India context, safety, policy and evaluation

The spec’s access model cannot go ahead as written. Anthropic’s consumer terms set a minimum age of 18, and Anthropic disables consumer accounts that show signs of a minor; the route Anthropic documents for minors is an organisation-built product on the API, with its own safeguards (Anthropic Consumer Terms 2025; Anthropic Age Assurance 2026; Anthropic Minors Guidelines 2026). India’s data law points the same way: from about May 2027 a fiduciary must get verifiable parental consent before processing a child’s data and may not track or behaviourally monitor children, with a narrow carve-out for educational institutions acting for educational purposes (DPDP Act 2023 s.9; DPDP Rules 2025 r.10, r.12, Fourth Schedule). On the evidence side, Indian technology-aided instruction works when it targets each student’s actual level and when an adult keeps time on task high, but in both Mindspark trials and in Avanti’s own RCT the gains did not show up on grade-level exams (Muralidharan et al. 2017; Muralidharan and Singh 2025; Chiplunkar et al. 2020). A one-centre pilot with about 40 students can detect only effects of roughly 0.7 SD on a test score, several times larger than anything this literature supports, so the pilot should be designed to learn about use and proximal outcomes, not to prove impact on JEE.

Evidence table

Only sources I downloaded and read are listed. “SD” means standard deviations of the outcome in the control group unless stated.

