
      • JNV students reach the JEE top tail more often than the national field: about 1.1 to 1.75 times the national share at or above the 90th percentile across 2021-2025.
      • The gender gap opens at the top of NCST and JEE: girls are 34.6% of NCST 2026 engineering takers and 11.0% of its top 1%.
      • In EI’s five-JNV maths pilot, Class 6 entrants score at about the ASSET-norm average (4.3% in its top 5%; mean at the 55th percentile), not at its top. The pilot is cross-sectional and cannot yet show whether JNV schooling erodes their standing; EI’s within-student panel can (Section 8.3).
    • Predictive; Descriptive; Observational.

D. What it means for a national effort

11. Find broadly and re-test; put the effort into what students get after they are found.
    • Identification: universal where possible, repeated at transitions, local norms, marks plus a test.
    • Nurture: changed instruction for students with weak alternatives (teaching at their level (Section 5.3), more supervised time, frequent testing, tutoring), plus scholarships at transitions. Residence is justified where distance or home circumstances would otherwise keep a student out; the evidence does not support it as an academic lever in itself.
12. The most useful studies to commission are regression-discontinuity studies of JNV and CoE admission, and a residential test.
    • JNV’s multi-quota admission supports a fuzzy RD at the cutoffs, if non-admits’ later records can be linked.
    • A cheaper first step is an RD at the NCST cutoffs for CoE admission, which would show what the CoE itself adds.
    • Assigning otherwise similar students to residential and non-residential tracks by lottery or cut-off would measure what residence adds.
    • Design recommendation.

---

2. What “talent” means

Every system answers three questions before it identifies anyone: talent in what, compared with whom, and measured when. The answers decide who is found, often more than the test.

Ability, achievement or potential. Gagné, Renzulli, the Munich model and Stanley’s SMPY define giftedness differently (Expert), and no study compares programmes built on one against another. The choice matters in practice. In Broward, among low-income and English-learner (“Plan B”) students, those found only through screening had the same mean IQ as Plan B gifted students before screening (124.5 against 124.2) but lower achievement (0.81 against 1.15 SD): referral had selected on visible achievement. Across all students the newly found had lower IQs (124 against 132), reflecting Plan B’s lower bar (Card & Giuliano 2016, PNAS). In the same district, low-income and minority high achievers placed in gifted classrooms on past scores gained 0.4 to 0.5 SD, while IQ-qualified students gained little, “even infra-marginal” ones (Card & Giuliano 2014).

How high a bar. Israel calls the top 2% gifted (Knesset Research and Information Center 2017), Taiwan +2 SD, Singapore’s GEP about 1% (last Primary 4 intake 2026); England identified about 22% at some point in Years 7 to 10 against guidance of 5 to 10%; Saudi Arabia’s MMCAT labelled 36% of 2025 testers gifted. These labels are not comparable on one scale. For reference, IQ 120 is about the 91st percentile on an SD-15 scale.

Compared with whom. Re-scoring NWEA MAP data from ten US states against school norms raised Black students’ representation index in reading from .22 to .76 at a 5% cut (Peters et al. 2019; Descriptive simulation). JNVST is already a local-norms design (Section 4.4). Measured when: Section 3.4.

A working definition for India. Children whose demonstrated achievement or reasoning in a domain is well ahead of peers with similar opportunities, measured more than once. Domain scores are worth reporting, since extreme discrepancies between domains are common among able students, but they are less reliable than a composite (CogAT difference-score reliabilities of about .57 to .67; Lohman, Gambrell & Lakin 2008). Selection can use the composite or the strongest domain, and a profile should be confirmed on a second occasion before it steers placement.

---

3. Identification

Recommendation. Screen widely and often; select narrowly and late. In practice:

1. Test everyone in the target group, not volunteers or nominees. Opt-in and referral miss the most (Broward; about three quarters of JNV JEE top scorers in 2024 never sat the opt-in NCST 2022).
2. Use a two-stage design. Stage 1 is a broad, cheap screen that keeps roughly the top 10% of each district or block, with category and gender norms, which JNVST already does. Stage 2 is a harder test, with enough ceiling to separate students within the top 1 to 5%, that decides the scarce places. The first gate is wide because top-tail membership churns (about 40% of a top-3% group leaves within a year); the second needs ceiling because grade-level exams bunch the best students at full marks (235 of 1,371 NCST 2022 takers scored 99 or 100 in board maths).
3. Make the stage-2 instrument a hard, curriculum-based maths and reasoning test, NCST-style, combined with school marks. Above-level items can sit inside it to spread the very top; they are not proven to beat a hard grade-level test. Do not rely on nonverbal (figural) items alone.
4. Re-test at transitions and keep re-entry open: Class 5 (JNVST), Class 8, and Class 10 (NCST). No early score should close the door.
5. Validate against outcomes and watch the gender tail. Link every stage to later results (as NCST → JEE does) and check who is lost at the top, where girls fall from about 35% of takers to 11% of the top 1% on NCST.

The thresholds (top 10%, top 1 to 5%) are design choices drawn from the evidence below, not tested cut-offs.

3.1 The instruments

Instrument	Evidence	Label
Grade-level test	Discriminates across most of the range; ceiling for a small top group. SMPY’s first gate is a grade-level top 3 to 5% screen (Lubinski 2016)	Predictive
Above-level test	Spreads the top 1%; predicts doctorates, patents and income 20 years later, modestly (Wai et al. 2005), with accomplishment still accruing at 40 years (Lubinski, Benbow & Kell 2014). No head-to-head test against a high-ceiling grade-level test	Predictive
Adaptive / level-spanning	No floor or ceiling; showed a five-to-six grade-level spread within each grade in its Delhi sample (Muralidharan, Singh & Ganimian 2019)	Descriptive
Nonverbal (Raven’s, NNAT)	Not culture-free: selecting on it “excludes many of the most academically capable Black students” (Lohman 2005); low-income students score about 6 IQ-like points lower after ethnicity and language controls (Lohman & Gambrell 2012), and 8.1 after ethnicity controls in a kindergarten NNAT sample (Carmen & Taylor 2010, as reported in Lohman & Gambrell 2012). Adds little to verbal prediction. Spatial ability has its own STEM value (Wai et al. 2009). Half of JNVST is figural (Section 8.1)	Predictive
Teacher judgement	Correlates about .6 with tests (Kaufmann 2024); at .6, a teacher’s top-5% list holds roughly a third of the test’s top 5%	Meta-analysis
Nomination gate	Can push false negatives past 60% (McBee, Peters & Miller 2016)	Simulation
Dynamic assessment	China’s SCGY ends selection with a one-week trial of learning rate (Dai & Steenbergen-Hu 2015); no outcome comparison	Descriptive

A first-stage screen should be lenient, and the combination rule matters: for two tests correlating .80, each with a top-3% cut, requiring both admits 1.35%, accepting either 4.65%, averaging about 2.4% (Lohman & Korb 2006).

3.2 Above-level testing and SMPY

SMPY’s 1,650 top-1% adolescents from two cohorts had, four decades later, “published 85 books and 7,572 refereed articles, secured 681 patents” (Lubinski, Benbow & Kell 2014). Within the top 1%, the top quartile at 13 earned more doctorates (32.1% against 20.0%) and patents (7.5% against 3.8%) than the bottom quartile (Wai et al. 2005), and the ability pattern predicts the domain of accomplishment (Park et al. 2007; Makel et al. 2016). This is prediction: there is no comparison group of equally able children who were not identified. By high school SMPY’s top students were “bumping their heads on the ceiling of the SAT” (Lubinski & Benbow 2006), so for India above-level testing fits Classes 5 to 8, and by Classes 10 to 12 a JEE-level paper is the above-level instrument. Lubinski & Benbow (2006, footnote 4) note that one SMPY participant, identified well before adolescence on an above-level SAT-M, won the Fields Medal at 31, and that two future Nobel laureates, Alvarez and Shockley, were rejected by Terman’s study. These are single cases and do not compare outcomes.

3.3 Ceilings

A test has a ceiling when too many strong students score at or near full marks, so it can no longer tell them apart: a child who would have scored 110 and one who would have scored 100 both get 100. For finding the very top students this is the main weakness of ordinary school exams. A ceiling depends on the test and on who takes it, and can be fixed with harder items or extended scoring (Pearson’s extended WISC-IV norms; Zhu et al. 2008). In Avanti’s data, JNVST admits matched to JEE 2025 (n = 4,015) average 85.0 of 100 (SD 8.0), with 12.9% at 95 or more. Among the 1,371 NCST 2022 engineering takers who sat JEE 2024, 42% scored 95 or more in board maths and 6.5% scored 100 (34% and 5.2% in the whole linked engineering pool, n = 1,957). NCST and JEE percentiles show no ceiling (Section 8.2). India already has above-level instruments: JEE for Classes 11 and 12, and the Ei ASSET Talent Search for Grades 4 to 8.

3.4 Stability of the top tail

Rank is fairly stable: IQ at 11 and 77 correlate .63 (.73 corrected; Deary et al. 2000), and school achievement (marks and tests pooled) has a two-year rank stability of ρ = .70 at Grade 5, .72 for tests and .67 for marks (Scherrer, Breit & Preckel 2025; abstract only, so whether it is corrected for unreliability is unknown).

Top-tail membership is not. For the Iowa Tests, with K-R 20 reliability of .98 and a one-year correlation of .91, Lohman & Korb (2006, Table 1 and Figure 2) estimate under bivariate normality that about 60% of the composite top 3% in Grade 3 stay there in Grade 4, about half on single subjects, and about half on the composite by Grade 8. Their text says “only about 40%”, which their table and figure do not support. Internal consistency misses day-to-day error and real change, so even a very reliable test churns at the tail. At r = .70, about 47% of a top-10% group stay top 10%, and 27% of a top-1% group.

In Pakistan’s LEAPS panel, the bottom fifth at Grade 3 gained 1.75 SD by Grade 6 against 0.71 SD for the top fifth; measurement error explains part of this, and the gap narrows without reversing once it is addressed (Bau, Das & Chang 2021). The authors note that value-added models typically find half to two thirds of last year’s learning persisting (λ of 0.5 to 0.7, from the wider literature); that is persistence of learning, a different quantity from rank stability.

Implication. At ρ ≈ .70 from Class 6 to Class 8, fewer than half of a Class 6 top-10% list would still be top 10%. A single test at age 10 cannot be the only gate.

When to identify. Earlier identification buys a longer window for support and locks in more error, since rank is less stable the younger the child. Learning gaps open in the first two or three years of school, and a school year’s productivity differs across countries (Singh 2020), so an early score also reflects the schooling a child happened to receive. The evidence supports repeated identification (for example at Classes 5, 8 and 10) and does not support moving the first gate earlier than JNVST’s Class 5. No outcome evidence on timing.

3.5 Predictive validity compared

Early measure → later outcome	r	n	Sample	Source
IQ at 11 → IQ at 77	.63 (.73 corrected)	101	Scottish survivors	Deary et al. 2000
Achievement (marks and tests), Grade 5 → 7	ρ .70	meta	363 studies	Scherrer et al. 2025
JNVST 2018 → JEE Main 2025 percentile	.072 [.041, .103]	4,015	JNV admits who sat JEE	Avanti, re-derived
JNVST 2018 → same, within school	.140	4,015 (469 schools)	same	Avanti, re-derived
JNVST 2018 → Class 10 maths	.317	4,015	same	Avanti, re-derived
Class 10 board maths → JEE Main 2025	.367	4,015	same	Avanti, re-derived
NCST 2022 (engineering) → JEE Main 2024	.585 (ρ .628)	1,371	NCST 2022 takers who sat JEE	Avanti warehouse
Class 10 board maths → JEE Main 2024	.595	1,371	same	Avanti warehouse
NCST 2024 → JEE Main 2026, JNV CoEs	.311 (ρ .535)	172	Avanti CoE students	Avanti warehouse

The samples differ (doubly selected admits, a self-selected pool, a selected and coached group), so the coefficients are not directly comparable (Section 8).

3.6 Coaching

The best-studied case is the US SAT: gains “more in the neighborhood of 30 points” (Briggs 2009). Meta-analyses of retest and coaching effects on other ability tests exist (Hausknecht et al. 2007; Kulik, Bangert-Drowns & Kulik 1984; Becker 1990); they were not read, and they bear directly on format-naive first-time takers such as JNVST candidates. No source read shows that any talent-search format resists coaching.

In CMS-E 2025 (weighted, Class 11 to 12), 37.5% of rural and 47.9% of urban government-school students took private coaching, at about ₹12,100 a year among the coached, against ₹28,700 for urban private-school students; coaching bundled into junior-college fees is not captured. In JEE Advanced 2016, self-reported coaching rises from 30.3% of registrants to 44.5% of admits, and parental income above ₹8 lakh from 8.9% to 16.4%. Any new instrument can carry a cheap coaching study: randomise familiarisation and measure the gain.

---

4. Who gets missed

4.1 Universal screening

In Broward County (Card & Giuliano 2016, PNAS; Quasi-experimental: the policy’s introduction and removal, matched schools, and a placebo across 22 Florida districts), every second grader from 2005 sat a short nonverbal test in class; those above 130, or 115 for Plan B students, were referred for a full IQ test. Eligibility rules and referral were unchanged.

Third graders identified as gifted	2004-05	2006-07	Trend-adjusted odds ratio
All	3.3%	5.5%	1.45
Plan B (low-income / English learner)	1.4%	4.3%	2.74
Hispanic	2.1%	5.7%	2.18
Black	1.1%	2.7%	1.74
White	5.8%	7.6%	1.12 (n.s.)

