How much talent is in the Navodaya system?

Avanti Fellows · Research & Data

How much talent is in the Navodaya system?

The JNV potential for M1-level and M2-level JEE CoEs, NEET B1, CA and CLAT, from the 2026 Class-10 board results and NCST 2026, against today's CoE footprint

Created 7 October 2026 · draft for review

About this paper. This is an internal analysis by Avanti Fellows, shared for transparency. It has not been peer-reviewed. It was prepared with substantial help from AI tools (Anthropic’s Claude), which wrote much of the code and drafted much of the text under the direction of Avanti staff. The numbers are produced by code from the underlying data, and the sources and methods are documented so they can be checked, but errors are possible. Read the findings as preliminary, and contact Avanti Fellows before citing them.

The answer first

For families earning ₹4 lakh or less, the Navodaya system could fill about 51 more CoEs than today’s 50, holding the 2026 Class-10 board marks as they are:

• JEE: +35.4. 55 M1-level and 20 M2-level batches against about 40 JEE batches today. Bhopal, Shillong, Hyderabad and Chandigarh have the largest gaps.
• NEET: +3.2, almost all in Shillong (+10.0). Elsewhere today’s NEET CoEs roughly match the students who could win a government MBBS seat. The B1 numbers come from a separate model, the government-MBBS prediction (Section 5), and the North-East numbers rest partly on indicative cutoffs.
• CA and CLAT: 5 and 7, all new, since no CoE runs either today.

Summary: CoEs needed, for families earning ₹4 lakh or less

Everything in this summary uses Avanti’s selection rule (family income ≤ ₹4 lakh). A CoE here is one 40-seat Class-11 batch. Need is potential minus today’s CoEs; a negative number is a surplus. The overall need adds the four types together within each region, so a JEE surplus can offset a NEET or CA gap in the same row.

How the tables are built:

• M1, M2 and B1 CoEs are batches whose expected rate is at least 50% M1, 70% qualifying, or 50% B1. B1 means a government MBBS seat in the student’s own state, from the government-MBBS model. The NEET tables also show the older JEE-M1 proxy as a cross-check (“proxy bound”, Section 5).
• CA and CLAT are proxies: students with M2-level 10th marks who choose Commerce (CA) or Arts (CLAT), ÷ 40. None of today’s 50 CoEs is a CA or CLAT centre, so their need is the whole potential. Most areas have too few candidates for a CoE of their own. So the national total is rounded up to whole CoEs (CA: 4.2 → 5; CLAT: 6.2 → 7), and they are placed where candidates are densest, one at a time to the area with the most candidates per CoE already placed. Students from elsewhere travel to them.
• Today is the Ops master’s 50 CoEs (Avanti, ENF, Dakshana), each scaled by its own intake. Today’s JEE CoEs are not split by tier, so the JEE need covers M1 and M2 together.
• The two all-India CoEs fill first. Dakshana Bengaluru Urban (3.5 JEE batches) and Dakshana Pune (2.95 NEET batches) recruit nationally. They are assumed to take the country’s top students: the best 140 for JEE and the best 118 medical students. So they are always full, and their need is zero. Every region and state is sized on the students who remain. The two centres sit in their own row, not against their host region.
• At saturation means every potential CoE in the tables below is running: the Class-11 seats are the students placed in JEE (M1, M2) and NEET (B1) batches plus the CA and CLAT candidates. Girls at saturation is their share of those seats if they are filled strictly in merit order. Each type table carries the same measure for its own seats. Avanti fills girls’ seats from a separate list, so the real share would be higher. Today’s actual share is 37% (Section 11).

By NVS region

1. Overall

Region	CoEs needed (all types)	Class-11 seats at saturation	Girls at saturation (merit order)
Bhopal	+19.0	929	29%
Chandigarh	+3.0	189	37%
Hyderabad	+1.2	588	31%
Jaipur	−1.2	111	21%
Lucknow	+3.5	305	25%
Patna	+4.1	377	24%
Pune	−1.0	401	28%
Shillong	+22.0	996	46%
All-India CoEs (Dakshana Bengaluru Urban, Pune)	+0.0	258	—
National	+50.6	4,156	33%

2. JEE CoEs

Region	M1 CoEs	M2 CoEs	Today	Need	Girls at saturation
Bhopal	13	6	5.0	+14.0	22%
Chandigarh	4	0	0.5	+3.5	32%
Hyderabad	5	8	11.8	+1.2	27%
Jaipur	2	0	2.2	−0.2	8%
Lucknow	5	0	2.8	+2.2	11%
Patna	6	1	4.3	+2.7	15%
Pune	4	5	8.0	+1.0	24%
Shillong	13	0	2.0	+11.0	34%
All-India CoEs (Dakshana Bengaluru Urban, Pune)	3.5	—	3.5	+0.0	—
National	55.5	20	40.1	+35.4	24%

3. NEET CoEs

Region	B1 CoEs (govt MBBS)	Proxy bound	Today	Need	Girls at saturation
Bhopal	2	8	1.0	+1.0	56%
Chandigarh	0	2	0.5	−0.5	—
Hyderabad	0	2	2.0	−2.0	—
Jaipur	0	1	1.0	−1.0	—
Lucknow	1	1	1.8	−0.8	55%
Patna	1	3	1.6	−0.6	52%
Pune	0	1	3.0	−3.0	—
Shillong	11	18	1.0	+10.0	58%
All-India CoEs (Dakshana Bengaluru Urban, Pune)	2.95	—	2.95	+0.0	—
National	17.9	36	14.8	+3.2	57%

4. CA CoEs

Region	CA candidates (in CoEs of 40)	CA CoEs placed	Today	Need	Girls at saturation
Bhopal	0.9	2	0	+2	62%
Chandigarh	0.3	0	0	+0	68%
Hyderabad	0.7	1	0	+1	67%
Jaipur	0.3	0	0	+0	60%
Lucknow	0.7	1	0	+1	54%
Patna	0.6	1	0	+1	54%
Pune	0.4	0	0	+0	59%
Shillong	0.4	0	0	+0	62%
National	4.2	5	0	+5	60%

5. CLAT CoEs

Region	CLAT candidates (in CoEs of 40)	CLAT CoEs placed	Today	Need	Girls at saturation
Bhopal	1.3	2	0	+2	55%
Chandigarh	0.4	0	0	+0	63%
Hyderabad	1.0	1	0	+1	60%
Jaipur	0.5	0	0	+0	55%
Lucknow	1.0	1	0	+1	49%
Patna	0.8	1	0	+1	49%
Pune	0.6	1	0	+1	53%
Shillong	0.5	1	0	+1	55%
National	6.2	7	0	+7	55%

By state

1. Overall

State	CoEs needed (all types)	Class-11 seats at saturation	Girls at saturation (merit order)
Madhya Pradesh	+13.0	561	30%
Manipur	+9.0	368	52%
Odisha	+5.0	277	26%
Uttar Pradesh	+4.5	218	21%
Arunachal Pradesh	+4.0	163	48%
Meghalaya	+4.0	161	50%
West Bengal	+4.0	217	29%
Assam	+4.0	220	34%
Chhattisgarh	+3.0	172	37%
Andhra Pradesh	+2.0	176	25%
Tripura	+2.0	83	37%
Punjab	+2.0	93	37%
Himachal Pradesh	+2.0	88	37%
Jammu And Kashmir	+1.0	48	19%
Kerala	+0.8	138	41%
Puducherry	+0.5	41	31%
Bihar	+0.2	108	16%
Uttarakhand	+0.0	48	22%
Telangana	+0.0	172	30%
Haryana	+0.0	56	25%
Maharashtra	+0.0	271	29%
Chandigarh	−1.0	1	74%
Gujarat	−1.0	89	24%
Karnataka	−1.0	141	31%
Nagaland	−1.0	40	55%
Jharkhand	−1.1	52	22%
Rajasthan	−1.2	54	15%
All-India CoEs (Dakshana Bengaluru Urban, Pune)	+0.0	258	—
National	+54.6	4,312	33%