Study	Design and population	n	Effect (unit)	What it means for the coach
Chiplunkar, Dhar and Nagesh 2020 (“Too little, too late”)	RCT, 24 government schools in Chennai, Grade 9, Avanti blended after-school programme (videos, peer worksheets, facilitator), one year	1,044 baseline, 880 endline	ASER reading +0.52 SD, ASER math +0.86 SD; no effect on Grade 10 board exam scores; no effect on non-cognitive skills	Avanti’s own evidence that large gains on foundational skills need not move a grade-level high-stakes exam. Measure the outcome you care about, not only a proximal one.
Muralidharan, Singh and Ganimian 2017 (NBER WP 22923; published AER 2019)	RCT (individual lottery), after-school Mindspark centres in Delhi, Grades 6 to 9, 45 min adaptive software + 45 min group instruction, 6 days a week	619	ITT after 4.5 months: math +0.36 SD, Hindi +0.22 SD; IV for 90 days’ attendance: +0.59 and +0.36 SD; attendance 58%; math gains only on below-grade-level questions; no significant gain on grade-level school math exams	Level-targeted practice works in India, but gains sit where the content was pitched. A coach pitched below JEE level will not show up in a JEE score.
Muralidharan and Singh 2025 (NBER WP 34205)	Cluster RCT, 80 public schools in Rajasthan (40 treatment), Grades 1 to 8, Mindspark in school-day computer labs, devices shared in pairs	~6,500 students in treated schools	After 18 months: math +0.22 SD, Hindi +0.20 SD; no significant effect on school exams; usage roughly halved when the lab coordinator was spread across 3 to 4 schools; gains tracked time on platform	Shared devices with individual logins can work. Time on task is the leading indicator, and it falls when the supervising adult is withdrawn.
Banerjee et al. 2017 (TaRL, NBER WP 22931)	Series of RCTs of Pratham’s Teaching at the Right Level in Bihar, Uttarakhand, Haryana, Uttar Pradesh	Several trials	Grouping by level happened in 0 to 4% of regular Bihar classes, but in over 90% of Haryana schools once a dedicated hour and monitoring were in place	Level-targeted teaching fails when it is an add-on with no protected time. The coach needs a protected slot and someone checking.
Rodriguez-Segura 2021 (author draft of WBRO 2022 review)	Review of EdTech studies with causal identification in developing countries	Review	One-laptop-per-child evaluations found null learning effects; self-led learning software and support for instruction are the most effective categories	Devices and access alone do nothing; the software and the routine around it carry the effect.
De Simone et al. 2025 (World Bank PRWP 11125)	RCT, Benin City, Nigeria, first-year senior secondary (~15 years), 12 x 90-min after-school sessions with Microsoft Copilot (GPT-4), students in pairs on shared lab computers, teacher-guided, guardian consent forms signed	657 T / 671 C assigned; 422 T / 337 C took the final test	ITT +0.31 SD on composite test, +0.23 SD on English (abstract), +0.21 SD on the end-of-term curricular English exam; effects larger for higher prior achievers	Closest analogue to the spec: a general LLM, shared devices, adolescents, 6 weeks. Differential attrition is a real caveat.
Henkel et al. 2024 (Rori, Ghana)	Cluster-randomised, 11 schools (5 T, 6 C), Grades 3 to 8, one hour a week on a phone with a WhatsApp math tutor, 8 months, preliminary	637 baseline, 477 endline	+0.36 effect size on math growth	Low-bandwidth, shared-phone delivery can work; few clusters and 25% attrition make the estimate fragile.
Verma et al. 2025 (MILU benchmark, arXiv 2411.02538 v3)	Benchmark of 42 LLMs on exam-style multiple-choice questions in English and 10 Indian languages	79,617 questions	Best model GPT-4o: 81.75% English, 78.62% Hindi, 73.44% Marathi; language-specific Indic models near chance	Accuracy drops outside English, more for Marathi than Hindi. Claude was not tested and the items are not JEE physics.
Fang et al. 2018	Meta-analysis of big-fish-little-pond effect on academic self-concept	33 studies, 56 effects, N = 1,276,838	Mean beta = -0.28; high school -0.32; Asian samples -0.35 (7 effects)	JNV CoE students are in a doubly selected pond. Any feature that makes the local cohort the comparison group pushes self-concept down.
Azmat, Bagues, Cabrales and Iriberri 2016 (IZA DP 9853; Management Science 2019)	RCT, Spanish university, decile rank feedback every 6 months for 3 years	977	Second year: GPA -0.05 SD, 0.4 fewer modules passed (0.1 SD); effect faded by graduation; driven by students who had underestimated their rank; satisfaction rose	Showing rank can lower short-run performance, and the sign depends on what the student believed before. Not a neutral feature.
Murphy and Weinhardt 2013 (CEP DP 1241; REStud 2020)	Observational, administrative data on over 2 million pupils in England; identifies rank effects conditional on test score	>2,000,000	Moving from 25th to 75th percentile of rank within primary school raises later test scores by 0.2 within-pupil SD; mechanism via self-concept	Local rank itself affects later achievement. A within-centre leaderboard assigns half the cohort a low rank.
Evans and Yuan 2020 (CGD WP 545)	Meta-review of education RCTs in low- and middle-income countries	130 RCTs with learning outcomes	Median 0.10 SD; 90th percentile 0.38 SD; South Asia median 0.08 SD; studies with >5,000 students median 0.05 SD vs 0.10 SD for small ones	Sets the prior for what the pilot can expect.
Kraft 2019 (EdWorkingPaper 19-10)	Methods paper on interpreting effect sizes from causal K-12 studies	n/a	Benchmarks: <0.05 small, 0.05 to <0.20 medium, >=0.20 large; researcher-designed aligned tests often give effects 2 to 4 times larger than broad tests	Report chapter-level accuracy and JEE-level outcomes separately; do not sell an aligned-test gain as a JEE gain.
Duflo, Glennerster and Kremer 2007 (NBER t0333)	Methods toolkit for randomised evaluation	n/a	MDE formula; MDE scales with 1/(take-up difference); baseline scores as covariates raise precision; spillovers bias estimates	Basis for the power numbers below.
Anthropic Consumer Terms (effective 8 Oct 2025), Age Assurance (18 May 2026), Minimum Age (16 Mar 2026)	Primary policy text	n/a	Consumer Claude is 18+; accounts with indicators of minor activity are disabled	Students aged 16 cannot hold consumer Claude seats.
Anthropic Guidelines for Organizations Serving Minors (16 Mar 2026); Usage Policy (page states effective 12 Nov 2026)	Primary policy text	n/a	Organisations giving minors access via the API must add age verification, moderation, monitoring, safety education, follow the child-safety system prompt, comply with local law, disclose AI	The compliant route is an Avanti-built product on the API.
Anthropic U.S. K-12 Terms (effective 28 Aug 2026); Commercial Terms (effective 17 Jun 2025)	Primary policy text	n/a	K-12 terms cover US schools and their employees only, and state Anthropic does not collect data directly from students; commercial terms allow products for a customer’s end users and are not for consumer use	No Anthropic education offering I found covers Indian students under 18 holding their own seats.
Digital Personal Data Protection Act 2023, s.2(f), s.9, Schedule	Primary statute	n/a	Child = under 18; verifiable parental consent; no detrimental processing; no tracking or behavioural monitoring; penalty up to Rs 200 crore	The logging layer and any tracking plugin fall inside s.9.
DPDP Rules 2025 (G.S.R. 846(E)) and commencement notification (G.S.R. 843(E)), 13 Nov 2025	Primary regulation	n/a	Rule 10 sets how parental consent is verified; Fourth Schedule Part A exempts an educational institution from s.9(1) and s.9(3) only for tracking and monitoring for its educational activities or child safety; s.9 and Rules 3, 5 to 16 commence 18 months after publication	Plan for compliance by May 2027, or earlier if the timeline is shortened.
NTA Examination Calendar (revised, Dec 2026 cycle onwards)	Primary notice	n/a	JEE Main Session 1: 20 to 24 and 28 Jan 2027; Session 2: 2 to 7 Apr 2027 (tentative)	Fixes the outcome window.
Avanti warehouse, own query (this review)	Correlation of consecutive major-test scores, standardised within test, JNV CoE Grade 12 engineering, academic year 2025-2026	409 students, 27 tests, 4,967 pairs	r = 0.59 between consecutive tests; r = 0.66 from the mean of the previous three tests to the next	The baseline covariate explains about 35 to 43% of variance, which sets the power numbers.