Before screening, 13 of 140 schools had no gifted third graders; afterwards none. About 80% of those found only through screening were Plan B, more often Hispanic and children of non-English-speaking parents. Rates fell back when screening was cut.

4.2 Teacher discretion

Holding test scores and other characteristics at their means, Black US students have less than half the predicted probability of White students of assignment to gifted services (2.8% against 6.2%), and, for gifted reading, 6.2% for a Black student with a Black teacher against 2.1% with a non-Black teacher (Grissom & Redding 2016; Correlational). India’s nominated NIAS centre “quickly realized that the students of the centre largely belong to the middle- and upper-middle class” (Kurup 2021).

4.3 England’s gifted and talented register

England ran a G&T policy from 1999 to 2010, and schools recorded a G&T flag annually in national pupil data from 2006 to 2010. Linking one cohort (565,770 pupils), Jerrim & Palma Carvajal (2026) report associations only:

• The most advantaged pupils were 19 pp more likely to be identified, “around six percentage points” after prior achievement.
• Conditional on age-11 attainment, G&T status goes with a 7 pp higher chance of a Russell Group place, 2 pp after GCSE results; in Model 2 predictions, 40.5% against 26.8% in the top SES quartile and 26% against 24% for disadvantaged pupils.
• Earnings associations at 24 to 25 are near zero; for the most disadvantaged, lower benefit claiming (7% against 14%).
• The programme was “discontinued in favour of reallocating funds to support disadvantaged students’ access to higher education”.

The transferable lesson is administrative: because the flag sat in the national record, the programme could be evaluated sixteen years after it ended.

4.4 Local norms, quotas and reserved places

• Local norms with a local programme. Broward schools had to run a gifted class and fill spare seats with their own top scorers; Black and Hispanic non-gifted participants gained about 0.5 SD, persisting to Grade 6 (Card & Giuliano 2016, NBER w22104; Causal).
• Local norms as a rule change who is found, and need universal testing underneath: “districts will find it impossible to take the top 5% or 15% of any group if <100% of the group is tested” (Peters et al. 2019).
• Reserved places. A NSW Department official, quoted by ABC, said 21.5% of 2025 Year 7 acceptances came from four equity groups across fully and partially selective schools; ABC’s analysis puts the lowest socio-educational quartile at about 2% of fully selective schools’ 2024 enrolment, mostly admitted before the 2022 model. The figures measure different things; the fair statement is that reserved places have not yet changed the make-up of fully selective schools. In Singapore, which identified through a Primary 3 screening in schools, GEP students came from about 60% of primary schools (MOE, 2022).
• Statutory routes. Korea (Art. 5(2)) and Taiwan (Art. 46) allow adjusted identification for disadvantaged children; neither has been evaluated.

Where JNVST stands. JNVST already has this architecture: seats per district, rural-open seats block by block, SC and ST seats by district population, 27% OBC, and a one-third floor for girls. Its gap is that it is opt-in. About 22 to 28 lakh children register, and nobody has published how application rates vary by district, gender or category. A short classroom test for every Class 5 child feeding JNVST is the direct Indian analogue of Broward, and is untested.

4.5 Girls

Tail definitions differ by row, so tail sizes are not comparable across rows.

Measure (JNV students)	Girls’ share of takers	Upper tail (definition)	Extreme tail (definition)
CBSE Class 10 Maths Standard, 2022-2026	37.5 to 41.7%	32.6 to 39.0% (top 10% by rank)	31.1 to 35.9% (≥99 marks, 1.3 to 2.4% of takers)
NCST 2026, engineering (present)	34.6%	14.8% (top 10%)	11.0% (top 1%)
NCST 2025, engineering	32.9%	18.9% (top 10%)	7.1% (top 1%)
JEE Main 2024 freshers	32.0%	19.8% (≥90th national percentile)	13.4% (≥99th)

Source: Avanti warehouse. Girls are 48 to 52% of Maths Basic takers against about 42% of the cohort, so the Standard row is conditional on course choice. Nationally, 22.7% of women who sat JEE Advanced 2025 qualified against about 32% of men, and women were 20.15% of IIT admits. Girls took 2 of 24 Indian IMO places in 2023 to 2026. In Broward, girls were 56% of the screened-in against 45% of the earlier gifted cohort, a difference the paper cannot distinguish from zero (SE 0.13).

On NCST 2026, girls skip more questions (26% against 23%) and are less accurate when they answer (62% against 70%). In the NCST 2022 cohort, with NCST total and maths subscore controlled, the female coefficient on JEE is about −0.5 percentile points (SE about 1.0 to 1.1): NCST predicts JEE equally for girls and boys. JEE shares NCST’s format (negative marking, speed), so a format effect common to both is not ruled out. Re-weighting NCST is not supported for predicting JEE; whether it is warranted for identifying talent needs a check against an outcome in another format, such as board marks or college grades. Upstream, among JNV students with 90%+ Class 10 marks, 23% of girls choose the engineering track against 39% of boys, and there 9% of girls qualify for JEE Main against 21% of boys (an internal Avanti analysis). Section 5.8 covers causal Indian evidence on keeping girls in the pipeline.

4.6 Information after identification

A $6 semi-customised information package with fee waivers raised applications and enrolment at matched colleges (Hoxby & Turner 2013; Causal); the College Board’s national version, sent to 785,000 low- and middle-income students in the top half of PSAT/SAT scorers, found “no changes in college enrollment patterns” apart from a 0.02 SD rise in college quality for Black and Hispanic students (Gurantz et al. 2021). A certain, early promise of free tuition at Michigan raised application from 26% to 68% (Dynarski et al. 2021). Low-income US high achievers mostly do not apply to selective colleges (Hoxby & Avery 2013). Information works when it is certain and comes from a trusted institution with human support.

4.7 How much talent is lost

• Bell et al. (2019). Children of top-1% parents are ten times as likely to become inventors as children of below-median parents. In New York City schools, third-grade maths scores explain less than a third of the gap between top-quintile families and the rest. The counterfactual that inventors “would quadruple” if under-represented groups invented at the rate of high-income white men is unconditional on scores, and “does not necessarily imply that aggregate welfare would be higher”.
• Agarwal & Gaulé (2020). IMO participants from low-income countries produce 34% fewer publications than equally scoring peers from rich countries, “a lower bound” (Correlational).
• India. 36% of the top 1,000 JEE scorers and 62% of the top 100 have migrated (Choudhury, Ganguli & Gaulé 2023), which any argument that a national programme for high-potential students pays off for India has to address.

---

5. Nurture: what causally develops talent

5.1 Selective schools: regression-discontinuity evidence

Study	Setting	Result for the marginal admit
Abdulkadiroglu, Angrist & Pathak 2014	Boston, NYC exam schools	“little effect of exam school offers on most students’ achievement”
Dobbie & Fryer 2014	NYC specialised high schools	Harder courses; no gain in college enrolment or graduation
Bui, Craig & Imberman 2014	Large anonymous Southwestern US district, gifted and magnet (RD, lottery)	No achievement effect at the RD margin; magnet lottery raised science only; grades and class rank fall
Card & Giuliano 2014	Florida, IQ-qualified	No effect on reading or maths
Zen 2016	NSW (honours thesis)	“scattered and mostly insignificant”
Lucas & Mbiti 2014	Kenya national schools	Rules out effects of 0.1 SD or more, though peers are 0.5 SD stronger
Pop-Eleches & Urquiola 2013	Romania, about 2,000 cut-offs	+0.05 SD (significant; 0.02 to 0.10 across samples)
Beuermann & Jackson 2018	Barbados, plus meta-analysis	No test effect; long-run well-being index +0.187 SD (+0.298 for women; women’s educational attainment +0.376 SD); meta-analytic test effect about zero
Jackson 2010a	Trinidad and Tobago	Large exam gains, about twice as large for girls
Bucarey et al. 2014	Chile, Instituto Nacional	+0.25 SD on the university entrance test
Dustan, de Janvry & Sadoulet 2017	Mexico City	Dropout +9.4 pp (base 42%), more for weaker students and long commutes; raises end-of-high-school maths scores for the marginal admit
Cabrera-Hernández et al. 2026	Mexico City, long run	Lower formal employment and wages at 5 and 10 years; gaps close by 15

All are RD designs. Three readings follow. At the margin, selection explains most of what elite schools look like: Kenya’s schools’ “sterling reputations reflect the selection of students”. Direct peer quality explains about a tenth of school value-added on average in Trinidad and Tobago, “but at least one-third among the most selective schools” (Jackson 2010b, NBER w16598), so nulls at the cut-off say little about inframarginal admits. Gains appear in attainment more than test scores, and more for girls (Barbados, Trinidad); Mexico City runs the other way, with maths scores up for the marginal admit and completion down. The marginal admit can be harmed: dropout and delayed formal work in Mexico City; worse peer interactions in Romania; and in India, SC and ST students at one elite engineering institution “fall behind their same-major peers”, with evidence of mismatch in selective majors (Frisancho & Krishna 2012; Correlational with selection correction; abstract read).

A rural JNV applicant’s alternative is usually a weak government school, and JNV bundles selection with full boarding. A JNV effect larger than Kenya’s is a hypothesis: Kenya’s null held even for admits with the weakest alternative schools (Lucas & Mbiti 2014). Section 5.3 reviews the causal evidence on boarding schools. No causal study of JNV was found.

5.2 Where programmes raise outcomes

Study	Selection and change	Result	Design
Card & Giuliano 2014/16 (Florida)	High achievers below the IQ cut, separate class	+0.4 to 0.5 SD for low-income and minority students (TOT)	RD
Cohodes 2020 (Boston AWC)	Test score; accelerated curriculum, Grades 4-6	College enrolment +15 pp, marginally significant, concentrated among Black and Latino students	Fuzzy RD
Booij, Haan & Plug 2016 (Dutch gymnasium)	Pull-out project programme	Grades +0.30 SD; more science; higher-return university fields	Fuzzy RD
Canaan, Mouganie & Zhang 2022 (Chinese high school)	Tracked high-achieving class	Maths +0.23 SD; +17 pp highly selective university entry	RD
Cohodes et al. 2024 (US)	Online or residential STEM programme	Online +9 pp at the most competitive colleges; residential an insignificant +5 pp more, and about +9 pp on six-year STEM degrees (p < 0.10). About $15,000 a student for six weeks residential, $2,000 for one-week and online versions	RCT
Lavy & Goldstein 2022 (Israel)	Top 2%; gifted classes	“Tiny” effect on achievement; BA attainment +3.3 pp (n.s.) but earlier (by 25: +9 pp); MA +9 pp; more PhDs (marginal); no earnings effect	Matched comparison

The thread is achievement-based selection, students with weak alternatives, and changed instruction. Where mechanisms are analysed, peers explain little, and authors credit different parts of the instruction: teachers in China, expectations and peer pressure in Florida (the authors’ interpretation), and in Boston a bundle the author does not decompose (Section 5.3). Heller-Sahlgren (2018/19) concludes that most studies from which causal inferences can reliably be drawn “do not reveal that current gifted education programmes on average work as intended”. Both hold: the average programme does little for the average admit, and specific designs do a lot for specific students.

5.3 What does the work: residence, level and pace, or standards

JNVs and Avanti’s CoEs bundle three things: students live at school; teaching runs at a set level and pace (the CBSE syllabus at grade pace in JNVs, accelerated to JEE level in CoEs); and a school culture of standards and expectations. A fourth factor, how weak the school a student would otherwise attend is, shapes all three. No study read separates them in one design. What exists is a set of partial contrasts, each holding one part fixed and varying another.