2. JEE CoEs

State	M1 CoEs	M2 CoEs	Today	Need	Girls at saturation
Madhya Pradesh	5	7	2.0	+10.0	26%
Manipur	4	0	0.0	+4.0	34%
Odisha	5	0	2.5	+2.5	16%
Uttar Pradesh	4	0	2.2	+1.8	10%
Arunachal Pradesh	2	0	0.0	+2.0	42%
Meghalaya	2	0	0.0	+2.0	34%
West Bengal	3	1	1.0	+3.0	21%
Assam	3	0	0.5	+2.5	22%
Chhattisgarh	3	0	0.5	+2.5	26%
Andhra Pradesh	1	3	2.0	+2.0	22%
Tripura	1	0	0.0	+1.0	28%
Punjab	2	0	0.0	+2.0	31%
Himachal Pradesh	2	0	0.0	+2.0	34%
Jammu And Kashmir	1	0	0.0	+1.0	12%
Kerala	1	2	2.2	+0.8	37%
Puducherry	1	0	0.5	+0.5	30%
Bihar	2	0	2.2	−0.2	5%
Uttarakhand	1	0	0.5	+0.5	15%
Telangana	1	3	3.0	+1.0	28%
Haryana	1	0	0.5	+0.5	10%
Maharashtra	2	4	6.0	+0.0	25%
Chandigarh	0	0	0.5	−0.5	—
Gujarat	2	0	2.0	+0.0	21%
Karnataka	1	2	4.0	−1.0	26%
Nagaland	1	0	1.5	−0.5	55%
Jharkhand	1	0	1.1	−0.1	12%
Rajasthan	1	0	1.8	−0.8	2%
All-India CoEs (Dakshana Bengaluru Urban, Pune)	3.5	—	3.5	+0.0	—
National	56.5	22	40.1	+38.4	24%

3. NEET CoEs

State	B1 CoEs (govt MBBS)	Proxy bound	Today	Need	Girls at saturation
Madhya Pradesh	1	1	0.0	+1.0	48%
Manipur	5	6	0.0	+5.0	66%
Odisha	1	5	0.5	+0.5	48%
Uttar Pradesh	0	1	1.2	−1.2	—
Arunachal Pradesh	2	3	0.0	+2.0	52%
Meghalaya	2	2	0.0	+2.0	66%
West Bengal	1	2	0.0	+1.0	50%
Assam	2	3	0.5	+1.5	48%
Chhattisgarh	1	1	0.5	+0.5	62%
Andhra Pradesh	0	0	0.0	+0.0	—
Tripura	1	1	0.0	+1.0	45%
Punjab	0	1	0.0	+0.0	—
Himachal Pradesh	0	1	0.0	+0.0	—
Jammu And Kashmir	0	1	0.0	+0.0	—
Kerala	0	1	0.0	+0.0	—
Puducherry	0	0	0.0	+0.0	—
Bihar	0	1	0.5	−0.5	—
Uttarakhand	0	0	0.5	−0.5	—
Telangana	0	0	1.0	−1.0	—
Haryana	0	1	0.5	−0.5	—
Maharashtra	0	1	2.0	−2.0	—
Chandigarh	0	0	0.5	−0.5	—
Gujarat	0	1	1.0	−1.0	—
Karnataka	0	1	1.0	−1.0	—
Nagaland	0	1	0.5	−0.5	—
Jharkhand	0	0	1.1	−1.1	—
Rajasthan	0	0	0.5	−0.5	—
All-India CoEs (Dakshana Bengaluru Urban, Pune)	2.95	—	2.95	+0.0	—
National	18.9	35	14.8	+4.2	57%

4. CA CoEs

State	CA candidates (in CoEs of 40)	CA CoEs placed	Today	Need	Girls at saturation
Madhya Pradesh	0.4	1	0	+1	62%
Manipur	0.1	0	0	+0	66%
Odisha	0.4	1	0	+1	61%
Uttar Pradesh	0.6	2	0	+2	53%
Arunachal Pradesh	0.0	0	0	+0	69%
Meghalaya	0.0	0	0	+0	54%
West Bengal	0.2	0	0	+0	57%
Assam	0.2	0	0	+0	60%
Chhattisgarh	0.1	0	0	+0	64%
Andhra Pradesh	0.2	0	0	+0	63%
Tripura	0.0	0	0	+0	47%
Punjab	0.1	0	0	+0	73%
Himachal Pradesh	0.1	0	0	+0	67%
Jammu And Kashmir	0.1	0	0	+0	56%
Kerala	0.2	0	0	+0	77%
Puducherry	0.0	0	0	+0	53%
Bihar	0.3	0	0	+0	52%
Uttarakhand	0.1	0	0	+0	59%
Telangana	0.1	0	0	+0	60%
Haryana	0.2	0	0	+0	64%
Maharashtra	0.3	1	0	+1	61%
Chandigarh	0.0	0	0	+0	79%
Gujarat	0.1	0	0	+0	50%
Karnataka	0.2	0	0	+0	64%
Nagaland	0.0	0	0	+0	76%
Jharkhand	0.1	0	0	+0	55%
Rajasthan	0.1	0	0	+0	54%
National	4.2	5	0	+5	60%

5. CLAT CoEs

State	CLAT candidates (in CoEs of 40)	CLAT CoEs placed	Today	Need	Girls at saturation
Madhya Pradesh	0.6	1	0	+1	57%
Manipur	0.1	0	0	+0	55%
Odisha	0.5	1	0	+1	53%
Uttar Pradesh	0.9	2	0	+2	48%
Arunachal Pradesh	0.0	0	0	+0	65%
Meghalaya	0.0	0	0	+0	51%
West Bengal	0.3	0	0	+0	49%
Assam	0.3	0	0	+0	54%
Chhattisgarh	0.2	0	0	+0	59%
Andhra Pradesh	0.2	0	0	+0	58%
Tripura	0.0	0	0	+0	43%
Punjab	0.2	0	0	+0	69%
Himachal Pradesh	0.1	0	0	+0	63%
Jammu And Kashmir	0.1	0	0	+0	51%
Kerala	0.2	0	0	+0	69%
Puducherry	0.0	0	0	+0	48%
Bihar	0.4	1	0	+1	48%
Uttarakhand	0.1	0	0	+0	53%
Telangana	0.2	0	0	+0	54%
Haryana	0.2	0	0	+0	59%
Maharashtra	0.5	1	0	+1	55%
Chandigarh	0.0	0	0	+0	72%
Gujarat	0.1	0	0	+0	46%
Karnataka	0.3	1	0	+1	59%
Nagaland	0.0	0	0	+0	74%
Jharkhand	0.2	0	0	+0	51%
Rajasthan	0.2	0	0	+0	50%
National	6.1	7	0	+7	54%

By state, batches are filled within each state rather than each region. The state totals therefore differ from the region totals (53 M1-level batches against 52), and small states rarely fill a batch on their own. States and UTs with no potential batch and no CoE today are left out.

At a glance: students at each level, and the CoEs they could fill. Each cell shows students, with CoEs in brackets (students ÷ 40).

Family income	M1	M2	B1 (govt MBBS)	CA (proxy)	CLAT (proxy)	JEE CoEs (M1 + M2)
All students	2,492 (62)	1,013 (25)	715 (18)	235 (6)	342 (9)	88
≤ ₹8 lakh	2,268 (57)	932 (23)	650 (16)	198 (5)	289 (7)	80
≤ ₹4 lakh (Avanti rule)	1,907 (48)	812 (20)	556 (14)	170 (4)	248 (6)	68

How a student is counted, coached at CoE level:

• M1: a chance of reaching M1 of at least 50%.
• M2: not M1, with a chance of qualifying of at least 70%.
• B1: a medical-stream student with a chance of at least 50% of scoring above the government MBBS bar for their own state and category (Section 5).
• CA / CLAT (proxy): M2-level 10th marks plus a Commerce or Arts choice. There are no CA or CLAT outcomes, so this stands in for them.