What the evidence says

Level-targeted technology works in India, at the level it targets

The strongest Indian evidence is Mindspark. In the Delhi efficacy trial, winning a lottery for after-school Mindspark centres raised math scores by 0.36 SD and Hindi by 0.22 SD after 4.5 months, with 58% attendance (Muralidharan et al. 2017). The same trial found math gains only on questions below grade level, and no significant effect on grade-level school math exams, because the software pitched content where students actually were. The scaled trial in 80 Rajasthan schools found 0.22 SD in math and 0.20 SD in Hindi after 18 months, and again no significant effect on school exams (Muralidharan and Singh 2025).

Avanti’s own RCT shows the same pattern at Grade 9. The blended programme raised ASER math by 0.86 SD and reading by 0.52 SD, and had no effect on the Grade 10 board exam (Chiplunkar et al. 2020). The authors’ reading is that remediation from a Grade 2 baseline was too little and too late to reach grade-level exam content.

The JNV CoE cohort is different: it is selected on a test, and these students are much closer to grade level than the Delhi or Chennai samples. So the “below grade” gap will be smaller. The lesson still transfers: a coach that sends a student to easier material can raise accuracy on that material without changing a JEE score. The coach has to pitch practice at the edge of JEE-level difficulty for each chapter, and the evaluation has to measure at JEE level.

The adult in the room carries the usage

In both Mindspark trials the programme was blended: software plus a supervising adult. The 2017 paper reports a parallel trial in which group tutoring without technology had no effect, and reads the result as technology raising an instructor’s productivity (Muralidharan et al. 2017). In the 2025 trial, usage fell to about half when each lab coordinator was spread across 3 to 4 schools, and recovered only after implementers saw the usage data and intervened (Muralidharan and Singh 2025). Gains tracked time on the platform.

Teaching at the Right Level shows the same thing from the teacher side. When teachers got training and materials but no dedicated time, 0 to 4% of regular Bihar classes were grouped by level; in Haryana, with a reserved hour and monitoring, over 90% of schools grouped by level (Banerjee et al. 2017). Reviews of EdTech in developing countries find that providing hardware alone, as in one-laptop-per-child programmes, produced null learning effects (Rodriguez-Segura 2021).

Shared devices and low connectivity are workable

The scaled Mindspark trial paired students on one computer where labs were too small, with separate logins and separate headphones (Muralidharan and Singh 2025). The Nigeria LLM trial ran in pairs on lab computers with routers and mobile signal, through frequent internet and power cuts, and still found 0.31 SD on its composite test after six weeks (De Simone et al. 2025). The Ghana Rori trial used one hour a week on a basic phone over low-bandwidth data (Henkel et al. 2024). None of these trials is in JEE preparation, and two of the three (Ghana, Nigeria) have attrition or few clusters that weaken them, but together they show the device constraint is not the binding one. Usage and supervision are.

LLM tutoring in LMICs: early and positive, on easier targets

The Nigeria trial is the closest analogue to the spec: a general-purpose LLM (Microsoft Copilot on GPT-4), adolescents around 15, a teacher circulating, curriculum-aligned prompts, 12 sessions. The ITT effect was 0.31 SD on the study’s own assessment and 0.21 SD on the regular end-of-term English exam, which covered broader content (De Simone et al. 2025). Effects were larger for students with higher prior achievement. Two caveats: only 422 of 657 treatment and 337 of 671 control students sat the final test, and English is a different target from JEE physics problem solving. Guardians signed consent forms before students took part.

Language

On MILU, a benchmark of Indian exam-style questions, the best model tested (GPT-4o) scored 81.75% in English, 78.62% in Hindi and 73.44% in Marathi (Verma et al. 2025). The drop is moderate for Hindi and larger for Marathi. The benchmark is general knowledge, the models are from 2024, and Claude was not evaluated, so this sets a direction, not a size, for the coach. I found no opened source on code-mixed (Hinglish or Marathi-English) mathematical or physics reasoning; that remains a lead.

Selective schools, self-concept and showing ranks

Students in high-achieving schools report lower academic self-concept than equally able students elsewhere. Across 33 studies the effect averages beta = -0.28, and it is larger in high school (-0.32) and in Asian samples (-0.35, 7 effects) (Fang et al. 2018). JNV students are selected at entry and CoE students are selected again, so the local comparison group is strong twice over.