Study	Setting	What varied	Result	Design
Curto & Fryer 2014	SEED boarding charter, Washington DC	Boarding plus a “No Excuses” school, against DC public schools	+0.198 SD reading, +0.230 SD maths a year; almost all among girls	Lottery
Behaghel, de Chaisemartin & Gurgand 2017	French internats d’excellence	Boarding, smaller classes, supervised study, against the usual school	Over 0.2 SD a year in maths, from year 2 only; top tercile about 0.57 SD a year; other two terciles not significant	Lottery
Shi 2020	North Carolina selective STEM boarding school, Grades 11-12	Admission offer at the cut-off	SAT maths +2 percentile points; gains concentrated among minority, rural and lower-achieving students	RD
Kadzamira et al. 2023	Malawi National boarding secondary schools (roughly the top 5% of PSLCE candidates at the end of Grade 8)	Admission at district-by-sex cut-offs; boarding vs day by value-added	+0.40 SD Grade 10 exam (0.385 to 0.403; attrition lower bound 0.23 to 0.28); more for girls and weak-primary districts. Value-added: National +0.57 SD and District Boarding +0.13 SD vs community day schools	RD; value-added
Ajayi 2014	Ghana selective secondary schools, many boarding	Admission to a more selective school	Marginal gains in completion and exams; boarding goes with completing on time, not with scores	RD; correlational (boarding)
Park, Shi, Hsieh & An 2015	Rural county magnet high schools, Gansu, China (boarding not documented)	Admission at the cut-off	+0.39 SD college entrance exam; +27.8 pp qualifying for college	RD
Foliano, Green & Sartarelli 2019	Christ’s Hospital, England, bursary-funded selective boarding for low-SES high achievers	Against matched grammar-school (selective day, far fewer resources) and independent day-school pupils	+18 pp (30%) on five or more A/A* GCSEs; no difference by SES or sex	Matching
Seftor, Mamun & Schirm 2009	US Upward Bound (87% of projects had a residential summer in 1992, per Moore 1997 as cited in Seftor & Calcagno 2010)	Random assignment	No detectable effect on college enrolment for the average applicant; +6 pp enrolment for low-expectation students	RCT
Seftor & Calcagno 2010	Upward Bound Math-Science, six-week residential summer	Matched comparison	+12.0 pp four-year enrolment; authors warn of selection bias	Matching
Cohodes, Ho & Robles 2022	US STEM summer programmes, one elite host	Six-week residential, one-week residential, online	STEM degree 64%, 64%, 55% (n.s.) against 52%; online raised elite graduation almost as much	RCT (offer)
Zárate 2023	Peru, 25 selective boarding schools	Dorm neighbours’ achievement and social centrality	Academic peer effects zero on average; girls’ reading −0.120 SD per 1 SD stronger neighbour	Field experiment
Clotfelter, Ladd & Vigdor 2015	Charlotte, top two quintiles	Algebra I in Grade 7	Lower Algebra I scores; Algebra II gain in one specification; Geometry n.s.	Policy-induced variation
Canaan, Mouganie & Zhang 2022	One Chinese high school	High-achieving class: faster pace, stronger teachers, smaller class	Maths +0.23 SD; elite university +17 pp	RD
Dobbie & Fryer 2013	35 New York City charters	Five school practices	Practices explain about half of school effectiveness; high expectations adds nothing in maths once the other four are held fixed	Correlational
Angrist, Pathak & Walters 2013	Massachusetts charters	Urban against non-urban	Urban charters raise scores; non-urban middle schools lower them by about 0.16 SD a year in ELA and maths (−0.22 and −0.25 SD one year after the lottery)	Lottery
Guryan et al. 2023	Chicago, Grades 9-10	Daily 2:1 maths tutoring	+0.16 SD and +0.37 SD for those tutored (0.08 SD intent-to-treat in trial 1)	Two RCTs

Residence. Two lotteries show that boarding plus a changed school raises scores. SEED’s maths gain of 0.23 SD a year sits at or just below the 0.26 to 0.54 range that other “No Excuses” charter lotteries report, which the authors read as boarding adding little; the comparison is across studies, not a test. SEED’s gains come almost entirely from girls (0.382 SD reading and 0.265 maths, against −0.138 and 0.037 for boys) and, in reading, from below-median students (0.314 against −0.048). It spends $39,275 a pupil against $20,523 in DC schools. The French internats serve applicants near the 45th percentile; the gain appears only after two years and mostly comes from the top third, and weaker students show none even then. The bundle includes classes about six students smaller and six hours a week of supervised study against 1h15, and costs €21,600 against €10,700, the difference “mostly due to the boarding component”. Well-being fell in year 1, which the authors offer, as “somewhat speculative”, as the reason gains start late. In North Carolina, SAT maths and college-selectivity gains are concentrated among minority, rural and lower-achieving students and those from weaker sending schools; advantaged students show no maths gain but gain as much in college enrolment (5 to 6 pp pooled).

The designs that come closest to isolating residence find little academic effect. In the US STEM trial, one week in residence bought the same STEM-degree gain as six weeks, and the online arm, which included a short campus visit, raised elite graduation almost as much; the residential arms raised self-reported life skills more, by amounts the authors call “not large”, and they attribute the graduation gains mainly to college quality and application information. These are summer programmes, so extrapolating to boarding schools is a stretch. In Peru’s selective boarding schools, randomised dorm neighbours had no academic effect on average and lowered girls’ reading; this tests peer composition among boarders, not boarding. Training dormitory staff in ten rural Chinese primary schools improved welfare and behaviour with no detectable effect on scores, a badly under-powered test (Yue et al. 2014). In rural China, boarders have worse health and Chinese scores in unadjusted comparisons (Wang et al. 2016), and an IV study finds boarding raises memory and attention but has no effect on overall IQ or non-cognitive skills (Chang et al. 2023; the instrument, the school’s share of boarders, may not satisfy the exclusion restriction).

In India, KGBV exposure is associated with a BMI gain of about 1% among underweight girls, from a sample defined by the outcome and with an imprecise placebo (Chatterjee 2020). An unrefereed paper finds that ST women exposed to a nearby EMRS completed about 1.25 fewer years of schooling, which the author attributes to crowding out of day schools and the household cost of residence (Tripathi 2025). The NPEGEL and KGBV girls’ bundle raised girls’ upper-primary enrolment by 6 to 7 pp at the Educationally Backward Block cut-off, but KGBV boarding added about 1% of school capacity and existing day schools absorbed about three quarters of the gain (Meller & Litschig 2015; RD). No causal learning evaluation of JNV, EMRS, the Telangana or Andhra residential societies, KISS, Sainik schools or ashram schools was found in this search. Where outcomes were measured independently they are weak: in CBPS’s own tests, students in Maharashtra’s tribal ashram schools were less likely than private-aided school students to pass or score 60%+ in Classes V and IX, with no difference by management type in other classes (CBPS 2017), and Sainik schools achieved a 9.7% NDA success rate over three years against a 20% annual intake target (CAG, as reported by The Wire 2026). Peru’s high-achievement boarding schools (COAR) have admission-cut-off evaluations that could not be downloaded; they are the most relevant unread source.

Level and pace. The level-matching literature is almost entirely about students who are behind, so it says little about top students. For pace above grade level, the one direct test gives mixed results (Charlotte). In the Chinese high-achieving class the authors cannot quantify pace and think teachers “may be the largest driver”; Boston’s AWC does not isolate its accelerated maths; in Florida the authors had no data on pacing and argue that a level-and-pace match “cannot readily explain” why minority students gained and white students did not.

Standards and expectations. In Dobbie & Fryer’s 35 charters, frequent feedback, data use, tutoring and extra time go with effectiveness; high expectations, coded as academic focus plus behavioural expectations, loses significance in maths once the other four are held fixed, and in reading data use and time lose it too. Lesson plans in effective charters are not more often at or above grade level. Houston’s injection of the five practices raised maths by 0.206 SD a year, pooling randomised and non-randomised schools, while replacing 19 of 20 principals (Fryer 2014). Teacher-expectation effects are real, small, may be larger for stigmatised groups, and expectations may predict outcomes mainly because they are accurate (Jussim & Harber 2005); a model-based estimate puts the elasticity of college completion to teacher expectations at about 0.12 (Papageorge, Gershenson & Kang 2020). One weekly hour of instruction raises PISA scores by 0.15 SD within pupil in OECD countries and about 0.06 SD in developing ones, though the paper’s introduction gives 0.025 (Lavy 2010, NBER w16227). No study comparing CBSE and state-board standards was found in this search.

The alternative. Across all three channels, gains concentrate where the alternative is weak. The same charter model raises scores in cities and lowers them outside, and differences in what compliers would score without a charter account for all of the urban advantage in maths and most in reading (Angrist, Pathak & Walters 2013); that measure mixes family background with school quality. KIPP Lynn helps the weakest students most (Angrist et al. 2010), and so do Shi’s rural and weak-school admits.

Avanti’s data. All JNV groups board, so residence is held fixed. At equal Class 10 maths and science marks, with category and gender controlled, CoE students score +13.9 JEE Main percentile points above Nodal students in 2024 (71 against 467), +16.8 in 2025 (328 against 604) and +15.0 in 2026 (525 against 420); robust standard errors are not clustered by school, so the intervals (10.8 to 17.0, 14.5 to 19.2, 12.2 to 17.7) are too narrow, and this replicates the order of magnitude of an internal Avanti impact analysis (+14.5 in 2025 and +22.3 in 2026, +18.7 pooled over those two years, with school-clustered errors). The contrast carries selection on NCST, eligibility filters (family income, willingness, preferring Avanti’s coaching), dedicated teachers, more tests and different peers. In 2024, the least-selected year, Nodal students are not associated with a detectable difference from JNV students with no Avanti link (−0.9, −2.5 to +0.7), though that group includes some coached students; online students score below Nodal (−3.2 in 2024, −6.5 in 2025), plausibly negative selection (untested). Varying residence is not possible: Delhi’s non-residential Live Class students score like Nodal students at equal CBSE marks, higher in the top bands on 7 to 12 students, but only 32% of Blended Live Class Grade-12 enrolees appeared in JEE 2025. State-board and CBSE marks are not a common scale (students with 95%+ state-board marks average the 44th to 56th JEE percentile, against 82 to 94 for JNV CBSE). Non-admits are not linked, so the data cannot measure the alternative.

What it implies. For JNVs, which combine residence and selection with grade pace, the evidence credits any effect to the school relative to a weak alternative, supervised study time and removed distance, and gives boarding itself no separate credit; this is untested. CoEs add accelerated instruction, frequent testing and dedicated teachers inside an existing residence, and are associated with a large gap whose components are unknown. Testing intensity is a hypothesis: among online JNV students more full-length tests go with qualification (about 7% at none against 22% at three or more; odds ratio about 1.5 a test with controls), concentrated among high-board students, while the CoE lift is largest for low-board students. Non-residential models can carry what has support (teaching at the student’s level, tutoring, testing, time; see above); residence earns its cost where distance or home circumstances would otherwise keep a student out. Two risks need monitoring: year-1 losses for weaker admits and peer effects on lower-achieving girls. What would settle it is assignment between residential and non-residential tracks by lottery or cut-off, with programme intensity recorded per centre.

5.4 Acceleration

Accelerated students “significantly outperformed their nonaccelerated same-age peers (g = 0.70) but did not differ significantly from nonaccelerated older peers (g = 0.09)” (Steenbergen-Hu, Makel & Olszewski-Kubilius 2016). The 0.70 pools three meta-analyses, two drawing almost entirely on the same studies (100% and 88.9% overlap with Rogers 1991, as reported there); the only recent one gives g = 0.40 [−0.05, 0.85]. The studies are mostly quasi-experiments with equivalent groups, published 1932-1974 and 17 of 23 from the 1960s (Kulik 1992), and on our reading a same-age comparison builds in a year of extra content. SMPY participants’ “amount of acceleration did not covary with psychological well-being” at 50 (Bernstein, Lubinski & Benbow 2021; Correlational). Acceleration is efficient and low-risk, which justifies subject acceleration; it does not have “one of the largest effect sizes for any educational intervention”. A Nation Deceived (Colangelo, Assouline & Gross 2004) is advocacy (Expert).

5.5 Grouping high achievers

Grouping helps high achievers only when the curriculum changes with it. US meta-analyses give special grouping for the gifted g = 0.37 and between-class grouping 0.04 to 0.06 (Steenbergen-Hu et al. 2016); Kulik (1992) attributes the between-class null to classes following “the same course of study”, while accelerated classes gained nearly a year. Quasi-experimental (meta-analyses).

5.6 Enrichment, compacting, tutoring, social comparison

• Enrichment. Kim (2016; abstract) reports g = 0.96, largest for summer residential programmes; Warne (2016) argues this is “likely distorted by publication bias”. Cohodes et al. (2024) is partly consistent with Kim’s moderator, for STEM degrees. Treat 0.96 as an upper bound.
• Compacting. A randomised trial across 27 districts (783 students) found 40 to 50% of material could be dropped with no loss on most out-of-level tests (Reis et al. 1993).
• Tutoring. Randomised evidence gives a pooled 0.37 SD (Nickow, Oreopoulos & Quan 2020). Bloom’s “2 sigma” (Bloom 1984) rests on two dissertation experiments with general students.
• Social comparison. Self-concept falls in higher-achieving schools (β = −0.20, Marsh & Hau 2003, abstract; β = −0.28, Fang et al. 2018), more “among high school students, in Asia”. One matched German study found no cost from gifted classes (Preckel et al. 2019, abstract). Self-concept, rank and dropout among girls and lower-ranked admits should be monitored in any residential cohort of high-potential students.

5.7 Scholarships and cash

Study	Design	Population	Result
Kremer, Miguel & Thornton 2009 (Kenya)	RCT	Grade 6 girls; top 15% by exam win fees and a grant (merit)	+0.12 to 0.19 SD; boys and low-pretest girls also gained
Londoño-Vélez et al. 2020, 2026 (Colombia, Ser Pilo Paga)	RD, DiD	Top 9.3% of all test-takers (later tightened), below a wealth threshold	Enrolment up 56.5 to 86.5%; SES gap among high achievers closed; earnings +18 log points after nine years
Duflo, Dupas & Kremer 2021 (Ghana)	RCT lottery	Admitted, not enrolled for lack of fees	Completion +27 pp; tertiary +4.4 pp, mostly women
Barrera-Osorio et al. 2011 (Bogotá)	RCT	Conditional transfers	All arms pooled: tertiary matriculation +23 pp; tertiary-conditioned arm +49.7 pp (self-reported, grade-11 subsample; 2008 NBER WP)

Ser Pilo Paga pays full tuition at any accredited college. NMMSS uses a state Class 8 test and a ₹3.5 lakh income ceiling and pays ₹12,000 a year for Classes 9 to 12 to about 100,000 new awardees a year. Its guidelines make students of Kendriya Vidyalayas, Jawahar Navodaya Vidyalayas, government residential schools and private schools ineligible, so it complements JNV, and an NMMSS RD from state merit lists would estimate effects for government day-school students only. None has been done.