Today’s 50 CoEs hold about 40 JEE batches and 14.8 NEET batches a year. The B1 students are medical-stream students: they are the pool for NEET CoEs, and are not counted in the JEE columns.

• JEE, M1 level. The strongest students fill 71 batches of 40 in which at least half are expected to reach M1. Those batches yield about 2,041 M1.
• JEE, M2 level. A further 26 batches qualify JEE at 70% or better, without a high M1 rate. They yield about 831 qualifiers.
• Today’s footprint. The 50 CoEs of 2026-27 (Avanti, ENF and Dakshana) hold about 40 JEE batches a year. Filled with the best students in the country, they would yield about 1,322 M1. The system supports more than twice today’s JEE capacity (97 batches against 40) before batch quality falls below these bars.
• NEET B1 (a government MBBS seat in the student’s own state). 715 students clear it at 50% or better, about 18 CoEs (556 and 14 at ≤ ₹4 lakh), against about 14.8 NEET batches today. Under the ₹4 lakh rule the regions short of NEET CoEs need about 11 more, and most of them are in Shillong (Section 5).
• CA and CLAT. At M2-level performance, 92%-99% of JNV students choose Science. That leaves about 235 CA candidates (Commerce) and 342 CLAT candidates (Arts) nationally.

With Avanti’s income rule, family income ≤ ₹4 lakh, the pool is 34,129 students. The M1 ceiling falls to about 2,707, and the batches to 55 M1-level and 20 M2-level (75). At ≤ ₹8 lakh it is 3,144 M1, with 65 M1-level and 24 M2-level batches (89). Even under the ₹4 lakh rule, the system supports well over twice today’s 40 JEE batches (Section 4.5).

On the method: Model B ranks the newly selected Class 11 in line with their first 12 internal tests (Section 7). Across the whole NCST pool, NCST adds information at the very top of the curve. But once added to Model B for students we have already selected, it does not improve the forecast (Section 8).

Methodology and assumptions

Every choice behind the numbers above is listed here in one place. Each one can be changed in sources/assumptions.yaml and the analysis rerun; the sections that follow explain them in more depth.

The students

#	Assumption	Why
1	The pool is every JNV student who sat CBSE Class 10 in 2026: 45,170. Their 10th marks are taken as given.	“No intervention in grades 9 and 10” means holding these marks fixed.
2	Board marks are the verified CBSE results, not NCST’s self-reported marks.	The self-reported marks average 59 in maths, against 85-plus on the boards for the same students.
3	Stream (engineering or medical), category (including EWS), family income and state come from NCST 2026. The 4,906 students who did not sit NCST have no stream and are counted only where stated.	NCST is where students declare stream and income.
4	JEE needs Maths Standard. The 2,954 engineering students who took Maths Basic are left out.	Under CBSE rules they cannot take Maths in Class 11.
5	Region comes from the board file, and state is the JNV’s state in NCST.	These are the only geographies recorded for every student.
6	Income filters: NCST family income ≤ ₹4 lakh (Avanti’s selection rule) or ≤ ₹8 lakh. A missing income counts as excluded, as it is in selection.	The ₹4 lakh rule is the one CoE selection uses (Section 4.5).

The outcomes and the model

#	Assumption	Why
7	M1 and M2 use the bar for the student’s home state (their JNV’s state) and category, from M1 and M2 by home state: M1 is the loosest IIT/NIT/IIIT seat open to a student from that state, M2 a seat at a good government college (in the state-exam states, qualifying JEE Main). M2 here is reaching the M2 bar but not M1. States with no state bar keep the national category bars. Section 4.6 compares the two.	The national bars are benchmarks, not seats: a student from a small state with its own NIT can win a seat far below them, and in the north a good government seat is an NIT seat. Validation keeps the national bars, because the outcomes it tests were defined with them.
8	Every student is scored with Model B, assuming CoE-level coaching.	Model B is our best-validated forecast (Section 10). The question asks what CoE coaching could achieve.
9	A student below Model B’s training floor (board 69) is scored as if at 69.	Model B applies the same floor to its own forecast (data-assistant#422). A board mark below the range the model was trained on would otherwise be extrapolated.
10	B1 is a government MBBS seat in the student’s own state. Every medical-stream student is scored with the government-MBBS model: NEET marks out of 720, judged against the in-state government MBBS bar for their domicile and category. Where that bar sits at the qualifying gate, expected seats are capped at the government seats open to the state’s students. The JEE-M1 proxy is kept as a cross-check.	Our students cannot afford private colleges, and in-state seats are the ones they chase. BDS is left out: in small states its bar collapses to the qualifying gate.
11	CA and CLAT are proxies. A candidate has M2-level 10th marks (an aggregate of at least 85.8, the bottom quarter of JEE M2-batch students) and chooses Commerce (CA) or Arts (CLAT). The chance of each choice comes from 2023-25 JNV students in the same marks band, separately for girls and boys.	There are no CA or CLAT outcomes. Stream choice for the 2026 cohort is not yet recorded.

Counting CoEs

#	Assumption	Why
12	A CoE is one 40-seat Class-11 batch. A standard CoE is approved for 80 students across XI and XII.	That is the Ops master’s standard intake.
13	Batches are filled best-first within each region (or state). A batch counts as M1-level while its expected M1 share is at least 50%. The remaining students then form M2-level batches while their expected qualifying share is at least 70%. B1-level batches need an expected B1 share of at least 50%. A last partial batch is dropped.	At these bars a batch is clearly worth a CoE. Section 4.4 varies them: the M1/M2 split moves, but the JEE total barely does.
14	In the at-a-glance student counts, a student counts at a level by their own chance: 50%+ for M1 or B1, or 70%+ to qualify for M2. CoEs there are students ÷ 40.	This uses the same bars, so it agrees with the batch counts.
15	CA and CLAT CoEs: the national candidate total is rounded up to whole CoEs (CA 4.2 → 5, CLAT 6.2 → 7). They are placed one at a time, each in the area with the most candidates per CoE already placed. Students elsewhere travel to them.	Few areas have enough candidates for a CoE of their own.

Today’s CoEs

#	Assumption	Why
16	Today’s CoEs are the 50 in the Ops master “NVS CoEs 26-27 (ALL)”: Avanti, ENF and Dakshana. Centres starting in 2026-27 are included, and one is still awaiting approval.	It is the only list covering all three agencies.
17	Each centre counts as approved strength ÷ 2 (for XI and XII) ÷ 40 batches. A JEE/NEET centre is split half and half, using the current-year stream.	Larger centres (Dakshana 280 and 236, Bundi 100) count for more than one batch.
18	The two all-India Dakshana CoEs fill first with the country’s top students: 140 for JEE (Bengaluru Urban) and 118 medical students (Pune). They are always full, so their need is zero. Every region and state is sized on the students who remain.	These centres recruit nationally and select the strongest candidates (Akshay, 7 Oct 2026).
19	Need = potential − today, per type. A negative number is a surplus. The overall need adds the types within a region, so a surplus in one type offsets a gap in another.	It shows where to add CoEs, and where there are already more than the students can fill.

Saturation and girls

#	Assumption	Why
20	Saturation means every potential CoE is running. Its seats are the students placed in JEE and NEET batches plus the CA and CLAT candidates.	This is the size of the system with no limit on centres.
21	Girls at saturation is girls’ share of those seats in strict merit order.	It shows what merit alone gives. Avanti fills girls’ seats from a separate list, so the real share is set by policy and is higher.
22	Girls in today’s CoEs comes from the Ops master’s current-session strength (XI + XII, 2025-26), all agencies.	It is the only girl count covering ENF and Dakshana as well as Avanti.