Rank information is not neutral either. In a Spanish university RCT, giving students their decile rank every six months lowered GPA by 0.05 SD and modules passed by 0.1 SD in the following year; the effect faded by graduation and was driven by students who had underestimated their rank (Azmat et al. 2016). In English administrative data, a pupil’s rank within primary school, holding test score fixed, predicts later achievement: moving from the 25th to the 75th percentile of local rank adds 0.2 within-pupil SD, through self-concept (Murphy and Weinhardt 2013). Neither study is Indian or JEE. Together they say local rank affects behaviour, and the direction depends on what students believed before. A JEE aspirant will see a national percentile anyway; the design question is whether the coach adds a local, within-centre comparison on top. The evidence argues against adding one. (That a cutoff-referenced framing avoids the effect is my inference, not a tested result.)

Policy: Anthropic

Anthropic’s Consumer Terms say a user “must be at least 18 years old” or older if local law requires (Anthropic Consumer Terms 2025, s.2). Anthropic’s help centre states that consumer Claude is available only to people over 18, that it detects signals of under-18 use, and that it disables such accounts until age is verified through Yoti (Anthropic Age Assurance 2026). The same terms forbid making an account available to anyone else, which rules out shared logins on centre devices.

For minors, Anthropic’s documented route is an organisation’s own product built on the API. Its guidelines ask such organisations for age verification, content moderation, monitoring and reporting, safety education, use of Anthropic’s child-safety system prompt where provided, public statement of compliance with child-safety and privacy law, and disclosure that users are talking to an AI; Anthropic audits for this (Anthropic Minors Guidelines 2026). The Usage Policy makes these guidelines binding for products serving minors (Anthropic Usage Policy, page dated effective 12 Nov 2026). The U.S. K-12 terms cover US schools, let a school make the service available to “its own employees”, and state that Anthropic does not collect data directly from students (Anthropic K-12 Terms 2026). I found no Anthropic page that permits a Team or Enterprise seat for a user under 18; this needs a written answer from Anthropic before the pilot, not an assumption.

Policy: India’s DPDP Act and Rules

The Act defines a child as anyone under 18 (s.2(f)). Section 9 requires verifiable parental consent before processing a child’s personal data (s.9(1)), bars processing likely to harm a child’s well-being (s.9(2)), and states that a fiduciary “shall not undertake tracking or behavioural monitoring of children” (s.9(3)). Breach of s.9 carries a penalty of up to Rs 200 crore (Schedule, item 3).

Rule 10 requires “appropriate technical and organisational measures” to obtain the parent’s verifiable consent, checking that the parent is an identifiable adult through details already held, details the parent provides, or a token such as DigiLocker (DPDP Rules 2025, r.10). Rule 12 and the Fourth Schedule, Part A, item 3, exempt an educational institution from s.9(1) and s.9(3), with processing “restricted to tracking and behavioural monitoring” for its educational activities or for the safety of enrolled children. Part B, item 3 exempts creating an account only to the extent it is used for email. The s.9(2) duty not to harm a child’s well-being has no exemption.

Timing: the commencement notification brings s.9 into force 18 months after 13 Nov 2025, that is around 13 May 2027, and Rules 3 and 5 to 16 on the same date (G.S.R. 843(E); r.1(4)). So the law does not yet bind a November 2026 pilot, but it will bind any scale-up in academic year 2027-2028. Law-firm reports say MeitY consulted in early 2026 on shortening this window; I did not open them and did not find a notified amendment, so treat the date as unconfirmed.

Two points need a lawyer, not a literature review. First, whether Avanti, which coaches inside JNV campuses, is itself an “educational institution” under the Fourth Schedule definition (“an institution of learning that imparts education”), or whether the exemption sits with the JNV or NVS as fiduciary. Second, whether logging which external sites a student visits, through a Chrome plugin, counts as processing “for the educational activities of such institution”. The plain text supports educational-purpose logging by an institution; it does not obviously support a third-party plugin that records browsing.

Evaluation: what a one-centre pilot can detect

Realistic effects: the median learning effect across 130 LMIC education RCTs is 0.10 SD, 0.08 SD in South Asia, and the 90th percentile is 0.38 SD (Evans and Yuan 2020). Kraft (2019) classes 0.05 to 0.20 SD as medium for causal K-12 studies and warns that tests aligned to the intervention give effects 2 to 4 times larger than broad tests. The best Indian technology results (0.2 to 0.37 SD) came from months of daily use, on content at the students’ level. Ten weeks of 2 to 3 hours a week, measured on JEE Main, should be expected to produce well under 0.2 SD.

