The Estimate

TAM-INS.05 · The Insufficient · The Approximate Mind

The man with the notebook, first introduced in Part 74 of this series, has been doing arithmetic. Not the kind he was trained for, not the systems-level calculations of his healthcare career, but the simpler and more difficult arithmetic of whether an idea can survive contact with a budget.

He has been writing numbers on the backs of pages already filled with questions. The questions face inward. The numbers face outward. He has a habit of underlining totals twice, a habit his son once pointed out to him, and he said it was because the first underline is the number and the second is the commitment to take the number seriously.

He has a figure. It is smaller than people assume. That is part of the problem.

What Is Being Costed

#

The previous four essays described an architecture. A skeptic that questions whether the categories in any specification are real. Seven philosophical operations that catch what a single skeptic would miss. A bias-in-intent layer that asks who commissioned the specification and why. A retroductive method that works backward from outcomes to undocumented mechanisms. And beneath all of them, the Intersectional Systemic Harm Index as the operational layer that demonstrates, case by case, that the compound is the mechanism.

This essay asks what it would cost to build a working pilot of this architecture in one domain, in one country, for long enough to find out whether the idea works in practice or only in essays.

The domain is maternal health. The country is India. The reasons for both choices are practical, not symbolic. India has the population density, the compound barriers, the stratum gaps between what is documented and what is real, and the institutional infrastructure to test every component against conditions severe enough to reveal whether the architecture produces genuine value or elaborate documentation.

Maternal health is chosen because the outcome gap is measurable, consequential, and insufficiently explained by documented mechanisms. India’s maternal mortality ratio has improved dramatically over two decades but remains deeply uneven across states, districts, and populations in ways that the documented variables do not fully account for. The residual, the gap between what the risk models predict and what actually happens, is large enough to constitute retroductive evidence that undocumented mechanisms are operating.

The Components and Their Costs

#

The epistemic AI. A domain-specific small language model trained on the Indian maternal health literature, including published research, gray literature from state health departments, ASHA worker field reports where digitized, and documented traditional birth practices from ethnographic sources.

Training the model requires assembling the corpus first. The published literature is accessible. The gray literature is scattered across state health directorates, district hospitals, and NGO archives. The traditional knowledge is oral, partially documented in anthropological studies, and requires careful ethical engagement with the communities that hold it.

Corpus assembly: six to eight months of work by a team of three to four researchers with domain expertise and linguistic range across Hindi, Marathi, Tamil, and at least two other state languages. This is the most labor-intensive component and the one most likely to be underestimated. Realistic cost: ₹40-80 lakh ($50K-$100K).

Model training on the assembled corpus: ₹4-25 lakh ($5K-$30K) in compute. The model does not need to be large. It needs to be deep in one domain.

Values framework library: structured representation of ethical traditions relevant to Indian maternal health, including but not limited to utilitarian, capabilities, care ethics, Gandhian, Ambedkarite, and Dalit feminist frameworks. This is a knowledge engineering task requiring sustained engagement with scholars across traditions. Realistic cost: ₹80 lakh-₹2 crore ($100K-$250K), primarily human expertise.

Integration and orchestration across the four interrogation modes: ₹40-80 lakh ($50K-$100K) in engineering.

Epistemic AI subtotal: ₹1.6-3.8 crore ($200K-$480K). Realistic midpoint: ₹3.2 crore ($400K).

The skeptic. This is not one model but seven operations running in parallel, each trained on a different corpus of category failures.

The Pyrrhonian spine: a classification system trained on documented cases where the unit of analysis turned out to be wrong. This corpus does not exist in compiled form. Cases must be identified across medical history, agricultural policy, public health, and development economics, drawn from the literature on optimization failures that the series has been examining since Part 74. Building this corpus is itself a research contribution. Realistic cost: ₹60 lakh-₹1.2 crore ($75K-$150K) for assembly, ₹8-20 lakh ($10K-$25K) for training.

Six tradition-specific operations, each requiring a curated training set developed in collaboration with scholars in each tradition. The Madhyamaka anti-reification operation requires engagement with Buddhist philosophical scholarship. The Ubuntu relational ontology operation requires engagement with African philosophy. The feminist standpoint operation requires engagement with gender studies scholars who have worked specifically on Indian maternal health. Each operation is a focused knowledge engineering effort. Realistic cost per tradition: ₹24-48 lakh ($30K-$60K). Total for six: ₹1.4-2.9 crore ($180K-$360K).