What is not modelled: students’ willingness, coaching preference and travel; the regional or state intake rules of individual CoEs (beyond the all-India pair); and any change in grades 9-10. Each of these would lower the numbers that can actually be reached.

1. The question, precisely

“What is the total potential in the Navodaya system without a significant intervention to improve performance in grades 9 and 10?”

We read “without intervention in grades 9 and 10” literally: the 2026 Class-10 board marks are taken as given. The question is then what those students could reach with the best coaching we run in Classes 11 and 12. The 2026 marks have already happened, so this is a statement about one real cohort, not a projection of future cohorts. The same arithmetic applies to any year’s board file.

Outcomes and how each is measured:

Outcome	Definition	Source of the bar
M1	JEE Main percentile at or above the bar for the student’s home state and category (M1 and M2 by home state). The national M1b floor (Gen 97.23, EWS 94.96, OBC 92.68, SC 80.48, ST 75.33; the bar the golden record’s nit_q uses) is the fallback and the comparison in Section 4.6.	m1-m2-bars state bars; dim_nta_cutoffs, 2026
Qualify	At or above NTA’s category qualifying cutoff (Gen 93.41, EWS 82.42, OBC 80.92, SC 63.92, ST 52.02)	dim_nta_cutoffs, 2026
M2	Qualifies, but below M1	the two bars above
B1	NEET marks at or above the government MBBS bar for the student’s own state (domicile) and category. Seats are capped where the bar sits at the qualifying gate. See the government-MBBS model.	neet_dim_marks_matrix_2026 (College DB); NMC seat matrix
CA / CLAT	M2-level 10th performance, opting for Commerce (CA) or Arts (CLAT)	Akshay’s definition, 7 Oct 2026

A batch is one year’s Class-11 intake at a CoE, 40 seats. A standard CoE is approved for 80 students across XI and XII.

2. The pool

	Students
JNV students who sat CBSE Class 10 in 2026	45,170
linked to an NCST 2026 record (stream known)	40,264 (engineering 21,087, medical 19,177)
not in NCST (no stream declared)	4,906, of whom 332 have Maths Standard and a board average of 90+
engineering students with Maths Standard (eligible for JEE)	18,133
engineering students who took Maths Basic (cannot take Maths in Class 11)	2,954
medical students scored for NEET (board marks present)	19,159

Board marks are the verified CBSE 2026 results (jnv_fact_board_results_10th), linked to NCST through the identity map. We do not use NCST’s own grade-10 column: it is self-reported before the boards and averages 59 in maths, against the board’s 85-plus for the same students.

3. Method, through one student

Take an OBC girl from a JNV with 92 in maths and 88 in science, so a board average of 90.

1. Model B predicts her JEE percentile as a distribution, not a point. (Model B is the hierarchical model in jee-main-prediction, described in its methodology note and fitted on 2,063 past CoE and Nodal students from 2024-26, first JEE attempts only.) Coached at CoE level, her median is the 88.4th percentile. The spread is wide: Model B’s uncertainty is about ±1.1 on the logit scale, because two students with the same marks routinely land 20 points apart.

2. The distribution gives her probability of clearing each bar.

   • M1 (OBC M1b 92.68): P = 0.32.
   • Qualifying (OBC 80.92): P = 0.71.
   • Coached only at Nodal level, the same student has a 0.06 chance of M1 and 0.28 of qualifying. The gap between those two pairs is what a CoE adds for her.

3. Counts are sums of probabilities. A batch of 40 students like her is expected to produce 40 × 0.32 ≈ 13 M1 and 28 qualifiers. Nobody is rounded to “will” or “won’t”.

4. Batches are filled best-first within each NVS region.

   • The first batch takes the region’s 40 highest-P(M1) students, the next takes the next 40, and so on.
   • A batch counts as M1-level while its expected M1 share is at least 50%.
   • When the M1 batches run out, the remaining students are re-ranked by P(qualify). A batch counts as M2-level while its expected qualifying share is at least 70%.
   • NEET batches are B1-level while their expected B1 share is at least 50%, using the government-MBBS model (Section 5).

   The thresholds live in sources/assumptions.yaml and are varied in Section 4.4.

For comparison, a General boy with 98 and 96 (board 97) has P(M1) = 0.54 and P(qualify) = 0.83 at CoE level. For General students, even strong board marks leave M1 not far from a coin flip.

How Model B treats each factor. Model B predicts the JEE percentile on a logit scale. On that scale each factor adds a fixed amount, and the result is turned back into a percentile. Near the middle of the range, +0.1 on the logit scale is about +2.5 percentile points; near the top it is much less, because a percentile cannot pass 100. The fitted effects, from 2,063 CoE and Nodal students in 2024-26 (794 girls):

Factor	Effect (logit)	What it means
Board average, per mark (around 85)	+0.10	Steady climb with marks, slightly curved
Studying at a CoE (vs Nodal)	+1.13	The largest single effect
Girl (vs boy, same marks, category, centre)	−0.48 (s.e. 0.05)	Clearly below zero
Category, vs EWS: General / OBC / SC / ST	−0.22 / −0.30 / −0.51 / −0.49	Percentiles; each category’s own cutoff is applied afterwards
PCM vs PCMB stream	+0.13	No meaningful effect
Cohort 2025 / 2026, vs 2024	+0.27 / −0.22	Year-to-year level, which feeds the uncertainty band
School (random effect)	spread 0.42	Centre-to-centre differences, shrunk for small centres
What none of these explain	spread 0.96	Two students alike on all of the above still land far apart

Is there a systematic gap between girls and boys? Yes, and the model’s single term is slightly generous to girls at the top. What the girls’ effect means in percentile points, for an OBC student:

Board average	CoE: boy	CoE: girl	Gap	Nodal: boy	Nodal: girl	Gap
80	80.2	71.5	−8.6	58.0	46.2	−11.8
85	86.9	80.4	−6.4	68.1	57.0	−11.1
90	92.5	88.4	−4.1	78.5	69.4	−9.1
95	96.3	94.1	−2.1	87.4	81.2	−6.2

The percentile gap narrows toward the top only because percentiles crowd together near 100. The underlying effect does not shrink. Three checks, each adding one term to Model B:

• CoE vs Nodal. The extra girls’ effect at a CoE is −0.18 (p = 0.08). That is not clearly different from zero, so read the CoE lift as about equal for girls and boys, with a hint that it favours boys.
• Higher board marks: the gap is wider. The girls’ effect grows by −0.019 per board mark (p < 0.001). So a girl at 95 is further behind a boy at 95, on the underlying scale, than a girl at 80 is behind a boy at 80. Model B uses one girls’ term for everyone, so it slightly overstates high-scoring girls’ chances, and with them the girls’ share of the M1 layer in Section 11.
• By cohort. The 2024 effect is −0.36; 2025 and 2026 add −0.22 and −0.11 (p = 0.06 and 0.34). Raw, girls’ JEE percentiles trailed boys’ by 3.8-13.0 points across cohorts, though they entered only 0-2 board points lower.

Model B’s methodology note explains these terms and the checks in its section on girls.

One correction applied: the model’s support floor. Model B’s board curve is quadratic, so below its training range it would extrapolate. Its training data rarely goes below a board average of 69, its 5th percentile, so every student below that is scored as if at 69, the same rule Model B’s own forecast now uses (data-assistant#422). That is monotone, and still generous: in the 2024 JNV-wide data, M1 rates below a board of 70 were 0-4% in every programme. The 6,537 JEE students below the floor are reported separately and excluded from the headline totals.

4. JEE: M1 and M2 potential

4.1 The ceiling: everyone coached

Among the 11,596 JEE-eligible students inside the model’s support:

If every one of them got…	Expected M1	Expected qualifiers
CoE-level coaching	3,406 (P10-P90 2,623-4,263)	5,090
Nodal-level coaching	1,122	2,177
difference a CoE makes	+2,284	+2,913

The band carries Model B’s year-to-year variation: a cohort can land well above or below an average year.