Power: using the toolkit formula (Duflo et al. 2007), with 40 students split 20/20, two-sided 5% test and 80% power, the minimum detectable effect is 0.91 SD with no covariate. Avanti’s own data give the baseline covariate: across 409 JNV CoE Grade 12 engineering students in academic year 2025-2026, consecutive major-test scores standardised within test correlate at r = 0.59, and the mean of the previous three tests predicts the next at r = 0.66 (own query, file /Users/akshay/avanti-data/jnv_coe_g12_major_test_retest_r2_2025_26.csv). With R-squared of 0.35 to 0.43 the MDE is 0.69 to 0.73 SD. If only 70% of students assigned to the coach actually use it, the MDE rises to about 0.98 SD, since it scales with one over the take-up difference (Duflo et al. 2007). The power to detect a true 0.15 SD effect is about 9%. At 400 students it would be about 0.21 SD.

A matched comparison centre does not fix this. With one treated and one comparison centre, the treatment is perfectly confounded with everything else that differs between the centres (teachers, timetable, the principal), and there is no way to estimate the between-centre variance. It is a case study with a comparison, useful for spotting large failures, not a test of impact.

Randomising students within the centre is statistically better but has two problems in a residential school. Students share notes, devices and study groups, so control students will be exposed; Duflo et al. (2007) note that spillovers bias the estimate. And teachers may resist a design that withholds a resource from half a batch before JEE.

A design that fits the constraint is randomising within student: for each student, randomly assign which of their eligible target chapters the coach works on, and compare accuracy and attempt rate on those chapters with the student’s own untreated chapters in later major tests. With 40 students and 6 eligible chapters each, the rough MDE is 0.29 to 0.36 SD of chapter-level accuracy, and about 0.23 to 0.28 SD with 10 chapters. These are approximations that assume chapter outcomes are independent within a student after a student fixed effect; the unit is chapter-level, not student-level, SD; and the design measures the coach’s effect on targeted chapters, not on total score. Spillover across chapters (a student who practises more overall) biases it toward zero, which is the safe direction.

What is measurable by JEE Main Session 1 (20 to 28 Jan 2027, about 12 weeks from today; NTA calendar): weekly time on task and plan completion from logs, which Mindspark showed is a valid leading indicator of learning (Muralidharan and Singh 2025); chapter-level accuracy and attempt rate on the regular major tests; and the Session 1 percentile, which can be reported but cannot answer whether the coach worked at n = 40. Session 2 (2 to 7 Apr 2027) gives a second, later outcome.

Where the evidence is weak or contested

• No randomised evidence exists for an LLM coach in JEE or NEET preparation, or for selected high-ability Indian students. Every Indian technology RCT here is Grades 1 to 9 and mostly below grade level.
• The two LMIC LLM trials (Nigeria, Ghana) are short or preliminary, one has large differential attrition (64% vs 50% of assigned students tested), and the other has 11 clusters. Neither is peer-reviewed in the version I read.
• Mindspark numbers differ slightly by version: the NBER working paper I read reports 0.36 and 0.22 SD; the Learning Framework quotes 0.37 and 0.23 SD, which I believe is the AER version but did not open.
• Rank-feedback evidence is mixed in sign. Azmat et al. (2016) cite earlier studies with positive effects of relative feedback for high school students and for Vietnamese students; I did not open those.
• The BFLPE meta-analysis measures self-concept, not exam performance. The link from self-concept to JEE outcome runs through other studies.
• The Indic-language benchmark tests general knowledge in 2024 models. It says little about Claude’s physics or chemistry reasoning in Hindi or Marathi, and nothing about code-mixed input.
• The DPDP Rules are new and untested. How “educational institution” and “tracking and behavioural monitoring” will be read has not been decided by the Data Protection Board or any court.
• Anthropic’s policy pages change. The Usage Policy page I downloaded carries an effective date of 12 Nov 2026, after today, which suggests a published upcoming version; the minors rules should be re-checked before launch.