5.8 Information, role models, technology

Four of these five studies are Indian (three randomised, one triple-difference). Information on returns raised schooling by 0.20 to 0.35 years in the Dominican Republic (Jensen 2010); BPO recruiting made Indian girls 3 to 5 pp more likely to be in school (Jensen 2012); reserving council headships for women for two election cycles closed the gender gap in adolescent aspirations by 32% (Beaman et al. 2012); a Haryana school programme shifted gender attitudes by 0.18 SD (Dhar, Jain & Jayachandran 2022); bicycles raised girls’ secondary enrolment in Bihar by 30% (Muralidharan & Prakash 2017). These are candidates for a retention layer for identified girls and first-generation students.

Avanti’s Haryana cluster RCT (240 government senior secondary schools, AEARCTR-0003665) is a caution on technology-led delivery. At about six months, students in ICT schools “performed 0.15 standard deviations lower in mathematics” than control, with no science effect, and observed instructional quality in ICT schools fell by 0.46 SD (de Barros 2020, J-PAL/Avanti Haryana RCT interim report). At the two-year endline, for a second cohort with a 2021 baseline and a changed design, there was no effect of at-home ICT in maths or in-classroom ICT in science (de Barros & Kumar 2024, J-PAL/Avanti Haryana RCT endline report).

5.9 The one Indian trial of selection rules

J-PAL’s Tamil Nadu trial with Art of Problem Solving (AEARCTR-0014090; registry status “in development” as of August 2025; 1,100 schools, 2,400 students) randomises online Pre-Algebra with tutoring, and randomises at school level the rule used to select students (tests, engagement, a custom test, grit), with outcomes to board exams, NMMS, college and STEM by 2030. It is the only study found that tests the selection rule itself. Tamil Nadu has no JNVs, which limits transfer. Anyone designing a national effort should obtain the pre-analysis plan and align outcomes with it.

---

6. International systems

“Outcome evidence” means evidence on what the programme did. The last column records figures checked against primary sources.

System	Structure	Outcome evidence	Checked figures
USA	Federal role limited to the Javits Act, the only federal gifted programme, which does not fund local programmes (NAGC); talent searches; district programmes	Causal RDs and screening (4.1, 5.1)	Duke ended TIP’s Talent Search in 2020 after a review already under way, accelerated when the pandemic closed its residential summer programme, a primary revenue stream (Gifted Atlanta 2020, secondary)
England	G&T register 1999-2010	Selection-on-observables (4.3)	The 19 pp SES gap in identification is raw, about 6 pp after prior achievement; the linked data cover one cohort of 565,770
Germany	LemaS (joint federal-state, €125m), gifted classes, boarding schools, competitions feeding Studienstiftung scholarships	Matched (Preckel et al. 2019)	LemaS is joint federal-state; Sankt Afra is a residential gifted school
Ireland	CTYI: above-level assessment, courses, early university entrance	None causal	CTYI’s CAT programme serves the 85th to 94th percentile
Russia	Kolmogorov school, Sirius, All-Russian Olympiad	Descriptive	About 500 reach the maths final; two IMO wins in 31 participations; “more than 60%” continue at MSU
Türkiye	BİLSEM: nomination, group screening, individual assessment	None causal; about 900,000 nominated for 20,000 places in 2019, so an RD is feasible	355 centres and 67,375 students in May 2022
Australia (NSW)	17 fully and 27 partially selective schools; opportunity classes; up to 20% of places held for four equity groups	Honours-thesis RD	1,840 opportunity-class places; 27 partially selective schools
Singapore	GEP (1984, about 1%, 9 schools); from 2027, school-based provision widened from about 7% to 10% of each cohort (in place since 2007) and modules at 15 centres	None published	MOE describes the widened group as higher-ability learners
South Korea	Gifted Education Promotion Act 2000; gifted and science high schools; nomination-based	Counts only	121,421 (1.87%) in 2013 to 63,167 (1.26%) in 2025: headcount halved while enrolment fell 23%, so the share fell by a third from a 1.91% peak in 2017 (KEDI).
China	SCGY (40 to 50 a year, one-week trial); Olympiad route to guaranteed admission	90.9% of graduates hold a higher degree; not separable from selection	Double Reduction made grade 1-9 tutoring non-profit and stopped new approvals of high-school academic tutoring; it did not eliminate coaching
Taiwan	Special Education Act 1984 (nine amendments), +2 SD bar	Art. 45 requires a follow-up guidance system; no outcome data from it were found	Gifted special-class enrolment fell from 44,568 to 24,490 (2009-2013) as arts classes moved to the Arts Education Act and provision moved to resource rooms and itinerant services, so it does not measure total provision
Israel	Opt-out screening in Grade 2; 20 to 25% to stage 2; top 2% with SES-differential bars; ministry circular	Lavy & Goldstein 2022	20 to 25% of the cohort reach stage 2; about 25% of students in gifted frameworks attend full-time classes (3,907 of 15,309); Lavy & Goldstein 2022 follow students to adulthood
Saudi Arabia	Mawhiba MMCAT, opt-in, 200-riyal fee; 88,431 registered and 79,965 tested in 2025	Pre/post only, no control group (one summer programme; Muammar 2022, abstract)	Opt-in and fee-paying; registrants are about 1.3% of 6.72 million students. By medals 2026 was its best IMO; by points and rank 2022 was stronger

Four cross-cutting points:

1. Legislation does not decide survival. Israel and Singapore ran national identification for about 40 years on ministry policy; Korea kept its statute while participation shrank; England ended its programme to fund widening access. A statute creates entitlements and procedures, and in no case examined guaranteed funding or scale.
2. Universal screening is rare. Israel screens whole cohorts (opt-out); Singapore screens at Primary 3 through its schools; Saudi Arabia’s test is opt-in and fee-paying; Broward screened from 2005 to 2010.
3. Credible outcome evidence exists for a handful of systems and is modest: near zero at the margin, some gains for disadvantaged and achievement-selected students, more on what and how long students study than on earnings.
4. The transferable lesson is administrative. England could evaluate its programme because the flag sat in the national record. For India, that means recording each screening score and the assignment rule, with parental consent. Any Aadhaar- or APAAR-linked tracking needs care: APAAR is Aadhaar-seeded, a national “high-potential” flag would label minors in a government record, the DPDP Act 2023 (s.9) requires verifiable parental consent for children’s data, and NCF-SE warns against “undue attention to students with special talents”.

---

7. India today

7.1 Policy

NEP 2020 para 4.43: “The NCERT and NCTE will develop guidelines for the education of gifted children. B.Ed. programmes may also allow a specialization in the education of gifted children.” Para 4.44 encourages “national residential summer programmes … with a rigorous merit-based but equitable admission process”; para 4.45 asks for Olympiads with “progression from school to local to state to national levels”, in rural areas and regional languages. NCF-SE 2023 (Part B, Section 4.4) reproduces 4.43 and warns that “schools must guard against giving undue attention to students with special talents at the cost of others”. No NCERT/NCTE guideline or B.Ed. specialisation was found. DST’s INSPIRE page describes summer camps for “about 50,000 science students of Class XI” (current scale unverified), and the HBCSE Olympiad ladder exists with unpublished reach.

7.2 Jawahar Navodaya Vidyalayas

Fact	Figure	Source
Schools	665 functional in 27 states and 8 UTs (Sep 2026); 689 sanctioned; none in Tamil Nadu	NVS Class IX prospectus 2027-28; PIB
Seats	At most 80 per school in Class VI, reducible to 40	JNVST 2026 prospectus
Candidates	2018-19: 27.8 lakh registered, 19.9 lakh appeared, 46,400 selected; 2022-23: 28.4, 19.3 lakh, 46,767 (about 2.4% of those appearing)	Careers360 tabulation of NVS reports and RTI (secondary)
Capacity and cost	Student capacity 3,10,517 (Sep 2025); grant-in-aid ₹5,370.79 crore (2024-25), about ₹1.73 lakh per capacity place (years differ; includes some capital)	PIB; own arithmetic
Reservations	Rural-open seats block-wise; SC/ST at district level, interchangeable, over and above open merit; OBC 27%; girls a one-third floor met by preference	Prospectus
Fees	Free board; ₹600 a month in Classes IX-XII for non-exempt students (₹1,500 or the CEA for wards of government employees); girls, SC/ST, Divyang and BPL exempt	Prospectus
Test and eligibility	80 questions, 100 marks (Mental Ability 50, figural; Arithmetic 25; Language 25), sectional minimums, ties broken on Mental Ability; 20 language options; Class V in any recognised school in the district	Prospectus
After Class VI	Regional-language medium to Class VIII, then English for maths and science; CBSE boards; 30% of Class IX students spend a year in a JNV in another language area	NVS Class IX prospectus 2027-28
Lateral entry	Class IX and XI tests. The Class IX test is English or Hindi only, run by an external agency, with seats varying by year (hand counts of the vacancy annexures: roughly 4,000 for 2026-27, 1,500 for 2027-28; no published total)	NVS prospectuses
Teachers	5,083 teaching posts vacant	Lok Sabha reply, via ETV Bharat, Mar 2026
Retention	About 44,000 to 46,000 JNV students a year in Class 10 board files and 34,000 to 36,000 in Class 12 (2022-2024), not reconciled as a retention measure	Avanti warehouse

JNVST is not the only point of entry: lateral-entry tests at Classes IX and XI add seats, and within JNV, NCST at Class 10 is a second national gate for coaching seats (Section 8.2). No causal evaluation of JNV was found, and NITI Aayog DMEO’s 2026 request for proposals to evaluate NVS specifies a mixed-methods design with no counterfactual.

7.3 National talent instruments

• NTSE. Since 1963; Class VIII from 2007, Class X from 2013; per Careers360’s report of an NCERT notice, “stalled till further orders” after March 2021. NCERT lists two third-party evaluations (2017, 2020); its summary reports only awardee and stakeholder ratings, and no comparison group is mentioned (reports not read).
• KVPY was subsumed into INSPIRE in 2022. INSPIRE-MANAK is an idea competition gated by school nomination.
• Olympiads. HBCSE “is the nodal centre of the country for Olympiad programmes in mathematics and sciences including astronomy”, with maths “under the aegis of” NBHM; informatics is not among HBCSE’s listed Olympiad subjects. India’s IMO team ranked 9th, 4th, 7th and 7th in 2023 to 2026.
• CBSE Aryabhata Ganit Challenge. Classes VIII to X in CBSE schools, which include all JNVs and KVs; Stage 1 free, ₹900 per school for Stage 2. It already reaches the JNV population at no test-development cost.
• Ei ASSET Talent Search tests “two levels above” grade for Grades 4 to 8; entry requires a top-15-percentile Ei ASSET score (a school-purchased test), CAT4 stanine 9, a paid online ASSET sitting or school nomination; no validity study.
• NIAS (from 2011): nomination, portfolio, Raven’s and Torrance; about 150 students in centres; no outcome evaluation.
• Others. Dakshana reports 496 of 518 scholars clearing JEE Main 2018 (self-reported). No evaluation of Super 30, Punjab’s Meritorious Schools, Haryana’s Super-100 or ashram schools was found; for KGBV and EMRS, see Section 5.3. Avanti’s CoEs: Section 8.2.

7.4 The Indian evidence

Question	Best Indian evidence	Type
Does JNV raise outcomes?	None	None
Does elite engineering admission pay off for disadvantaged marginal entrants?	“A strong, positive economic return” for the marginal lower-caste entrant, one state (Bertrand, Hanna & Mullainathan 2010); SC/ST students fall behind at one institution (Frisancho & Krishna 2012)	Quasi-experimental; correlational
Where does the right tail go?	Choudhury, Ganguli & Gaulé 2023	Correlational plus natural experiment
Do JNVST and NCST predict JEE?	Section 8	Predictive
Does residential schooling help?	NPEGEL + KGBV girls’ bundle: upper-primary enrolment +6 to 7 pp, mostly via day schools (Meller & Litschig 2015); KGBV exposure and underweight girls’ BMI, about +1% (Chatterjee 2020); EMRS exposure and ST women’s schooling, about −1.25 years (Tripathi 2025, unrefereed); no learning evaluation of any residential school	Causal (RD) for enrolment; quasi-experimental
Does blended video help?	Avanti Haryana: no effect at two-year endline (second cohort)	Causal
Does any school-age talent programme help?	None evaluated; TN/AoPS trial registered	None

Bagde, Epple & Taylor (2016) and Sekhri (2020) were not obtained. India has no causal evidence on any school-age talent programme, and a report should say so and not import global effect sizes.

---

8. Avanti’s own evidence, re-derived

Two audits re-computed Avanti’s figures, one from linked student records (4,079 JNVST 2018 admits matched to Class 10 2023 and JEE Main 2025) and the five-JNV ASSET pilot, the other from Avanti’s data warehouse (NCST 2022-2026, CBSE Class 10 and 12 results for JNV students, and JEE Main). All figures are aggregates. Nothing here is causal.

8.1 JNVST 2018 → JEE Main 2025

What holds. n = 4,015; r = 0.072 [0.041, 0.103], Spearman 0.072; Class 10 maths 0.367. Exact or perfect-confidence matches give 0.069 and 0.063, so match quality does not drive it. Within school r = 0.140, so there is a small, real signal. By section, Mental Ability r = 0.037 [0.006, 0.068] (ρ = 0.022), arithmetic 0.080 [0.049, 0.111], language 0.040; the intervals overlap, but given the nonverbal-test evidence (Section 3.1), the weight on the figural half of JNVST, which is also the tie-breaker, is worth testing before JNVST-style items are reused.