4.2 The yield curve: what each additional batch is worth

Batches are ranked nationally, best first. The share is the expected M1 rate within that batch of 40.

Batch	1st	10th	25th	40th	70th	80th	100th
Expected M1 share	100%	92%	79%	68%	50%	46%	38%
Cumulative M1	40	383	890	1,322	2,022	2,214	2,550

The 40th batch is where today’s footprint ends. The 70th is where batches stop being “half M1”.

4.3 By region: batches needed against today’s CoEs

All students

Region	JEE pool	M1-level batches	M2-level batches	Total	Today’s JEE batches	Batches needed
Bhopal	3,128	18	8	26	5.0	+21.0
Chandigarh	1,319	6	0	6	0.5	+5.5
Hyderabad	2,783	7	11	18	11.8	+6.2
Jaipur	1,513	3	0	3	2.2	+0.8
Lucknow	3,493	5	0	5	2.8	+2.2
Patna	2,637	7	1	8	4.3	+3.7
Pune	2,109	6	6	12	8.0	+4.0
Shillong	1,011	15	0	15	2.0	+13.0
All-India CoEs (Dakshana Bengaluru Urban, Pune)	—	3.5	—	3.5	3.5	+0.0
National	17,993	70.5	26	96.5	40.1	+56.4

Batches needed = Total − today’s JEE batches. A negative number is a surplus.

The same table with the income filters (Section 4.5):

≤ ₹8 lakh

Region	JEE pool	M1-level batches	M2-level batches	Total	Today’s JEE batches	Batches needed
Bhopal	2,941	16	7	23	5.0	+18.0
Chandigarh	1,235	5	0	5	0.5	+4.5
Hyderabad	2,560	6	10	16	11.8	+4.2
Jaipur	1,462	3	0	3	2.2	+0.8
Lucknow	3,402	5	0	5	2.8	+2.2
Patna	2,574	7	1	8	4.3	+3.7
Pune	1,903	5	6	11	8.0	+3.0
Shillong	968	14	0	14	2.0	+12.0
All-India CoEs (Dakshana Bengaluru Urban, Pune)	—	3.5	—	3.5	3.5	+0.0
National	17,045	64.5	24	88.5	40.1	+48.4

≤ ₹4 lakh (Avanti rule)

Region	JEE pool	M1-level batches	M2-level batches	Total	Today’s JEE batches	Batches needed
Bhopal	2,509	13	6	19	5.0	+14.0
Chandigarh	1,078	4	0	4	0.5	+3.5
Hyderabad	2,267	5	8	13	11.8	+1.2
Jaipur	1,332	2	0	2	2.2	−0.2
Lucknow	3,236	5	0	5	2.8	+2.2
Patna	2,419	6	1	7	4.3	+2.7
Pune	1,619	4	5	9	8.0	+1.0
Shillong	892	13	0	13	2.0	+11.0
All-India CoEs (Dakshana Bengaluru Urban, Pune)	—	3.5	—	3.5	3.5	+0.0
National	15,352	55.5	20	75.5	40.1	+35.4

How to read the comparison:

• Today’s JEE batches come from the Ops master, “NVS CoEs 26-27 (ALL)”: all 50 CoEs, each scaled by its own approved strength, ÷ 2 for XI and XII, ÷ 40 seats.
  • Most centres are 80, so 1 batch.
  • Dakshana Bengaluru Urban (280) counts as 3.5 JEE batches and Dakshana Pune (236) as 2.95 NEET batches.
  • Bundi and Kottayam (100) count as 1.25, Muzaffarpur (60) as 0.75 and Puducherry (40) as 0.5.
  • A JEE/NEET centre is split half and half, using the current-year stream column.
• The two all-India Dakshana CoEs fill first, with the country’s top students, so they never show a need. The regions are sized on everyone else. Bengaluru Urban’s 3.5 JEE batches do not count against Hyderabad or Pune, which are about at capacity. Under the ₹4 lakh rule, Jaipur has a JEE surplus.
• Bhopal, Shillong, Hyderabad and Chandigarh are the largest gaps. Bhopal alone could fill 18 M1-level batches and has 5.0 JEE batches today.
• Against the old FY27-FY30 plan. That plan sized only Cat-1 CoEs: 43 nationally, against our 71 M1-level batches. It got there by a different route, the v1 model plus girls +5.68 and a +5 buffer (see the provenance note). The region numbers are in regions.csv.

4.4 Sensitivity to the batch thresholds

M1 share threshold	M1-level batches	M2-level batches (at 70% qualify)
65%	43	38
50% (default)	71	26
35%	111	13

The total (M1 + M2) is stable at 81-124 batches. The threshold only decides which side of the M1/M2 line a batch is labelled. The full yield curve is in yield_curve.csv, so any other bar can be read off directly.

4.5 Income filters

Avanti’s CoE selection admits only families earning ₹4 lakh a year or less, and a missing income counts as rejected. ₹8 lakh is the wider band other programmes can use. Income is NCST’s self-reported annual_family_income, so the 4,906 students not in NCST drop out of both filtered scenarios. The CA/CLAT bar is held at its unfiltered value: income changes who is eligible, not what M2-level performance means.

	All students	≤ ₹8 lakh	≤ ₹4 lakh (Avanti rule)
JNV Class-10 students	45,170	38,104	34,129
JEE pool (engineering, Maths Standard)	18,133	17,185	15,492
Ceiling: expected M1, everyone CoE-coached	3,406	3,144	2,707
Ceiling: expected qualifiers	5,090	4,704	4,074
M1-level batches	71	65	55
… their expected M1	2,041	1,869	1,585
M2-level batches	26	24	20
B1-level students, government MBBS (CoEs)	715 (18)	650 (16)	556 (14)
CA candidates / CLAT candidates	235 / 342	198 / 289	170 / 248

The ₹4 lakh rule removes 23% of the M1-level batches (71 → 55). The largest drops are in Bhopal (18 → 13) and Chandigarh (6 → 4). Region by region under each scenario: income_scenarios_by_region.csv.

4.6 National bars against state bars

Until 8 October 2026 this note sized CoEs with one national bar per category. It now uses the bar for each student’s home state. The two runs side by side, at the ₹4 lakh income rule:

Region	M1-level batches	M2-level batches	JEE CoEs needed	All CoEs needed
Bhopal	13 → 13	8 → 6	+16.0 → +14.0	+21.0 → +19.0
Chandigarh	3 → 4	2 → 0	+4.5 → +3.5	+4.0 → +3.0
Hyderabad	7 → 5	5 → 8	+0.2 → +1.2	+0.2 → +1.2
Jaipur	3 → 2	3 → 0	+3.8 → -0.2	+2.8 → -1.2
Lucknow	7 → 5	6 → 0	+10.2 → +2.2	+12.5 → +3.5
Patna	7 → 6	6 → 1	+8.7 → +2.7	+10.1 → +4.1
Pune	4 → 4	4 → 5	+0.0 → +1.0	-2.0 → -1.0
Shillong	4 → 13	2 → 0	+4.0 → +11.0	+15.0 → +22.0
National	48 → 52	36 → 20	+47.4 → +35.4	+63.6 → +50.6

• The total falls from 63.6 to 50.6 CoEs (JEE 47.4 → 35.4). M1-level batches move from 48 to 52, M2-level from 36 to 20.
• M2 shrinks most in the north. Outside Delhi and Madhya Pradesh the only good government colleges there are the NITs, so the M2 bar is the M1 bar, and the M2-level batches that qualifying JEE used to fill disappear.
• Shillong grows. The home-state quota at the North-East NITs admits students far below the national bar, so many more of the region’s students are M1-level.
• M1 is tighter for reserved categories in the large states. The national OBC M1 bar is 92.68; the loosest NIT or IIIT seat open to an OBC student from a large state is about 95, and JoSAA 2025’s other-state OBC seats at the five bottom-tier NITs that define the national bar close around the 96th percentile on the same conversion. So the national OBC bar looks looser than the seats it stands for. That is unverified: the national bars come from a build that reads a local spreadsheet, and the question is open with its owners.