Design rules for the Avanti coach

1. Do not give students under 18 consumer Claude seats. Build the student experience on the API under Avanti’s commercial agreement, with Anthropic’s minors safeguards: age gate, moderation, logging for safety review, the child-safety system prompt, a public compliance statement, and an AI disclosure at the start of each session (Anthropic Consumer Terms 2025; Anthropic Minors Guidelines 2026). Teachers, who are adults, can use Claude seats with the connector.
2. Get Anthropic’s written answer on whether any Team, Enterprise or education plan may seat Indian students aged 16 to 17 before committing to the seat model (Anthropic K-12 Terms 2026 covers only US schools and employees).
3. Collect verifiable parental consent through the school before the pilot, even though s.9 is not yet in force, using a method that fits Rule 10 (details held by the school, or DigiLocker). The Nigeria trial did this with signed guardian forms (De Simone et al. 2025; DPDP Rules 2025 r.10).
4. Log only what the weekly loop uses (attempts, accuracy, chapters, plan items), state it in the consent notice, and get a legal opinion on whether Avanti or the JNV is the “educational institution” fiduciary for the Fourth Schedule exemption (DPDP Act 2023 s.9(3); DPDP Rules 2025 Fourth Schedule Part A item 3).
5. Drop the browsing-tracking Chrome plugin from the pilot. Pull completion data from platform exports or student self-report confirmed by a mastery check instead (DPDP Act 2023 s.9(3)).
6. Pitch practice at the edge of JEE-level difficulty for each target chapter, not at the student’s comfort level, and judge it on JEE-level items (Muralidharan et al. 2017; Chiplunkar et al. 2020).
7. Protect a timetabled slot and name the adult who is present in it; treat a drop in weekly time on task as the first alarm (Banerjee et al. 2017; Muralidharan and Singh 2025).
8. Allow individual logins on shared devices, and allow pair work on one device where devices are short (Muralidharan and Singh 2025; De Simone et al. 2025).
9. Never show within-centre ranks or leaderboards in the coach. Express standing against criterion targets such as category cutoffs and the student’s own trend (Fang et al. 2018; Murphy and Weinhardt 2013; Azmat et al. 2016).
10. Default technical content to English, accept questions in Hindi, Marathi or mixed language, and audit a sample of non-English explanations each week for errors before trusting them (Verma et al. 2025).
11. Pre-register the pilot as a learning pilot: primary outcomes are time on task, plan completion and chapter-level accuracy on targeted vs untargeted chapters (randomised within student); report the JEE Main percentile descriptively (Duflo et al. 2007; Kraft 2019).
12. Set the success bar against LMIC benchmarks (median 0.10 SD), not against Mindspark’s best results or US undergraduate effects (Evans and Yuan 2020).

Where the current spec conflicts with the evidence

• Access model (s.2). “Students use regular Claude … Claude’s existing young-user safety protections apply.” Regular Claude is 18+, and Anthropic’s protection for minors is to disable the account. The spec’s own “later phase” (short-lived session codes on shared devices, no individual Claude seats) is closer to the compliant design and should become the pilot design.
• Identity (s.7). “Each pilot student gets … their own Claude seat.” Same conflict. The Avanti email addresses are fine; the DPDP email-account exemption covers an account used only for email, not one used to log in to an AI service (DPDP Rules 2025, Fourth Schedule Part B item 3).
• Tracking plugin (s.8b). A plugin that records what students do across external sites is close to the “tracking or behavioural monitoring of children” that s.9(3) bars, and its exemption depends on who the fiduciary is. The spec lists “transparency” as the design question; the law makes consent and purpose limitation the question.
• Success metric (s.8c, s.10). “0.15 to 0.20 SD vs matched control is a realistic pilot target.” With 40 students the MDE is about 0.7 SD, and one matched centre cannot separate the coach from the centre. The target is realistic as an effect; it is not detectable in this pilot.
• Evidence base (s.8c, via the Learning Framework). The framework calls Mindspark evidence from “Indian secondary students”; the 2017 trial was Grades 6 to 9 and the 2025 trial Grades 1 to 8, and neither found gains on grade-level math exams. The 2025 paper’s authors are Muralidharan and Singh, not Muralidharan, Singh and Ganimian.
• Handoff concept, comparison design. “A matched CoE without the coach” and “AIAT score change against the comparison centre” are a two-centre comparison with the same identification problem. Its Friday quiz and level movement are good proximal measures and fit the within-student design.
• Handoff concept, device access. “Teachers only release tablets for tests.” The TaRL and Mindspark evidence says a protected slot with an adult present is the main implementation risk, not a detail to frame around.
• No rankings (s.8c). The spec is consistent with the evidence. Note that the warehouse test tables carry rank and af_rank columns; the connector should not return them to students.