Geography. Across the 12 largest states, mean JNVST and mean JEE correlate −0.75 (95% CI −0.93 to −0.32; ecological): Bihar’s admits average 96.0 on JNVST and 47.9 on JEE, Kerala’s 80.6 and 56.4. Admission is district-wise, so raw scores are not comparable across states, and pooling buries the within-district signal (between-school r = 0.038). Within school, the top JNVST group (25% of students, because of ties in small schools) holds 19 of the 45 students at the national 95th percentile: 1.9% against 1.1% overall, a lift of about 1.7 on small counts.

Range restriction. The Thorndike case-2 correction, r_c = r·u / √(1 − r² + r²u²), gives:

Scenario	u	From pooled r = 0.072	From within-school r = 0.14
Weak effective truncation	1.5	0.11	0.21
	2.0	0.14	0.27
Top 5% of applicants	2.69	0.19	0.35
Top 2% (nominal JNV ratio)	3.00	0.21	0.39

The within-school column is the relevant one. The pooled correlation is held down mainly by differences in score scales between districts, which a truncation correction does not address. Selection runs within district by quota cells with different cut-offs, a multivariate problem that case 2 handles only within one cell, and quotas weaken the effective truncation below the nominal 2%. Selection into JEE-taking (about 9% of admits) is not corrected. These are scenarios; the applicant distribution is not available.

Reading. JNVST ranks within a district carry modest signal for JEE seven years later; inside the admitted pool scores are too compressed and too early to guide talent decisions, and JNVST adds almost nothing once Class 10 marks are known (within-school R² 0.3563 to 0.3570). That is the wrong test of an entry instrument, whose job is to choose among ten-year-olds.

8.2 NCST (Class 10) → JEE Main

NCST 2022 (Dakshana’s applicant pool, 4,494 candidates, about 10% of the JNV Class 10 cohort) is the only cohort not truncated on the outcome (57% qualified). It is still selected twice: it is self-selected (1,339 of 1,410 JEE-linked engineering takers gave Dakshana as first preference), and only about 60% of its engineering takers sat JEE 2024 (1,410 of 2,338; 1,371 with a board mark), non-takers being much weaker on NCST (mean pool percentile 0.37 against 0.59). Both restrict range and probably understate r (correcting for JEE-taking alone, assuming direct selection on NCST, gives about 0.64). Its qualification rate, against about 32% for all JNV 2024 freshers, confirms positive selection. NCST-based coaching pushes the other way. Later cohorts are qualifier-only or partial.

Cohort	n	Qualified	r NCST	r board maths	R² board maths + science + category + gender	+ NCST total and maths
NCST 2022 → JEE 2024	1,371	57.3%	0.585	0.595	0.407	0.438 (0.437, total alone)
NCST 2023 → JEE 2025 (qualifiers only)	168	100%	0.619	0.544	0.489	0.569
NCST 2024 → JEE 2026 (partial file)	515	100%	0.661	0.579	0.442	0.536

In 2022 the Spearman difference between NCST and board maths is +0.031 (95% CI −0.006 to +0.067); one SD of NCST adds 8.1 JEE percentile points and one SD of board maths 8.8. At the top NCST separates students where board maths cannot, because 235 of the 1,371 score 99 or 100:

Group (same 1,371 students)	n	Share reaching JEE ≥95
NCST, top 10% of this sample	143	51.7%
NCST, pool-wide top decile	189	45.5%
NCST, top 235 of this sample	235	42.1%
Board maths ≥99 (all ties)	235	33.6%

Inside the CoEs. CoE seats are allocated on NCST with category floors (Cat 1: 200/200/150/110/100 for GEN/EWS/OBC-NCL/SC/ST, 70 for students with a disability) and a board-maths floor, then ranked by raw NCST in centre priority order (Avanti’s 2026 CoE selection rules). Avanti’s selection analysis records 385 SC and ST students with competitive NCST scores excluded by the maths floor. Among JNV CoE students (NCST 2024 → JEE 2026, n = 172, 169 with a board mark), r = 0.31 [0.17, 0.44] (ρ = 0.54) for NCST against 0.53 (ρ = 0.60) for board maths; the difference in r is imprecise (95% CI −0.46 to 0.00) and small on ranks. CoE students sit around the 80th NCST percentile, the 2026 file is partial (91% qualified here), and coaching may reshuffle students within a band; the data cannot separate these. In the pooled result NCST partly decides who is coached, which inflates the slope. So NCST at Class 10 forecasts JEE, and whether it measures something or routes students to coaching is unknown.

An internal Avanti impact analysis, controlling for Class 10 marks and gender with school-clustered errors, finds CoE students about 14.5 JEE percentile points above Nodal students in 2025 and 22.3 in 2026, about 18.7 pooled (study sample 1,779 across 2024 to 2026), which it calls “an observational comparison, not a randomised trial”. An internal Avanti prediction analysis reports 88 to 91% JEE Main qualification for 501 CoE students in 2026.

Reach was the biggest leak in 2022. Of JNV JEE 2024 freshers at or above the national 95th percentile (n = 872), 75.2% had no linked NCST 2022 record; their median board maths mark was 94 (IQR 90 to 98). About 18% of NCST 2022 records (824 of 4,494) link to no board or exam record; even if all those failed links were JEE sitters, about 70% of these freshers would still have no NCST record. NCST takers were about 1.8 times over-represented among them (24.8% against 13.7% of all freshers), so opt-in drew a stronger group, yet most top scorers were outside it and had already been identified by their boards. This measures reach. NCST then reached 83.4% of the Class 10 cohort in 2025 and about 95% registered in 2026 (43,054 roll numbers; 88.1% present; about 230 Dakshana-track students tested separately are absent). Whether census-scale NCST keeps r ≈ 0.59 will be known from JEE 2027 and 2028; NCST 2025 carries no date of birth or father’s name, so as the data stand it cannot be linked.

8.3 The ASSET pilot (five JNVs, one sitting)

Class	n	Top 5% of norm [95% CI]	Top 20%	Girls in top 20%	Boys in top 20%
6	301	4.3% [2.5, 7.2]	24.6%	19.4%	28.5%
7	339	1.8% [0.8, 3.8]	13.9%	10.3%	16.0%
8	357	2.0% [1.0, 4.0]	9.0%	8.9%	9.0%
9	270	0.4% [0.1, 2.1]	8.9%	4.9%	12.2%
10	343	2.6% [1.4, 4.9]	21.9%	11.5%	28.9%

The fall to 0.4% in Class 9 has been read as a collapse. The arithmetic reproduces; the finding does not hold as stated. It is one child in 270, and Classes 7 to 10 do not differ (χ² p = 0.20). Class 10 rebounds and mean percentile is U-shaped (55.0, 46.1, 41.9, 40.9, 49.9), which across different cohorts in one sitting points first to norming, form difficulty or composition. Two JNV facts add candidates: about 30% of Class IX students spend that year in a JNV in another language area, which fits a one-year dip; and maths and science switch to English from Class IX, which alone would not predict a rebound. The design is cross-sectional, five schools, one sitting whose date is not in the file (to confirm with EI, with test language and who was present). Class 6’s 4.3% is at par with the norm’s 5%. The gender gap is robust pooled (girls 11.0%, boys 18.9% in the top 20%; adjusted odds ratio for boys 2.0, p < 0.001) and imprecise in any one grade. Any estimate of talent lost should wait for EI’s within-student panel.

How the levels compare. ASSET here is maths only and normed separately in each grade (norm median scale score 484 to 487 in every class), so a lower percentile in a later class means standing lower against that grade’s ASSET takers, not less maths, and no grade-equivalent can be computed. EI’s deck does not name the norm group; its only benchmark is the norm itself.

Class	JNV mean percentile (z vs norm)	JNV at or above norm median	Top 20%	Top 5%	NAS 2021 maths, Central Govt % proficient (MH / GJ)	NAS 2021, State Govt (MH / GJ)
6	55.0 (+0.15)	59.5%	24.6%	4.3%	(Class 5: 27.1 / 28.8)
7	46.1 (−0.12)	45.1%	13.9%	1.8%
8	41.9 (−0.26)	38.4%	9.0%	2.0%	30.3 / 28.3	19.2 / 31.3
9	40.9 (−0.29)	37.4%	8.9%	0.4%
10	49.9 (+0.01)	46.9%	21.9%	2.6%	34.8 / 29.4	6.0 / 5.8

• Are the intakes very high potential? Class 6 entrants sit at about the average of the ASSET norm, not at its top. That norm is a fee-paying, school-subscribed, mostly private population whose place in the national distribution has not been measured, so “ASSET top 5%” is not India’s top 5% in either direction. If the norm sat 0.5 to 1.0 SD above the national mean, an average JNV entrant would be around India’s top 16 to 31%; this is an illustration, not an estimate. Descriptive.
• Is the system losing them? Standing is lower in Classes 7 to 9 and back at the norm in Class 10, and the U-shape survives adjustment for school and gender. Schooling-driven erosion would keep falling or level off, not rebound, so cohort, form or norm differences by grade fit at least as well. Only EI’s within-student panel can settle it. Descriptive; cross-sectional.
• Schools differ more than grades. One of the five schools sits about 1 SD below the norm in every class, while the other four sit at or above it (Class 6 mean percentile 58.9).
• Context. NAS is a different test with a low, at-grade proficiency bar, so it is not a level comparison. It does show Central Government schools (which include JNVs) with about six times the State Government proficiency share by Class 10 in both states. PISA 2009+ gives a sense of what a state’s top tail looks like internationally: the 95th percentile of 15-year-olds in mathematics scored 458 in Himachal Pradesh and 468 in Tamil Nadu, below the OECD mean of 496 (Walker 2011; both states did not meet all PISA sampling standards). Descriptive.

8.4 JNV JEE outcomes over time

Coverage varies: the 2023 qualification flag is all false, the 2026 file holds only qualifier freshers, OBC is folded into General in four years, and 8,066 of 12,103 JEE 2025 rows carry no name. National percentile tails are the safer comparison. NTA ranks candidates on their better session, so the national share above a percentile is not exactly nominal; rank at the 90th and 99th percentile over the maximum All India Rank gives about 10.5% and 1.1% (2021 to 2025).

JEE Main year	JNV takers ≥90	≥95	≥99
2021	11.2%	4.9%	0.65%
2022	15.6%	7.2%	0.87%
2023	14.6%	7.0%	0.91%
2024	18.6%	9.7%	1.70%
2025	14.7%	7.0%	1.22%

JNV takers are about 1.1 to 1.75 times the national share at ≥90 and 0.55 to 1.5 times at ≥99. In 2024, 26.9% of girls and 43.9% of boys in Class 12 sat JEE as freshers, and per enrolled Class 12 student girls reached the 95th percentile at 0.95% against 3.8% for boys. Because of these coverage problems, no year-on-year series should be quoted from this file yet.

8.5 Earlier internal estimates

Several figures in Avanti’s earlier internal analyses were re-derived for this review, and the figures above supersede them. The main changes: JNVST does predict JEE, weakly (r = 0.072 pooled, 0.14 within school), where earlier notes said it did not; apparent patterns across JNVST score bands disappear once category is controlled, so JEE percentile is the better outcome; an earlier female-penalty estimate used gender inferred from first names and needs re-running on recorded gender; the Class 9 dip in the ASSET pilot, and any estimate of talent lost built on it, should wait for EI’s within-student panel (Section 8.3); the Bell et al. counterfactual is unconditional on test scores (Section 4.7); and 2.5 lakh JEE Main candidates are about 17 to 18% of the field in 2024 and 2025, where an earlier note put them at 10%. Claims in earlier drafts that could not be traced to a source are not used here.

---

9. Implications and open questions

9.1 What a national report can claim

• Who gets tested strongly shapes who is found (quasi-experimental, one US district; consistent with Avanti’s NCST reach data).
• Identification should be universal, locally normed and repeated.
• A demanding Class 10 test forecasts JEE about as well as board marks, and better at the top (one self-selected Avanti cohort).
• For marginal admits, selective-school attendance adds little to test scores in most settings (causal, many settings).
• Designs with causal gains select on achievement, reach students with weak alternatives and change instruction; which part of the instruction matters is not isolated.
• Boarding has no demonstrated academic effect separate from the school that comes with it, and roughly doubles cost (Section 5.3).
• Scholarships at transitions have a strong causal record in lower- and middle-income countries.
• India has no causal evidence on any school-age talent programme, including JNV.

9.2 What it should not claim

That boarding itself raises achievement, or that JEE-level expectations work without matched instruction; that above-level testing is proven superior to good grade-level or adaptive testing; that acceleration has “the largest effect of any intervention” or enrichment an effect near 1 SD; that early IQ justifies one-time identification; that legislation determines survival; that JNV “loses 93 to 96% of its talent” or JNVST “does not predict” JEE; that UK selection was 19 pp biased “after controlling for ability”; that SMPY produced a Nobel laureate or evaluates a programme; or any Aadhaar- or APAAR-linked tracking or national gifted flag without a consent and data-protection design.

9.3 Knowledge worth creating

Two research paths follow from the evidence. These are research proposals from Avanti Fellows; none has been agreed with NVS, the Ministry of Education or NITI Aayog.

Path 1: validate and expand JNVST.