5. NEET B1: a government MBBS seat in the student’s own state

B1 is sized with the government-MBBS model, its own analysis and paper, in the same way the JEE numbers come from Model B. In short:

• It predicts NEET marks out of 720, with Model B’s machinery: a bounded scale, a school effect, and the year-to-year spread carried into every forecast. It is trained on 788 past CoE and Nodal NEET sitters (2024-26), of whom 100 won B1. The CoE effect is +0.50 on the logit scale; unlike JEE, there is no clear girls’ gap (−0.03).
• Each student is judged against their own bar: the government MBBS floor for their domicile state and category, from the College DB. The bar runs from 177 marks in parts of the North-East to over 530 in Uttar Pradesh. Private, deemed and management seats are excluded, because our students cannot afford them. BDS is left out.
• Where the bar sits at the qualifying gate, the seats bind. In 23 state-category cells the government seats did not fill, so any qualifier could have had one. There the expected B1 among our students is capped at the government seats open to that state’s students: 85% of in-state government intake, plus the Central Pool and the regional institutes (RIMS Imphal, NEIGRIHMS Shillong).
• Validated leave-one-cohort-out, each student against the cutoff that actually applied to them: CoE students pooled 78 predicted against 79 actual. Nodal students are over-predicted (46 against 21), and single years miss by about two to one.

Applied to the 2026 medical-stream pool, coached at CoE level:

Region	Today’s NEET batches	B1-level CoEs: all	≤ ₹8 lakh	≤ ₹4 lakh	Proxy bound (≤ ₹4 lakh)	Need (≤ ₹4 lakh)
Bhopal	1.0	4	3	2	8	+1.0
Chandigarh	0.5	1	1	0	2	−0.5
Hyderabad	2.0	1	1	0	2	−2.0
Jaipur	1.0	0	0	0	1	−1.0
Lucknow	1.8	1	1	1	1	−0.8
Patna	1.6	1	1	1	3	−0.6
Pune	3.0	1	1	0	1	−3.0
Shillong	1.0	12	12	11	18	+10.0
All-India CoEs (Dakshana Bengaluru Urban, Pune)	2.95	2.95	2.95	2.95	—	+0.0
National	14.8	23.9	22.9	17.9	36	+3.2

There is room for more NEET CoEs, but mostly in one region. Today’s 50 CoEs hold about 14.8 NEET batches a year, counting a JEE/NEET centre as half and Dakshana Pune’s 2.95 all-India batches as always full. Under the ₹4 lakh rule:

• Net need: +3.2. Shillong and Bhopal are short of NEET CoEs; Hyderabad, Jaipur, Lucknow, Patna, Pune hold more NEET batches than they have B1-level students.
• Counting only regions with room: surplus seats in one region cannot serve another, so the gross gap is about 11 more NEET CoEs (14 for all students).
• Most of Shillong’s need comes from the states where the bar sits at the qualifying gate (Manipur, Arunachal Pradesh, Tripura, Meghalaya, Ladakh), where coached students clear the bar easily.

The seat cap, by domicile state:

Domicile state	Students in floor cells	Expected B1 before the cap	After the cap
Arunachal Pradesh	82	58.9	58.9
Ladakh	69	58.1	7.0
Manipur	244	195.5	195.5
Meghalaya	11	9.0	9.0
Mizoram	5	3.8	3.8
Nagaland	4	3.1	3.1
Tripura	18	13.3	13.3

Read the North-East numbers with care. 19 of the 23 gate cells are rated INDICATIVE in the College DB (official but thin or OCR-derived), not VERIFIED. And the cap counts every government seat open to the state, not the share our students would realistically win against the state’s other candidates, so it is an upper bound. The cap binds only where our students alone would exceed the seats: Ladakh (58 → 7).

The JEE-M1 proxy, kept as a bound. Before the NEET marks could be checked (data-assistant#420), B1 was sized by scoring medical-stream students who took Maths Standard with Model B and counting those with a 50% or better chance of JEE M1. That proxy gives 1,137 students and 36 B1-level batches at ≤ ₹4 lakh, against 15 from the government-MBBS model. The two disagree by region in the expected direction: the proxy is generous where state bars are high (Bhopal, Patna) and blind to the North-East’s low bars.

• 11,634 medical-stream students took Maths Standard; 6,545 are inside Model B’s support, and 1,451 are M1-capable (against 2,492 in the engineering stream).
• The government-MBBS model scores every medical-stream student with a board mark and a category, Maths Standard or not, so it does not share the proxy’s undercount.

6. CA and CLAT

There are no CA or CLAT outcomes in the warehouse, so these use your definition: a student performing like an M2-level JEE student who opts for Commerce (CA) or Arts (CLAT).

1. The bar. The 10th aggregate (5 subjects) of the students placed in JEE M2-level batches: the bottom quarter of them sits at 85.8%. 11,062 students in the 2026 pool are at or above it.
2. Stream choice is not observed yet for this cohort, and the 12th-board stream column is unusable. So the split comes from the last linked cohort: 10th board 2023 → 12th board 2025, n = 26,949. Stream is derived from subjects: Physics means Science; otherwise Accountancy or Business Studies means Commerce; otherwise Arts.
3. The result. Above the bar, 92%-99% chose Science, depending on the marks band. Girls chose Commerce and Arts a little more often than boys, so the shares are applied by gender.

	All students	≤ ₹4 lakh
CA (Commerce, M2-equivalent)	235	170
CLAT (Arts, M2-equivalent)	342	248

Bhopal is the largest region for both. Strong JNV students overwhelmingly choose Science. So a CA or CLAT programme at this bar is about 1-2 batches per region at most. A larger one has to recruit below the M2 bar, or change stream choice in Class 10.

7. Was Model B predictive? The newly selected Class 11

The 2026-27 JNV CoE Class-11 engineering cohort was selected in March 2026 and has now sat a median of 12 internal tests.

• Who was checked: 760 of the 991 enrolled. The rest lack either a verified board mark or a test record.
• The test measure: each student’s average standardized test score, scaled within each test (z-scored), so a hard test and an easy one count equally.

Predictor	Rank correlation with tests	Of the top 20% on tests, share found by the predictor’s top 20%
Board average (raw)	0.45	50%
Model B predicted percentile	0.41	44%
NCST 2026 score	0.24	28%

By predicted-percentile decile, the test score climbs almost monotonically, from −0.41 SD (bottom tenth) to +0.47 SD (top tenth).

Model B is predictive, and so far it ranks this cohort slightly worse than raw board marks. The reason is that Model B folds in category and gender. Those shift JEE percentiles, but they do not shift internal test percentages. The comparison that matters, JEE 2028, is two years away. NCST looks weak here partly by construction: these students were selected on NCST, which compresses its range. Put together in one regression, Model B carries 0.25 SD per SD and NCST adds 0.09 SD.

8. Is NCST a better predictor at the top of the curve?

The fair test uses students who sat NCST but were not all selected on it: the NCST 2022-24 takers who later sat JEE.

• Sample: 2,268 engineering students (2022: 1,490, 2023: 258, 2024: 520). Only 15% were later coached in a CoE or Nodal, and 31% reached M1.
• Putting the years on one scale: the NCST scale changed every year, so both predictors are taken as percentile ranks within year.

	NCST	Board average
Rank correlation with JEE percentile	0.61	0.59
M1 rate among the top 10% on this predictor	60%	48%
M1 rate among the top 30%	46%	42%
Effect on the 90th percentile of JEE (quantile regression, both in)	9.4	9.0
Among the top 30% on board, NCST’s correlation with JEE	0.42	—

Yes, at the very top.