References

1. Chiplunkar, G., Dhar, D., and Nagesh, R. (2020). Too little, too late: Improving post-primary learning outcomes in India. Working paper, 29 June 2020. Archived copy from gs://avantifellows-research-sources/avanti-impact-paper/chiplunkar_dhar_nagesh_2020.pdf. Opened 2026-10-26. Local file: dhar2020_toolittle.pdf.
2. Muralidharan, K., Singh, A., and Ganimian, A. J. (2017). Disrupting education? Experimental evidence on technology-aided instruction in India. NBER Working Paper 22923 (revised July 2017); published in American Economic Review 109(4), 2019. https://www.nber.org/papers/w22923 . Opened 2026-10-26 (archived copy from gs://avantifellows-research-sources/talent-identification/india_programs_evidence__muralidharan2019_mindspark.pdf). Local file: muralidharan2019_mindspark.pdf.
3. Muralidharan, K., and Singh, A. (2025). Adapting for scale: Experimental evidence on technology-aided instruction in India. NBER Working Paper 34205 (Sept 2025, revised Nov 2025). https://www.nber.org/system/files/working_papers/w34205/w34205.pdf . Opened 2026-10-26. Local file: muralidharan2025_adapting.pdf.
4. Banerjee, A., Banerji, R., Berry, J., Duflo, E., Kannan, H., Mukherji, S., Shotland, M., and Walton, M. (2017). From proof of concept to scalable policies: Challenges and solutions, with an application. NBER Working Paper 22931 (revised Oct 2017); Journal of Economic Perspectives 31(4). Opened 2026-10-26 (archived copy from gs://avantifellows-research-sources/talent-identification/level_pace__tarl2017jep.pdf). Local file: banerjee2017_tarl_jep.pdf.
5. Rodriguez-Segura, D. (2021). Educational technology in developing countries: A review of the evidence. Author draft dated 21 June 2021 of the paper published in World Bank Research Observer 37(2), 2022. https://www.ictworks.org/wp-content/uploads/2023/03/Edutech-developing-countries.pdf . Opened 2026-10-26. Local file: rodriguezsegura2021_edtech_draft.pdf.
6. De Simone, M., Tiberti, F., Barron Rodriguez, M., Manolio, F., Mosuro, W., and Dikoru, E. J. (2025). From chalkboards to chatbots: Evaluating the impact of generative AI on learning outcomes in Nigeria. World Bank Policy Research Working Paper 11125 (May 2025, updated Dec 2025). https://documents1.worldbank.org/curated/en/099548105192529324/pdf/IDU-c09f40d8-9ff8-42dc-b315-591157499be7.pdf . Opened 2026-10-26. Local file: desimone2025_nigeria.pdf.
7. Henkel, O., Horne-Robinson, H., Kozhakhmetova, N., and Lee, A. (2024). Effective and scalable math support: Experimental evidence on the impact of an AI-math tutor in Ghana. arXiv:2402.09809. https://arxiv.org/pdf/2402.09809 . Opened 2026-10-26. Local file: henkel2024_rori_ghana.pdf.
8. Verma, S., Khan, M. S. U. R., Kumar, V., Murthy, R., and Sen, J. (2025). MILU: A multi-task Indic language understanding benchmark. arXiv:2411.02538 v3 (4 Feb 2025). https://arxiv.org/pdf/2411.02538 . Opened 2026-10-26. Local file: milu2024_indic.pdf.
9. Fang, J., Huang, X., Zhang, M., Huang, F., Li, Z., and Yuan, Q. (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 . Opened 2026-10-26 (archived copy from gs://avantifellows-research-sources/talent-identification/nurture_programs__fang2018bflpe.pdf). Local file: fang2018_bflpe.pdf.
10. Azmat, G., Bagues, M., Cabrales, A., and Iriberri, N. (2016). What you don’t know… can’t hurt you? A field experiment on relative performance feedback in higher education. IZA Discussion Paper 9853; published in Management Science 65(8), 2019. https://docs.iza.org/dp9853.pdf . Opened 2026-10-26. Local file: azmat2019_rankfeedback.pdf.
11. Murphy, R., and Weinhardt, F. (2013). Top of the class: The importance of ordinal rank. CEP Discussion Paper 1241 (Sept 2013); published in Review of Economic Studies, 2020. https://cep.lse.ac.uk/pubs/download/dp1241.pdf . Opened 2026-10-26. Local file: murphy2020_rank.pdf.
12. Evans, D. K., and Yuan, F. (2020). How big are effect sizes in international education studies? CGD Working Paper 545. https://www.cgdev.org/sites/default/files/how-big-are-effect-sizes-international-education-studies.pdf . Opened 2026-10-26. Local file: evansyuan2022_effectsizes.pdf.
13. Kraft, M. A. (2019). Interpreting effect sizes of education interventions. EdWorkingPaper 19-10, Annenberg Institute at Brown University; published in Educational Researcher, 2020. https://edworkingpapers.com/sites/default/files/Interpreting%20Effect%20Sizes%20-%20August%2019%20FINAL_0.pdf . Opened 2026-10-26. Local file: kraft2019_effectsizes.pdf.
14. Duflo, E., Glennerster, R., and Kremer, M. (2007). Using randomization in development economics research: A toolkit. NBER Technical Working Paper 333. https://www.nber.org/system/files/working_papers/t0333/t0333.pdf . Opened 2026-10-26. Local file: duflo2007_toolkit.pdf.