1. The JNV regression discontinuity. Assignment is sequential and multi-quota (block rural-open seats, district open merit, SC/ST/OBC over and above, a girls’ floor met by preference, sectional minimums, two wait lists), so the design is a fuzzy multi-cut-off RD with many small cells, and cut-offs are unpublished. Feasibility turns on NVS applicant data with non-admits (NVS and Ministry of Education approval, DPDP-compliant consent); on linking non-admits, who scatter to 30-plus state boards and private schools, to be piloted in one or two states; and on outcomes selected by exam-taking, so Class 10 and 12 participation and dropout should be primary, with JEE and NEET scores conditional on sitting and bounded. It could be offered as a module to NITI Aayog DMEO’s NVS evaluation.
2. The opt-in gap. JNVST registration as a share of Class 5 enrolment in all recognised schools (UDISE+ DSP, all managements), by district, gender and category. NVS registration by district is needed; UDISE+ cannot apply the birth window.
3. A universal Class 5 screen feeding JNVST. Broward’s logic transfers. Class 5 schools belong to states, so it needs a state partner (SCERT or Samagra Shiksha), randomised by block or district within partner states.
4. JNVST validity within district, and equating across states, from applicant-level score distributions.
5. An above-level or adaptive add-on for admits in Classes 6 to 8 in the medium of instruction, or at NCST, with a randomised familiarisation arm. The Class IX lateral test is a secondary option: English or Hindi only, externally run and small.
6. The CoE RD on the NCST category floors, to separate what NCST measures from what coaching adds, tracking the 385 SC and ST students excluded by the maths floor to JEE 2028.

Path 2: an “ASER for high-potential students”. A sampled, level-spanning, repeated measure of how many children are at, above and well above grade level, by district, gender, caste and school type. It would give the denominator any fair-share argument needs, a measure of right-tail stability in Indian conditions, and a baseline for JNVST, NCST and ATS reach. Estimating a 2% share to ±0.5 pp at 95% confidence needs about 3,100 children per reporting unit under simple random sampling (√(0.02 × 0.98 / n) = 0.0025); a cluster design inflates this by the design effect (plausibly 2 or more, to be sized in a pilot), so state estimates need several thousand children per state and district estimates are expensive.

Either path. EI’s within-student ASSET panel and ATS → JEE/NEET validation, corrected for range restriction and exam-taking; an NMMSS RD; programme intensity (contact hours, tests a month, syllabus pace) recorded per CoE and Nodal centre, and a lottery or cut-off assignment between residential and non-residential tracks to measure what residence adds; a DIF and format study of NCST by gender and a re-estimate of the gender gap using recorded rather than name-inferred gender; coordination with the TN/AoPS trial; and data infrastructure (screening scores and assignment rules recorded with consent and published; the JEE file fixes in Section 8.4).

9.4 For a national effort

A national effort to find and develop high-potential students would be best served by a plan that states the causal evidence honestly (9.1, 9.2); chooses between Path 1 and Path 2 or sequences them, with the JNV RD first if its feasibility pilot succeeds; and draws on the Indian linked evidence in Section 8, including the Haryana RCT as the one Indian trial of a technology-led secondary model.

9.5 Open questions

• Does census-scale NCST keep its predictive validity (JEE 2027 and 2028), and how much of it is routing into coaching?
• What share of the eligible Class 5 cohort applies to JNVST, by district, gender and category?
• How many JNV admits leave between Class VI and XII, and do those near the cut-off, or far from home, leave more, especially girls?
• Is the ASSET Class 6 to 9 pattern erosion, norming, composition, migration or the medium switch?
• Does the big-fish-little-pond effect appear in CoE cohorts?
• How are teachers assigned within JNVs and CoEs, given about 5,000 vacant posts?
• What does ATS add at the right tail?

---

10. References

Published sources below were downloaded and read in an archived copy. “Full” means the full text was read; “abstract” means only the abstract, and claims are limited to it; “WP” means a working-paper version was read. All were opened on 2026-09-29, except the later groups, opened on 2026-09-30. Avanti’s administrative data and internal analyses were also used; they are not public, and every figure drawn from them is an aggregate re-derived for this review. [download] after a reference is the archived copy that was read and opens only for someone signed in with an avantifellows.org Google account; the DOI or URL in the reference is the official source.

Identification, stability and above-level testing

• Bernstein, B. O., Lubinski, D. & Benbow, C. P. (2021). Academic acceleration in gifted youth and fruitless concerns regarding psychological well-being. Journal of Educational Psychology 113(4):830-845. https://doi.org/10.1037/edu0000500. Full. Opened 2026-09-29. download
• Briggs, D. C. (2009). Preparation for College Admission Exams. NACAC. https://eric.ed.gov/?id=ED505529. Full. Opened 2026-09-29. download
• Dai, D. Y. & Steenbergen-Hu, S. (2015). Special class for the gifted young. Roeper Review 37(1):9-18. https://doi.org/10.1080/02783193.2014.975882. Full. Opened 2026-09-29. download
• Deary, I. J., Whalley, L. J., Lemmon, H., Crawford, J. R. & Starr, J. M. (2000). The stability of individual differences in mental ability from childhood to old age. Intelligence 28(1):49-55. https://doi.org/10.1016/S0160-2896(99)00031-8. Full. Opened 2026-09-29. download
• Kaufmann, E. (2024). Teachers’ judgment accuracy: a replication check by psychometric meta-analysis. PLOS ONE 19(7):e0307594. https://doi.org/10.1371/journal.pone.0307594. Full. Opened 2026-09-29. download
• Lohman, D. F. (2005). Identifying Academically Talented Minority Students. NRC/GT RM05216. Full. Opened 2026-09-29. download
• Lohman, D. F. & Gambrell, J. L. (2012). Using nonverbal tests to help identify academically talented children. Journal of Psychoeducational Assessment 30(1):25-44. https://doi.org/10.1177/0734282911428194. Full. Opened 2026-09-29. download
• Lohman, D. F., Gambrell, J. & Lakin, J. (2008). The commonality of extreme discrepancies in the ability profiles of academically gifted students. Psychology Science Quarterly 50(2):269-282. Full. Opened 2026-09-29. download
• Lohman, D. F. & Korb, K. A. (2006). Gifted today but not tomorrow? Journal for the Education of the Gifted 29(4):451-484. https://doi.org/10.4219/jeg-2006-245. Full. Opened 2026-09-29. download
• Lubinski, D. (2016). From Terman to today. Review of Educational Research 86(4):900-944. https://doi.org/10.3102/0034654316675476. Full. Opened 2026-09-29. download
• Lubinski, D. & Benbow, C. P. (2006). Study of Mathematically Precocious Youth after 35 years. Perspectives on Psychological Science 1(4):316-345. https://doi.org/10.1111/j.1745-6916.2006.00019.x. Full. Opened 2026-09-29. download
• Lubinski, D., Benbow, C. P. & Kell, H. J. (2014). Life paths and accomplishments of mathematically precocious males and females four decades later. Psychological Science 25(12):2217-2232. https://doi.org/10.1177/0956797614551371. Full. Opened 2026-09-29. download
• Makel, M. C., Kell, H. J., Lubinski, D., Putallaz, M. & Benbow, C. P. (2016). When lightning strikes twice. Psychological Science 27(7):1004-1018. https://doi.org/10.1177/0956797616644735. Full. Opened 2026-09-29. download
• McBee, M. T., Peters, S. J. & Miller, E. M. (2016). The impact of the nomination stage on gifted program identification. Gifted Child Quarterly 60(4):258-278. https://doi.org/10.1177/0016986216656256. Full. Opened 2026-09-29. download
• Park, G., Lubinski, D. & Benbow, C. P. (2007). Contrasting intellectual patterns predict creativity in the arts and sciences. Psychological Science 18(11):948-952. https://doi.org/10.1111/j.1467-9280.2007.02007.x. Full. Opened 2026-09-29. download
• Scherrer, V., Breit, M. & Preckel, F. (2025). The stability of students’ academic achievement in school: a meta-analysis. Educational Research Review 48:100687. https://doi.org/10.1016/j.edurev.2025.100687. Abstract. Opened 2026-09-29. download
• Wai, J., Lubinski, D. & Benbow, C. P. (2005). Creativity and occupational accomplishments among intellectually precocious youths. Journal of Educational Psychology 97(3):484-492. https://doi.org/10.1037/0022-0663.97.3.484. Full. Opened 2026-09-29. download
• Wai, J., Lubinski, D. & Benbow, C. P. (2009). Spatial ability for STEM domains. Journal of Educational Psychology 101(4):817-835. https://doi.org/10.1037/a0016127. Full. Opened 2026-09-29. download
• Zhu, J., Cayton, T., Weiss, L. & Gabel, A. (2008). WISC-IV Technical Report #7: Extended Norms. Pearson. Full. Opened 2026-09-29. download

Equity and screening

• Agarwal, R. & Gaulé, P. (2020). Invisible geniuses. AER: Insights 2(4):409-424. https://doi.org/10.1257/aeri.20190457. Full (accepted manuscript). Opened 2026-09-29. download
• Bell, A., Chetty, R., Jaravel, X., Petkova, N. & Van Reenen, J. (2019). Who becomes an inventor in America? QJE 134(2):647-713. https://doi.org/10.1093/qje/qjy028. Full (accepted manuscript). Opened 2026-09-29. download
• Card, D. & Giuliano, L. (2016). Universal screening increases the representation of low-income and minority students in gifted education. PNAS 113(48):13678-13683. https://doi.org/10.1073/pnas.1605043113. Full. Opened 2026-09-29. download
• Card, D. & Giuliano, L. (2016). Can tracking raise the test scores of high-ability minority students? NBER WP 22104 (AER 106(10)). https://www.nber.org/papers/w22104. Full (WP). Opened 2026-09-29. download
• Card, D. & Giuliano, L. (2014). Does gifted education work? For which students? NBER WP 20453. https://www.nber.org/papers/w20453. Full (WP). Opened 2026-09-29. download
• Dynarski, S., Libassi, C. J., Michelmore, K. & Owen, S. (2021). Closing the gap: the effect of reducing complexity and uncertainty in college pricing on the choices of low-income students. AER 111(6). NBER WP 25349. https://www.nber.org/papers/w25349. Abstract and introduction. Opened 2026-09-29. download
• Grissom, J. A. & Redding, C. (2016). Discretion and disproportionality. AERA Open 2(1). https://doi.org/10.1177/2332858415622175. Full. Opened 2026-09-29. download
• Gurantz, O., Howell, J., Hurwitz, M., Larson, C., Pender, M. & White, B. (2021). A national-level informational experiment to promote enrollment in selective colleges. JPAM 40(2). Full (WP with appendices). Opened 2026-09-29. download
• Hoxby, C. & Avery, C. (2013). The missing “one-offs”. Brookings Papers on Economic Activity, Spring. https://doi.org/10.1353/eca.2013.0000. Full (key sections). Opened 2026-09-29. download
• Hoxby, C. & Turner, S. (2013). Expanding college opportunities for high-achieving, low income students. SIEPR DP 12-014. Full (key results). Opened 2026-09-29. download
• Jerrim, J. & Palma Carvajal, M. (2026). The education and labour market outcomes of England’s gifted and talented children. Oxford Review of Education. https://doi.org/10.1080/03054985.2026.2632049. Full text (figures quoted from the running text), plus the authors’ FFT Education Datalab summary of 13 March 2026. Opened 2026-09-29. download 1, download 2
• Kurup, A. (2021). Challenges to identify and mentor gifted children in developing countries: the Indian experience. Current Science 120(3):472-478. https://doi.org/10.18520/cs/v120/i3/472-478; and NIAS reports R37 (2016) and PR/18 (2019). Full. Opened 2026-09-29. download 1, download 2, download 3
• Peters, S. J., Rambo-Hernandez, K., Makel, M. C., Matthews, M. S. & Plucker, J. A. (2019). Effect of local norms on racial and ethnic representation in gifted education. AERA Open 5(2). https://doi.org/10.1177/2332858419848446. Full. Opened 2026-09-29. download