• The top tenth on NCST reached M1 at 60%. The top tenth on board marks reached it at 48%.
• Among students already in the top 30% on board marks, NCST still separates JEE outcomes strongly.
• Across the whole range the two are about equally informative.
• Board marks and NCST measure different things, and the best selection rule uses both.

What we then did. These are the three steps proposed in the first draft.

1. NCST inside Model B. Model B was refitted with NCST percentile-within-year, and allowed a different NCST effect at CoEs. It used the 531 training students who have an NCST score (2022-24 NCST, sitting JEE 2024-26). Both versions were fitted on exactly those rows, and each cohort was predicted from the other two.

Held-out cohort	n	Actual M1	Predicted: board	Predicted: board + NCST	AUC: board	AUC: + NCST	Brier: board	Brier: + NCST
2024	141	73	71	67	0.84	0.83	0.158	0.160
2025	64	40	37	39	0.90	0.92	0.136	0.125
2026	326	192	213	227	0.81	0.81	0.179	0.194

NCST adds no reliable improvement once board marks are in the model.

• Discrimination is unchanged: AUC is the chance the model ranks a random M1 student above a random non-M1 student.
• Calibration is better in 2025 and worse in 2024 and 2026.
• NCST’s own coefficient is not distinguishable from zero (p = 0.21).

This does not contradict the top-of-curve result above. That came from the broad NCST pool, of whom only 15% were coached. Model B’s training students were already selected, partly on NCST, which removes most of the variation NCST could explain. Among students we have already selected, board marks carry most of what NCST knows. Across the whole pool, NCST helps most at the very top.

2. Ranking the selection pool on the combined probability. Not done. Step 1 gives no evidence that the combined model forecasts better, so the potential in this note stays on board marks alone. The case for NCST in selection rests on the broad-pool result, and it is the selection floors, not this forecast, that should use it.

3. A prospective test, frozen today. This year’s Class-11 JNV CoE engineering students now have predictions on record from both models. That is 709 students with a verified board mark, 689 of them with NCST 2026. Expected M1 on the same 689 students:

• 319 from board marks alone
• 337 with NCST

JEE 2028 will settle which was closer. One caution: NCST 2026 percentiles are taken against all national takers, while the training percentiles came from smaller Dakshana-run pools. The NCST version’s higher count partly reflects that change of scale. The file is prospective_class11_2026_27.csv. It holds student ids, so it is kept in the private bucket only.

9. Validation and what to trust

Model B at the M1 bar (leave-one-cohort-out). Model B was validated for qualifying, not for M1. Here, each cohort is predicted from the other two:

Held-out cohort	CoE: predicted / actual M1	Nodal: predicted / actual M1
2024	51 / 54	60 / 40
2025	183 / 217	41 / 55
2026	398 / 326	40 / 27
Pooled	632 / 597 (+6%)	140 / 122

• Pooled it is close (632 against 597, +6%); cohort by cohort it is not. 2025 was under-predicted by 16% (183 against 217) and 2026 over-predicted by 22% (398 against 326).
• That year-to-year swing is why every total in this note carries a band.
• Read any single region-year to about ±20%.

What this note does not claim:

• That every student would gain the selected-CoE lift. Model B’s CoE effect comes from students who were selected, and the students further down the yield curve are less like them. The yield curve’s lower batches are the least certain part.
• That the batches can be filled. Income is applied as a scenario (Section 4.5). Willingness, coaching preference and travel are not applied. They decide how much of the potential is actually reachable.
• That the North-East NEET numbers are firm. Most of the bars there are indicative, and the seat cap is an upper bound (Section 5).

10. The earlier models, and why this note uses Model B

Three earlier board-marks models came before Model B. Each was built for a fair purpose. Only one of them answers “how many could reach M1”, and it is not the one the FY27-FY30 plan used. The provenance note traces each step and reproduces the coefficients from the original data.

The lineage:

1. Rahul’s regression (April 2025). OLS of JEE Main percentile on programme, gender, category, stream and 10th maths and science.
   • Data: the 2024 NTA results for 6,978 JNV grade-12 students.
   • What it was for: to measure what each programme adds over Enable, and it does that well. CoE comes out +21 percentile points, Nodal +7, and girls 6 points lower.
   • The question it answers is an average difference. OLS is the right tool for that, which is why our impact paper still uses it.
2. The BigQuery column predicted_percentile_on_10th_marks. The same family of linear fit, now stored per student.
3. The v1 “New Prediction Model”. The same data, with squared marks terms added and the category cutoff used as an extra predictor. This is the model the Godrej deliverable and the FY27-FY30 plan used.
4. The FY27-FY30 plan. v1 predictions for each Grade-11 student, then three adjustments: the full CoE effect (+18) for every student, girls given +5.68 so they match boys, and a +5 percentile buffer. Students over the M1 bar were counted and divided by 40.

Where the old approach goes wrong when the question becomes forecasting:

Issue	v1 and the old plan	Model B
Who it learns from	All JNV JEE-takers in 2024. The CoE effect is one constant, estimated from 74 selected students at two centres.	Avanti’s own CoE and Nodal students, 2024-26, first attempts (2,063). The CoE effect differs with board marks.
The 0-100 bound	None. Held out, a linear fit predicted 139 of 886 percentiles above 100. A General student with a perfect board predicts 101.6.	Models the percentile on a bounded (logit) scale, so it cannot leave 0-100.
The cutoff	Used as a predictor and as the pass mark. This counts General students’ higher cutoff against them twice: their qualifying rate came out about 19%, against about 46% without the double count.	Used only as the pass mark, and varied across recent years.
Low board marks	The squared terms turn back up below about 62 in maths and 65 in science, so weaker students are predicted higher.	Has the same curvature. Model B now holds every student at its training floor (data-assistant#422), and so does this note.
Each student	One predicted percentile, then a hard yes/no at the bar. A student 1 point over counts as a sure qualifier.	A probability of clearing each bar. A batch’s count is the sum of those probabilities, so close calls count as close calls.
Year-to-year swings	None. One number, with no range.	Carries the between-cohort spread, so every total has a range.
Proof it works	“91% accuracy” on a batch where 90% qualified. Predicting “everyone qualifies” would also score about 90%.	Each cohort predicted from the other two: 784 predicted against 782 actual qualifiers, pooled across 2024-26 (methodology, Section 5.1). At M1, 632 against 597 for CoE students.
Optimistic adjustments	Girls +5.68, buffer +5, the lower EWS cutoff, and the full CoE lift for everyone, all stacked.	None. Income and stream are applied as explicit scenarios instead.

What that means in practice.

• On cohorts that qualify at 88-91%, the methods give similar totals. Few students sit near the bar, so rounding errors cancel out. The linear method’s pooled total was 784 against 782 actual.
• This note asks about everyone else. Its pool includes the whole Navodaya cohort, most of whom are close to or below the bar. That is where a hard threshold, an unbounded scale and a constant CoE effect go most wrong.
• Do not read the old plan’s 43 Cat-1 CoEs, against our 71 M1-level batches, as a comparison. The two differ in their pool (last year’s Grade 11 not in a CoE, against this year’s whole Class-10 cohort), their adjustments, and what they count (students over a bar, against batches whose expected M1 rate is at least 50%). Region by region they diverge.

Model B is not perfect either:

• It shared v1’s low-board curvature, now held at the training floor (data-assistant#422).
• At the M1 bar, single cohorts miss by up to about a fifth (Section 9).
• It learns from students who were selected.

Section 9 sets out how far to trust it.

11. Girls

Where girls sit in each layer. Students are counted as in the at-a-glance table, coached at CoE level.