15. Anthropic. Consumer Terms of Service, effective 8 Oct 2025. https://www.anthropic.com/legal/consumer-terms . Opened 2026-10-26. Local file: anthropic_consumer_terms.html.
16. Anthropic. Age assurance on Claude, help centre article dated 18 May 2026. https://support.claude.com/en/articles/15171100-age-assurance-on-claude . Opened 2026-10-26. Local file: anthropic_age_assurance.html.
17. Anthropic. Minimum age requirement access restriction, help centre article dated 16 Mar 2026. https://support.claude.com/en/articles/13117299-minimum-age-requirement-access-restriction . Opened 2026-10-26. Local file: anthropic_min_age.html.
18. Anthropic. Responsible use of Anthropic’s models: Guidelines for organizations serving minors, dated 16 Mar 2026. https://support.claude.com/en/articles/9307344-responsible-use-of-anthropic-s-models-guidelines-for-organizations-serving-minors . Opened 2026-10-26. Local file: anthropic_minors_helpcenter.html.
19. Anthropic. Usage Policy, page states effective 12 Nov 2026. https://www.anthropic.com/legal/aup . Opened 2026-10-26. Local file: anthropic_aup.html.
20. Anthropic. U.S. K-12 Terms of Service and Data Processing Addendum, effective 28 Aug 2026. https://www.anthropic.com/legal/k12-terms . Opened 2026-10-26. Local file: anthropic_k12_terms.html. Related: Claude for Education page, https://claude.com/solutions/education (local anthropic_education.html), and Claude for Teachers, https://claude.com/solutions/teachers (local claude_teachers.html).
21. Anthropic. Commercial Terms of Service, effective 17 Jun 2025. https://www.anthropic.com/legal/commercial-terms . Opened 2026-10-26. Local file: anthropic_commercial_terms.html.
22. Government of India. The Digital Personal Data Protection Act, 2023 (No. 22 of 2023). https://www.meity.gov.in/static/uploads/2024/06/2bf1f0e9f04e6fb4f8fef35e82c42aa5.pdf . Opened 2026-10-26. Local file: dpdp_act_2023.pdf.
23. Ministry of Electronics and Information Technology. Digital Personal Data Protection Rules, 2025, G.S.R. 846(E), Gazette of India Extraordinary No. 760, 13 Nov 2025. https://www.meity.gov.in/static/uploads/2025/11/53450e6e5dc0bfa85ebd78686cadad39.pdf . Opened 2026-10-26. Local file: dpdp_rules_2025_gsr846.pdf.
24. Ministry of Electronics and Information Technology. Notification G.S.R. 843(E) appointing commencement dates under the DPDP Act, Gazette of India Extraordinary No. 757, 13 Nov 2025. https://www.meity.gov.in/static/uploads/2025/11/c56ceae6c383460ca69577428d36828b.pdf . Opened 2026-10-26. Local file: dpdp_commencement_gsr843.pdf.
25. National Testing Agency. Examination Calendar, December 2026 cycle onwards (revised). https://nta.ac.in/Download/Notice/ExamcalendarRevised.pdf . Opened 2026-10-26. Local file: nta_exam_calendar.pdf.
26. Avanti Fellows warehouse, fact_student_test_results_overall, own query run 2026-10-26: test-retest correlation of major tests, JNV CoE Grade 12 engineering, academic year 2025-2026. Output /Users/akshay/avanti-data/jnv_coe_g12_major_test_retest_r2_2025_26.csv.
27. Supreme Court of India (2025). Sukdeb Saha v. State of Andhra Pradesh, SLP (Crl.) 6378/2024, record of proceedings and guidelines, 25 July 2025. https://api.sci.gov.in/supremecourt/2024/18489/18489_2024_3_1502_62660_Order_25-Jul-2025.pdf . Opened 2026-10-29. Local file: policy_safety/saha2025_sc_judgment.pdf. Mental-health guidelines binding on all educational institutions, including residential ones and hostels, until legislation replaces them: a mental-health policy, counsellors or referral links, displayed helplines, twice-yearly staff training, and no segregation by performance or public shaming.

• Archived copies, with a .txt extraction and SHA-256 sums, are in gs://avantifellows-research-sources/ai-coach/india_context/ (Avanti Google login); see SOURCES.md.

Leads, not opened

• Rodriguez-Segura (2022), published WBRO version, DOI 10.1093/wbro/lkab011. Publisher PDF returned 403; I used the author draft instead.
• Azmat and Iriberri (2010), Journal of Public Economics, on relative-feedback in high school; Tran and Zeckhauser (2012) on rank feedback for Vietnamese students. Known only from Azmat et al. (2016).
• CodeMixBench (arXiv 2505.05063) and MathMist, on code-mixed and multilingual reasoning. Seen in search results only.
• Law-firm and think-tank notes (SS Rana, Feb 2026; Takshashila, Jan 2026) reporting a MeitY proposal to shorten the DPDP compliance window to 12 months. Not opened; no notified amendment found.
• Any Anthropic statement on whether minors may hold Team or Enterprise seats. Not found on any page I opened; needs a direct question to Anthropic.
• Terms of Brilliant.org and Art of Problem Solving on users under 18 and on data collection, relevant to s.8b of the spec. Not checked.
• The AER (2019) published version of Mindspark, which reports slightly different coefficients from the working paper.