Nurture: acceleration, grouping, selective schools, enrichment

• Abdulkadiroglu, A., Angrist, J. & Pathak, P. (2014). The elite illusion. Econometrica 82(1). NBER WP 17264. https://www.nber.org/papers/w17264. Full (WP). Opened 2026-09-29. download
• Beuermann, D. W. & Jackson, C. K. (2018, revised 2019). The short and long-run effects of attending the schools that parents prefer. NBER WP 24920. https://www.nber.org/papers/w24920. Full (WP). Opened 2026-09-29. download
• Bloom, B. S. (1984). The 2 sigma problem. Educational Researcher 13(6):4-16. https://doi.org/10.3102/0013189X013006004. Full (pp. 4-7, OCR). Opened 2026-09-29. download
• Booij, A. S., Haan, F. & Plug, E. (2016). Enriching students pays off. IZA DP 9757. https://docs.iza.org/dp9757.pdf. Full. Opened 2026-09-29. download
• Bucarey, A., Jorquera, M., Muñoz, P. & Urzúa, S. (2014). El efecto del Instituto Nacional. Estudios Públicos 133. Full. Opened 2026-09-29. download
• Bui, S., Craig, S. & Imberman, S. (2014). Is gifted education a bright idea? AEJ: Policy 6(3). NBER WP 17089. https://www.nber.org/papers/w17089. Full (WP). Opened 2026-09-29. download
• Cabrera-Hernández, F., Dustan, A., Osuna-Gomez, D. & Padilla-Romo, M. (2026). Marginal admission to elite high schools: long-run effects on labor market outcomes. IZA DP 18369. https://docs.iza.org/dp18369.pdf. Full. Opened 2026-09-29. download
• Canaan, S., Mouganie, P. & Zhang, P. (2022). The long-run educational benefits of high-achieving classrooms. IZA DP 15039. https://docs.iza.org/dp15039.pdf. Full. Opened 2026-09-29. download
• Cohodes, S. R. (2020). The long-run impacts of specialized programming for high-achieving students. AEJ: Policy 12(1). Full (author version). Opened 2026-09-29. download
• Cohodes, S., Ho, H., Huffaker, S. & Robles, S. (2024). Residential versus online? Experimental evidence on diversifying the STEM pipeline. AEA Papers and Proceedings 114:507-511. https://doi.org/10.1257/pandp.20241016. Full. Opened 2026-09-29. download
• Colangelo, N., Assouline, S. & Gross, M. (2004). A Nation Deceived, Vols I and II. https://eric.ed.gov/?id=ED535137. Full. Opened 2026-09-29. download 1, download 2
• Dobbie, W. & Fryer, R. G. (2014). The impact of attending a school with high-achieving peers. AEJ: Applied 6(3). NBER WP 17286. https://www.nber.org/papers/w17286. Full (WP). Opened 2026-09-29. download
• Dustan, A., de Janvry, A. & Sadoulet, E. (2017). Flourish or fail? JHR 52(3). Full (author version). Opened 2026-09-29. download
• Fang, J. et al. (2018). The big-fish-little-pond effect on academic self-concept: a meta-analysis. Frontiers in Psychology 9:1569. https://doi.org/10.3389/fpsyg.2018.01569. Full. Opened 2026-09-29. download
• Heller-Sahlgren, G. (2018/19). What Works in Gifted Education? Centre for Education Economics. Full. Opened 2026-09-29. download
• Jackson, C. K. (2010a). Do students benefit from attending better schools? Evidence from rule-based student assignments in Trinidad and Tobago. Economic Journal 120(549):1399-1429. Read as NBER WP 14911 (2009, “Ability-grouping and academic inequality: evidence from rule-based student assignments”). https://www.nber.org/papers/w14911. Full (WP). Opened 2026-09-29. download
• Jackson, C. K. (2010b). Can higher-achieving peers explain the benefits to attending selective schools? Evidence from Trinidad and Tobago. NBER WP 16598, December 2010 (previously circulated as “Peer quality or input quality?”). https://www.nber.org/papers/w16598. Full (WP). Opened 2026-09-29. download
• Kim, M. (2016). A meta-analysis of the effects of enrichment programs on gifted students. Gifted Child Quarterly 60(2):102-116. https://doi.org/10.1177/0016986216630607. Abstract. Opened 2026-09-29. download
• Kulik, J. A. (1992). An Analysis of the Research on Ability Grouping. NRC/GT. https://eric.ed.gov/?id=ED350777. Full. Opened 2026-09-29. download
• Lavy, V. & Goldstein, A. (2022). Gifted children programs’ short and long-term impact. NBER WP 29779. https://www.nber.org/papers/w29779. Full (WP). Opened 2026-09-29. download
• Lucas, A. M. & Mbiti, I. M. (2014). Effects of school quality on student achievement: discontinuity evidence from Kenya. AEJ: Applied 6(3):234-263. https://doi.org/10.1257/app.6.3.234. Full. Opened 2026-09-29. download
• Marsh, H. W. & Hau, K.-T. (2003). Big-fish-little-pond effect on academic self-concept. American Psychologist 58(5):364-376. https://doi.org/10.1037/0003-066X.58.5.364. Abstract. Opened 2026-09-29. download
• Nickow, A., Oreopoulos, P. & Quan, V. (2020). The impressive effects of tutoring on PreK-12 learning. NBER WP 27476. https://www.nber.org/papers/w27476. Full. Opened 2026-09-29. download
• Pop-Eleches, C. & Urquiola, M. (2013). Going to a better school. AER 103(4). NBER WP 16886. https://www.nber.org/papers/w16886. Full (WP). Opened 2026-09-29. download
• Preckel, F. et al. (2019). High-ability grouping. Child Development 90(4):1185-1201. https://doi.org/10.1111/cdev.12996. Abstract. Opened 2026-09-29. download
• Reis, S. M. et al. (1993). The Curriculum Compacting Study. NRC/GT RM 93106. https://eric.ed.gov/?id=ED379847. Full. Opened 2026-09-29. download
• Steenbergen-Hu, S., Makel, M. C. & Olszewski-Kubilius, P. (2016). What one hundred years of research says about the effects of ability grouping and acceleration. Review of Educational Research 86(4):849-899. https://doi.org/10.3102/0034654316675417. Full. Opened 2026-09-29. download
• Warne, R. T. (2016). Publication bias currently makes an accurate estimate of the benefits of enrichment programs difficult. https://eric.ed.gov/?id=ED572464. Full. Opened 2026-09-29. download
• Zen, K. (2016). The impact of selective high schools on student achievement: evidence from New South Wales, Australia. Honours thesis, UNSW School of Economics. Full. Opened 2026-09-29. download

Lower-income countries: learning levels, tracking, scholarships, information, technology

• Barrera-Osorio, F., Bertrand, M., Linden, L. & Pérez-Calle, F. (2011). Improving the design of conditional transfer programs. AEJ: Applied 3(2). Read as NBER WP 13890 (2008), “Conditional cash transfers in education: design features, peer and sibling effects”; figures are from that version. https://www.nber.org/papers/w13890. Full (WP). Opened 2026-09-29. download
• Bau, N., Das, J. & Chang, A. Y. (2021). New evidence on learning trajectories in a low-income setting. IJED 84. Full (World Bank PRWP version). Opened 2026-09-29. download
• Beaman, L., Duflo, E., Pande, R. & Topalova, P. (2012). Female leadership raises aspirations and educational attainment for girls. Science 335:582-586. https://doi.org/10.1126/science.1212382. Full. Opened 2026-09-29. download
• Dhar, D., Jain, T. & Jayachandran, S. (2022). Reshaping adolescents’ gender attitudes. AER 112(3). NBER WP 25331. https://www.nber.org/papers/w25331. Full (WP). Opened 2026-09-29. download
• Duflo, E., Dupas, P. & Kremer, M. (2021). The impact of free secondary education: experimental evidence from Ghana. NBER WP 28937. https://www.nber.org/papers/w28937. Full. Opened 2026-09-29. download
• Jensen, R. (2010). The (perceived) returns to education and the demand for schooling. QJE 125(2):515-548. https://doi.org/10.1162/qjec.2010.125.2.515. Full. Opened 2026-09-29. download
• Jensen, R. (2012). Do labor market opportunities affect young women’s work and family decisions? Experimental evidence from India. QJE 127(2):753-792. https://doi.org/10.1093/qje/qjs002. Read as the earlier version: Economic opportunities and gender differences in human capital: experimental evidence for India, NBER WP 16021 (2010), https://www.nber.org/papers/w16021; figures are from that version. Full (WP). Opened 2026-09-29. download
• Kremer, M., Miguel, E. & Thornton, R. (2009). Incentives to learn. REStat 91(3). NBER WP 10971. https://www.nber.org/papers/w10971. Full (WP). Opened 2026-09-29. download
• Londoño-Vélez, J., Rodríguez, C. & Sánchez, F. (2020). Upstream and downstream impacts of college merit-based financial aid for low-income students. AEJ: Policy 12(2):193-227. https://doi.org/10.1257/pol.20180131. Full. Opened 2026-09-29. download
• Londoño-Vélez, J., Rodríguez, C., Sánchez, F. & Álvarez-Arango, L. E. (2026). Financial aid and upward mobility: evidence from Colombia’s Ser Pilo Paga. JPE 134(7):2119-2165. https://doi.org/10.1086/740226. Full. Opened 2026-09-29. download
• Muralidharan, K. & Prakash, N. (2017). Cycling to school. AEJ: Applied 9(3). NBER WP 19305. https://www.nber.org/papers/w19305. Full (WP). Opened 2026-09-29. download
• Muralidharan, K., Singh, A. & Ganimian, A. J. (2019). Disrupting education? AER 109(4). NBER WP 22923. https://www.nber.org/papers/w22923. Full (WP). Opened 2026-09-29. download
• Singh, A. (2020). Learning more with every year. JEEA 18(4). Full (accepted version). Opened 2026-09-29. download

India: policy, programmes and evidence

• Bertrand, M., Hanna, R. & Mullainathan, S. (2010). Affirmative action in education: evidence from engineering college admissions in India. JPubE 94. NBER WP 13926. https://www.nber.org/papers/w13926. Full (WP). Opened 2026-09-29. download
• CBSE (2025). Circular Acad-69/2025, Aryabhata Ganit Challenge 2025. Full. Opened 2026-09-29. download
• Choudhury, P., Ganguli, I. & Gaulé, P. (2023). Top talent, elite colleges, and migration. Journal of Development Economics 164:103120. https://doi.org/10.1016/j.jdeveco.2023.103120. Full. Opened 2026-09-29. download
• DST. INSPIRE programme and INSPIRE-MANAK pages. Full. Opened 2026-09-29. download 1, download 2
• Educational Initiatives. Ei ASSET Talent Search site. https://ats.ei.study/. Full. Opened 2026-09-29. download
• ETV Bharat (16 Mar 2026). Report of a Lok Sabha written reply on KV and JNV teaching vacancies. Full. Opened 2026-09-29. download
• Frisancho Robles, V. C. & Krishna, K. (2012). Affirmative action in higher education in India: targeting, catch up, and mismatch. NBER WP 17727. https://www.nber.org/papers/w17727. Abstract. Opened 2026-09-29. download
• HBCSE. Olympiad programme pages (about, stages, performance, 2025-26 schedules). https://olympiads.hbcse.tifr.res.in/. Full. Opened 2026-09-29. download 1, download 2, download 3, download 4, download 5
• IMO official. India, Russia and Saudi Arabia country and team results. https://www.imo-official.org/. Full. Opened 2026-09-29. download 1, download 2, download 3, download 4
• JEE (Advanced) 2016 and 2025 Joint Implementation Committee reports. Full. Opened 2026-09-29. download 1, download 2
• J-PAL / AEA RCT Registry. AEARCTR-0014090, Tamil Nadu and Art of Problem Solving. https://doi.org/10.1257/rct.14090. Full (public fields). Opened 2026-09-29. download
• Ministry of Education (2020). National Education Policy 2020, paras 4.43-4.45. https://www.education.gov.in/sites/upload_files/mhrd/files/NEP_Final_English_0.pdf. Full (relevant sections). Opened 2026-09-29. download
• Ministry of Education. NMMSS guidelines. Full. Opened 2026-09-29. download
• NCERT. NTSE information brochure; NTS page and third-party evaluation page (the two evaluation reports themselves not read). Full. Opened 2026-09-29. download 1, download 2, download 3
• NCERT (2023). National Curriculum Framework for School Education, Part B, Section 4.4. Full (section). Opened 2026-09-29. download
• NITI Aayog DMEO (2026). RFP for evaluation of KVS and NVS. Full. Opened 2026-09-29. download
• Navodaya Vidyalaya Samiti. JNVST 2026 Class VI prospectus and Class IX and XI lateral-entry prospectuses 2026-27 (mirror copies); Class IX lateral-entry prospectus 2027-28 (official). Full. Opened 2026-09-29. download 1, download 2, download 3, download 4
• PIB (10 Nov 2025). KVS and NVS driving educational equity, PRID 2188297. https://www.pib.gov.in/PressReleasePage.aspx?PRID=2188297. Full. Opened 2026-09-29. download
• Careers360 (2022). JNVST registration and selection data (NVS reports and RTI); KVPY and NTSE reports. Full (secondary). Opened 2026-09-29. download 1, download 2, download 3, download 4

Country systems (mostly government and programme sources, plus press reports where noted)

• Israel: Knesset Research and Information Center (2017). Full. Opened 2026-09-29. download
• Korea: Gifted Education Promotion Act (KLRI); KEDI Gifted Education Database statistics. Full. Opened 2026-09-29. download 1, download 2, download 3, download 4
• Singapore: MOE press releases (19 Aug 2024; 3 Mar 2026); parliamentary replies (9 Mar 2022; 10 Sep 2024); National Day Rally 2024 transcript. Full. Opened 2026-09-29. download 1, download 2, download 3, download 4, download 5
• Taiwan: Special Education Act (2023 English text) and identification regulations (2026); MOE Statistics Brief No. 20 (2015). Full. Opened 2026-09-29. download 1, download 2, download 3
• China: Double Reduction opinions (2021); MOE bonus-points reform (2014) and special-admissions Q&A (2025); USTC SCGY brochures (2025, 2026). Full. Opened 2026-09-29. download 1, download 2, download 3, download 4, download 5
• Saudi Arabia: Mawhiba pages; Saudi Gazette report on 2025 results; Argaam (2024), Ministry of Education student counts. Full. Opened 2026-09-29. download 1, download 2, download 3, download 4
• Muammar, O. (2022). Impact of Mawhiba’s first award-winning gifted summer program in Saudi Arabia on students’ achievement, skills, and satisfaction. Gifted Education International. https://doi.org/10.1177/02614294221110326. Abstract only. Opened 2026-09-29. download
• Germany: LemaS pages; Sankt Afra; Bundeswettbewerb Mathematik; Studienstiftung. Full. Opened 2026-09-29. download 1, download 2, download 3, download 4, download 5, download 6, download 7, download 8, download 9
• Ireland: CTYI (DCU) programme pages. Full. Opened 2026-09-29. download 1, download 2, download 3, download 4, download 5, download 6
• Russia: SUNC MSU pages; Sirius; All-Russian Olympiad pages. Full. Opened 2026-09-29. download 1, download 2, download 3, download 4, download 5, download 6
• Türkiye: MEB BİLSEM releases (2019, 2020, 9 May 2022, 2025) and 2025 directive. Full. Opened 2026-09-29. download 1, download 2, download 3, download 4, download 5
• Australia (NSW): Department of Education selective schools, OC and fair-access pages; ABC News (20 Aug 2025). Full. Opened 2026-09-29. download 1, download 2, download 3, download 4, download 5, download 6, download 7
• USA: NAGC Javits page; Gifted Atlanta (14 Oct 2020) on Duke TIP (secondary). Full. Opened 2026-09-29. download 1, download 2