Layer	All students: count	Girls	≤ ₹4 lakh: count	Girls
M1	2,492	21%	1,907	19%
M2	1,013	31%	812	31%
B1 (government MBBS)	715	56%	556	55%
B1, JEE-M1 proxy bound	1,451	42%	1,137	42%
CA (proxy)	235	62%	170	60%
CLAT (proxy)	342	56%	248	55%

Against the pools they come from:

Base pool	Girls
All JNV Class-10 students, 2026	44%
Engineering stream in NCST	35%
JEE-eligible (engineering, Maths Standard)	34%
Medical stream in NCST	56%

What this shows:

• Girls thin out at every step up the JEE ladder. They are 44% of the Class-10 cohort and 35% of those who choose engineering. Among the JEE-eligible they make up 31% of the M2 layer and only 21% of the M1 layer (19% under the ₹4 lakh rule).
• On the medical side they do not. Girls are 56% of the medical stream and 56% of the students likely to win a government MBBS seat, because the NEET model finds no clear girls’ gap on NEET marks. The old JEE-M1 proxy put them at 42%: it carried JEE’s gender gap into medicine, where the NEET data does not show one.
• CA and CLAT run the other way. Girls are 62% and 56% of these layers, because strong girls choose Commerce and Arts slightly more often than strong boys. The layers are small, about 235 and 342 students.

Why the M1 layer is so male.

• On 10th boards, girls in our CoEs are level with boys at the same centre. On NCST they are about 7.5 points behind, and across all 20,925 engineering NCST takers the gap widens toward the top (girls-CoE gap analysis).
• Model B also carries a girls’ effect, because girls with the same board marks have done worse on JEE.
• For the 2026 JEE cohort, NCST explains only a small part of that effect. Controlling for NCST as well as board marks moves girls’ odds of M1 from 0.33 to 0.36 of boys’ (still clearly below, p = 0.002; n = 281 linked students). So the shortfall is mostly not in either entry score.

Girls’ share of CoE seats, by region (≤ ₹4 lakh): what today’s CoEs hold, against what a strict merit order from this pool would give.

Region	Girls in today’s CoEs (2025-26)	Today’s JEE seats filled on merit	Today’s NEET seats filled on merit	All M1 + M2 potential	All B1 potential
Bhopal	23%	13%	65%	22%	56%
Chandigarh	65%	25%	55%	32%	—
Hyderabad	32%	27%	45%	27%	—
Jaipur	25%	10%	28%	8%	—
Lucknow	27%	6%	47%	11%	55%
Patna	35%	13%	52%	15%	52%
Pune	34%	22%	49%	24%	—
Shillong	88%	28%	60%	34%	58%

What it means for CoE seats.

• Filled strictly in merit order, today’s JEE seats would go 6%-28% to girls, depending on the region. That is the “filled on merit” columns above. Today’s CoEs actually hold 37% girls (2025-26). Shillong is highest, because Kohima and Kokrajhar are girls-only.
• Avanti does not fill them that way. Girls’ seats are filled from a separate NCST list, and girls are 47% (340 of 727) of this year’s CoE students with full entry records.
• So reaching parity in the CoEs means admitting girls further down the probability ranking than boys. The girls-CoE gap analysis shows it: at mixed centres the median admitted girl scored 31% of the NCST maximum, against 39% for boys.
• The gap that results is then Avanti’s to close, and that analysis finds the 2027 cohort’s gap closing.

12. Lift 1: do we keep the M1-level students?

Everything above sizes the CoEs the Navodaya system could fill. The first lift is simpler: keeping the students we already have. We reran the placement above on the 2025 Class-10 boards and followed every student it puts in an M1-level batch through Avanti’s enrolment record (attrition.py). The same rule is applied to this year’s cohort.

• The pool is wider than the headline. The 2025 boards have no NCST link, so no declared stream or income. For both years, every student with Maths Standard is scored as if on the engineering track, with no income filter and no all-India skim. Read the counts as “students with M1-level marks”: 4,240 in 2025 (106 batches) and 4,200 in 2026, not CoE seats.
• About a quarter cannot be traced. 882 of the 2025 group (21%; 37% in 2026) have a board record with no Avanti student id. That is a gap in the identity link, not students leaving: the online NVS programme enrols almost the whole JNV Class-11 cohort. The shares below are of the 3,358 traced students.

Where in Class 11	2025 cohort	Still enrolled in Class 12	Sat a full Lakshya test that year	2026 cohort (now)
CoE	580	99%	60%	339
Nodal	32	88%	84%	7
Lakshya (online NVS)	2,155	94%	27%	1,797
not enrolled that year	591	—	—	502
no Avanti id	882	—	—	1,553
Total M1-level	4,240			4,200

The leak is reach, not drop-out. Of the M1-level students we enrolled, 99% of those in a CoE and 94% of those on Lakshya were still enrolled in Class 12. But only 17% of traced M1-level students were in a CoE; 64% were on Lakshya only, and only 27% of those sat a single full-length test that year. A further 18% had an Avanti id but no enrolment in Class 11, concentrated in the Pune and Hyderabad regions; that may be an enrolment-coverage gap there rather than attrition, and is unverified.

This year, by region. Of 2,647 traced M1-level students in Class 11 now, 339 are in a CoE and 1,797 are on Lakshya only (16% of those have sat a full test since April):

Region	M1-level students	Traced	In a CoE	On Lakshya only	Not enrolled
Bhopal	1,160	791	95	693	2
Chandigarh	320	218	13	203	1
Hyderabad	440	308	82	69	156
Jaipur	160	81	17	62	2
Lucknow	320	115	5	108	2
Patna	480	232	40	187	2
Pune	360	166	45	44	74
Shillong	960	736	42	431	263

Bhopal has the most: 693 traced M1-level students on Lakshya only, enough to fill several CoEs from students we already reach. Lift 1 is finding these students early and moving them into intensive support.

Related analyses

This note builds on, and is checked against, the following:

Analysis	What this note takes from it
JEE-Main prediction (Model B): forecast paper, methodology	The model that scores every JEE student, and its leave-one-cohort-out validation
Provenance of the board-marks JEE models	Rahul’s 2024 regression → the BigQuery column → v1 → the FY27-FY30 plan that Section 4.3 is compared with
Godrej v1 prediction	The v1 “New Prediction Model” behind the old plan
CoE selection 2026-27	The ₹4 lakh income rule and the NCST and maths-percentile floors; the earlier finding that NCST is weak within a coached group (R² ≈ 0.19), which Section 8 extends to the top of the curve
Avanti impact paper and core impact analyses	The CoE-vs-Nodal lift, of which Model B’s CoE effect is the forecasting form
Academic Levels	The definitions of M1 and M2 and their cutoffs
Government-MBBS prediction: paper	The NEET model behind every B1 number: marks, in-state government bars and seat caps
Golden student record	The training outcomes (nit_q, NEET marks via neet_appeared, data-assistant#430) for both models
M1 and M2 bars by home state	What M1 and M2 mean state by state: the NIT/IIIT seat open to a student from each state, and the good-government-college (M2) bar, with the alumni check behind it

Replication data

All outputs are in the private bucket and open to Avanti accounts:

• regions.csv: the region table, including today’s footprint, the old plan and CA/CLAT.
• yield_curve.csv: every national batch, with its expected share and cumulative count.
• sensitivity.csv: batch counts under every threshold.
• income_scenarios.csv and income_scenarios_by_region.csv: the headline numbers under each income filter.
• ncst_in_model_b_loco.csv: Model B with and without NCST, leave-one-cohort-out.
• prospective_class11_2026_27.csv: the frozen 2026-27 Class-11 predictions (student ids, private).
• lead_table_le_4_lakh.csv: the opening table.
• girls_by_layer.csv: girls’ share of each layer and pool.
• lead_table_le_4_lakh_by_state.csv: the state table.
• summary.json: every headline number in this note.
• validate_m1_loco.csv: Model B’s M1 validation. The NEET model’s validation is in its own outputs.
• class11_tests_by_decile.csv: the Class-11 check.
• per_student_deid.csv: one row per scored student, with region, category, gender, board mark and every probability. No roll number, name or school.