Residence, level and pace, standards and expectations (opened 2026-09-30; Section 5.3)

• Angrist, J. D., Dynarski, S. M., Kane, T. J., Pathak, P. A. & Walters, C. R. (2010). Who benefits from KIPP? NBER WP 15740. https://www.nber.org/papers/w15740. Full (WP). Opened 2026-09-30. download
• Angrist, J. D., Pathak, P. A. & Walters, C. R. (2013). Explaining charter school effectiveness. AEJ: Applied 5(4):1-27. https://doi.org/10.1257/app.5.4.1. Figures from NBER WP 17332 (August 2011), https://www.nber.org/papers/w17332. Full (WP). Opened 2026-09-30. download
• Behaghel, L., de Chaisemartin, C. & Gurgand, M. (2017). Ready for boarding? The effects of a boarding school for disadvantaged students. AEJ: Applied 9(1):140-164. https://doi.org/10.1257/app.20150090. Full. Opened 2026-09-30. download
• Chang, F., Huo, Y., Zhang, S., Zeng, H. & Tang, B. (2023). The impact of boarding schools on the development of cognitive and non-cognitive abilities in adolescents. BMC Public Health 23:1852. https://doi.org/10.1186/s12889-023-16748-8. PMC10517520. Full. Opened 2026-09-30. download
• Chatterjee, S. (2020). From better schools to better nourishment: evidence from a school-building program in India. IZA Journal of Labor Policy 10:2. Full. Opened 2026-09-30. download
• Clotfelter, C. T., Ladd, H. F. & Vigdor, J. L. (2015). The aftermath of accelerating algebra: evidence from a district policy initiative. Journal of Human Resources 50(1). NBER WP 18161. https://www.nber.org/papers/w18161. Full (WP). Opened 2026-09-30. download
• Cohodes, S. R., Ho, H. & Robles, S. C. (2022, revised 2024). STEM summer programs for underrepresented youth increase STEM degrees. NBER WP 30227. https://www.nber.org/papers/w30227. Full (WP). Opened 2026-09-30. download
• Curto, V. E. & Fryer, R. G. (2014). The potential of urban boarding schools for the poor: evidence from SEED. Journal of Labor Economics 32(1):65-93. https://doi.org/10.1086/671798. Read as NBER WP 16746 (2011), “Estimating the returns to urban boarding schools: evidence from SEED”, https://www.nber.org/papers/w16746. Full (WP). Opened 2026-09-30. download
• Dobbie, W. & Fryer, R. G. (2013). Getting beneath the veil of effective schools: evidence from New York City. AEJ: Applied 5(4). NBER WP 17632. https://www.nber.org/papers/w17632. Full (WP). Opened 2026-09-30. download
• Fryer, R. G. (2014). Injecting charter school best practices into traditional public schools: evidence from field experiments. WP title: Injecting successful charter school strategies into traditional public schools: a field experiment in Houston. QJE 129(3). NBER WP 17494 (revised December 2013). https://www.nber.org/papers/w17494. Full (WP). Opened 2026-09-30. download
• Guryan, J., Ludwig, J., Bhatt, M. P., Cook, P. J., Davis, J. M. V., Dodge, K., Farkas, G., Fryer, R. G. et al. (2023). Not too late: improving academic outcomes among adolescents. AER 113(3). NBER WP 28531. https://www.nber.org/papers/w28531. Full (WP). Opened 2026-09-30. download
• Jussim, L. & Harber, K. D. (2005). Teacher expectations and self-fulfilling prophecies: knowns and unknowns, resolved and unresolved controversies. Personality and Social Psychology Review 9(2):131-155. Full. Opened 2026-09-30. download
• Lavy, V. (2010). Do differences in schools’ instruction time explain international achievement gaps? Evidence from developed and developing countries. NBER WP 16227 (published Economic Journal 125(588):F397-F424, 2015; the published numbers may differ). https://www.nber.org/papers/w16227. Full (WP). Opened 2026-09-30. download
• Papageorge, N. W., Gershenson, S. & Kang, K. M. (2020). Teacher expectations matter. Review of Economics and Statistics 102(2):234-251. NBER WP 25255 (2018). https://www.nber.org/papers/w25255. Full (WP). Opened 2026-09-30. download
• Shi, Y. (2020). Who benefits from selective education? Evidence from elite boarding school admissions. Economics of Education Review 74. https://doi.org/10.1016/j.econedurev.2019.07.001. ERIC ED602572. Full (author version). Opened 2026-09-30. download
• Tripathi, A. (2025). Impact of residential schools on educational attainment of Indigenous women: evidence from India. Working paper, Department of Economics, University of Calgary, October 2025 (unrefereed). Full. Opened 2026-09-30. download
• Wang, A., Medina, A., Luo, R., Shi, Y. & Yue, A. (2016). To board or not to board: evidence from nutrition, health and education outcomes of students in rural China. China & World Economy 24(3):52-66. Full. Opened 2026-09-30. download
• Yue, A., Shi, Y., Chang, F., Yang, C. & Wang, H. (2014). Dormitory management and boarding students in China’s rural primary schools. China Agricultural Economic Review 6(3):523-550. Full. Opened 2026-09-30. download
• Ajayi, K. F. (2014). Does school quality improve student performance? New evidence from Ghana. Working paper, January 2014. Full. Opened 2026-09-30. download
• CBPS (2017). Reviewing the status of education in tribal areas in Maharashtra: a comprehensive report. Centre for Budget and Policy Studies, Bangalore, June 2017. Full. Opened 2026-09-30. download
• Foliano, F., Green, F. & Sartarelli, M. (2019). Away from home, better at school. The case of a British boarding school. Working paper, 17 June 2019. Full. Opened 2026-09-30. download
• Kadzamira, E., Winiko, S., Kaombe, T. M. & Rossiter, J. (2023). Merit, inequality, and opportunity: the impact of Malawi’s selective secondary schools. CGD Working Paper 673. Full. Opened 2026-09-30. download
• Seftor, N. S., Mamun, A. & Schirm, A. (2009). The impacts of regular Upward Bound on postsecondary outcomes seven to nine years after scheduled high school graduation: final report. Mathematica Policy Research. ERIC ED505850. Full. Opened 2026-09-30. download
• Seftor, N. S. & Calcagno, J. C. (2010). The impacts of Upward Bound Math-Science on postsecondary outcomes 7-9 years after scheduled high school graduation: final report. Mathematica Policy Research. ERIC ED526942. Full. Opened 2026-09-30. download
• Meller, M. & Litschig, S. (2015). Adapting the supply of education to the needs of girls: evidence from a policy experiment in rural India. Barcelona GSE Working Paper 805. Full. Opened 2026-09-30. download
• Park, A., Shi, X., Hsieh, C.-T. & An, X. (2015). Magnet high schools and academic performance in China: a regression discontinuity design. Journal of Comparative Economics 43(4):825-843. Full. Opened 2026-09-30. download
• The Wire Staff (2026, 24 January). CAG report flags shortfall of intake from NCC, Sainik Schools in armed forces. The Wire. https://m.thewire.in/article/government/cag-report-flags-shortfall-of-intake-from-ncc-sainik-schools-in-armed-forces. Full. Opened 2026-09-30. download
• Walker, M. (2011). PISA 2009 Plus results: performance of 15-year-olds in reading, mathematics and science for 10 additional participants. ACER; updated edition August 2012. Partial (executive summary, Table B.3.1). Opened 2026-09-30. download
• Zárate, R. A. (2023). Uncovering peer effects in social and academic skills. AEJ: Applied 15(3):35-79. https://doi.org/10.1257/app.20210583. Full. Opened 2026-09-30. download

Growth, migration and talent loss (opened 2026-09-30; Section 1)

• Abarcar, P. & Theoharides, C. (2024). Medical worker migration and origin-country human capital: evidence from U.S. visa policy. Review of Economics and Statistics 106(1):20-35 (read as 2021 working paper). Full. Opened 2026-09-30. download
• Agrawal, A., Kapur, D. & McHale, J. (2011). Brain drain or brain bank? The impact of skilled emigration on poor-country innovation. Journal of Urban Economics 69(1):43-55 (read as NBER WP 14592). Full. Opened 2026-09-30. download
• Bloom, N., Jones, C. I., Van Reenen, J. & Webb, M. (2020). Are ideas getting harder to find? AER 110(4):1104-1144. https://doi.org/10.1257/aer.20180338. Full. Opened 2026-09-30. download
• Gibson, J. & McKenzie, D. (2011). The microeconomic determinants of emigration and return migration of the best and brightest: evidence from the Pacific. Journal of Development Economics 95(1):18-29. https://doi.org/10.1016/j.jdeveco.2009.11.002 (read as IZA DP 3926). Full. Opened 2026-09-30. download
• Hanushek, E. A. & Woessmann, L. (2008). The role of cognitive skills in economic development. Journal of Economic Literature 46(3):607-668. https://doi.org/10.1257/jel.46.3.607. Full. Opened 2026-09-30. download
• Hanushek, E. A. & Woessmann, L. (2012). Do better schools lead to more growth? Journal of Economic Growth 17(4):267-321. https://doi.org/10.1007/s10887-012-9081-x (read as NBER WP 14633). Full. Opened 2026-09-30. download
• Khanna, G. & Morales, N. (2017). The IT boom and other unintended consequences of chasing the American dream. CGD Working Paper 460. Full. Opened 2026-09-30. download

Haryana RCT reports (J-PAL/Avanti evaluation, AEARCTR-0003665)

• de Barros, A. (2020). Do students benefit from blended instruction? Experimental evidence from public schools in India. Interim report, J-PAL/Avanti Haryana RCT, 15 March 2020. Full. Opened 2026-09-29. download
• de Barros, A. & Kumar, R. (2024). Endline report, J-PAL/Avanti Haryana RCT (Avanti “Sankalp” evaluation), 5 March 2024. Full. Opened 2026-09-29. download

Cited by others but not verified here

These appear in earlier analyses, in other documents reviewed, or in this review as leads, and were not obtained or not read in full. Nothing in this review rests on them.

• Rogers, K. B. (1991), cited via Steenbergen-Hu et al. (2016); Rogers (2004).
• Hausknecht, J. P. et al. (2007); Kulik, J. A., Bangert-Drowns, R. L. & Kulik, C.-L. C. (1984); Becker, B. J. (1990). Coaching and retest meta-analyses, not read.
• Angrist, J. & Rokkanen, M. (2015), JASA. Not obtained.
• Bellei (2009), Chile full-day school; Jacob (2001), graduation exams; Cortes, Goodman & Nomi (2015), JHR (read only as the authors’ own summary of the same study, “A double dose of algebra”, Education Next 13(1), 2013); Hannum, Liu & Wang (2020), China’s rural school consolidation (read, not relied on); Chatterjee (2017), KGBV (abstract only); Nadal-Fernandez, Varghese & Kumar (2026), Telangana residential schools (secondary summary only); the COAR attendance-effect paper (download failed). Not read in full; nothing rests on them. download 1, download 2, download 3, download 4, download 5
• Bagde, Epple & Taylor (2016), AER; Sekhri (2020), AEJ: Applied. Not obtained.
• NCERT third-party evaluations of NTS: Academy of Management Studies (2017); Centre for Market Research & Social Development (2020). Not read.
• Gross (2004); Pfeiffer (2008) handbook chapters; Pfeiffer, Shaunessy-Dedrick & Foley-Nicpon (2018); Neihart et al. (2002/2016); Renzulli (1978); Heller, Perleth & Lim (2005); Gagné (2009). Not obtained or metadata only. download
• Stanley (1991); Reis et al. (1998); Cross, Gust-Brey & Ball (2002); Feng et al. (2005); VanTassel-Baska & Wood (2010); Kulik & Kulik (1984); Marsh et al. (1995); Shaw et al. (2006); Deary, Pattie & Starr (2013); Abstracts only. download 1, download 2, download 3, download 4, download 5, download 6, download 7, download 8
• Kell, Lubinski & Benbow (2013), Psychological Science 24(5):648-659, https://doi.org/10.1177/0956797612457784. Full text archived; not cited. download
• Naglieri & Ford (2003); Südkamp et al. (2012) (read via Kaufmann 2024).
• House of Commons Children, Schools and Families Committee (2010), HC 337; Houng & Justman (2018); Allende & Valenzuela (2024). Blocked.
• Zhang (2016); Roy (2017); Kurup & Maithreyi (2012); Kurup, Chandra & Binoy (2015); Ørberg (2018). Not obtained or not located.
• MOE Singapore letter on GEP screening (2025): the download returned a page shell without the letter.
• Chetty, Deming & Friedman (2023); Jones & Summers (2020); Gentry et al. (2015); Duckworth et al. (2007); Dweck (2006); Raven et al. (2000); Jain & Jain (2022-2024). Cited in earlier analyses; not read.