Integration layer: ₹40-80 lakh ($50K-$100K).

Skeptic subtotal: ₹2.5-5.1 crore ($315K-$635K). Realistic midpoint: ₹4 crore ($500K).

The ISHI layer. The compounding algorithm, productionized for deployment in the pilot’s geography.

Algorithm development: barrier scoring, interaction modeling, excess detection. Software engineering, not AI training. ₹40-80 lakh ($50K-$100K).

Data integration: connecting to available data sources in the pilot districts. In India, this means HMIS records, RCH portal data, ASHA worker reports, Ayushman Bharat claims data where accessible, and potentially Aadhaar-linked service utilization records subject to privacy compliance under the DISHA framework. Each integration is a data engineering challenge with its own regulatory and technical requirements. ₹80 lakh-₹1.6 crore ($100K-$200K).

Calibration study: 12-18 months of longitudinal tracking to determine whether compound scores predict maternal outcomes better than individual risk factors. This requires field staff, data collection infrastructure, and relationships with the PHCs and district hospitals in the pilot area. ₹80 lakh-₹1.6 crore ($100K-$200K).

ISHI subtotal: ₹2-3.2 crore ($250K-$500K). Realistic midpoint: ₹3.2 crore ($400K).

The human layer. The component every technology budget underestimates and every honest assessment must center.

Domain expert panel: 15-20 people who carry the knowledge the system is designed to make visible. ASHA workers who have served the pilot districts for years and know every household’s actual situation. Auxiliary nurse midwives who have delivered babies in conditions no clinical guideline was written for. District health officers who have watched three generations of health programs arrive and depart. Retired public health officials who remember what each program could not see. Agricultural extension officers in the same geography who understand the economic pressures on the households the maternal health system serves.

These people are the ground truth. They must be compensated at rates that reflect the value of their knowledge, not the rates the development sector typically offers community participants. Monthly honoraria, travel support, and time compensation for an 18-month engagement. ₹1.2-2.4 crore ($150K-$300K).

Philosophical consultants: scholars in each of the seven traditions, engaged for corpus development and ongoing evaluation of whether the operations are capturing what their traditions actually argue. ₹80 lakh-₹1.6 crore ($100K-$200K).

Human layer subtotal: ₹2-4 crore ($250K-$500K). Realistic midpoint: ₹3.2 crore ($400K).

Infrastructure and operations. Compute for running all three layers in parallel over 18 months: ₹40-80 lakh ($50K-$100K).

Project management, institutional coordination, regulatory navigation, and the relationship-building that determines whether the pilot has access to the data and the communities it needs: ₹1.2-2 crore ($150K-$250K). This line item is where most pilots fail. The technical work can be brilliant. If the district collector does not return your calls, the pilot does not run.

Field deployment: hardware at pilot sites, connectivity (significant in rural India), training for PHC staff who will interact with the system’s outputs: ₹80 lakh-₹1.6 crore ($100K-$200K).

Infrastructure subtotal: ₹2.4-4.4 crore ($300K-$550K). Realistic midpoint: ₹3.2 crore ($400K).

The Total

#

Eighteen months. Two districts. One domain. All three layers running in parallel.

Total realistic estimate: ₹16.8 crore ($2.1 million).

Range: ₹10.4-21.6 crore ($1.3M-$2.7M).

What the Number Means

#

India’s National Health Mission budget is roughly ₹32,000-40,000 crore ($4-5 billion) annually. The Ayushman Bharat digital health infrastructure has absorbed hundreds of crores. A single AI deployment by a major health technology company in India runs ₹40-120 crore ($5-15M).

This pilot is a rounding error on any of those budgets.

That is part of the problem. The amount is small enough to fund and small enough to ignore. It does not register as a major investment. It does not trigger the institutional attention that major investments receive. It sits in the zone where it is too large for a single research grant and too small for a government program, too unconventional for most philanthropic foundations and too practical for most academic funding bodies.

The architecture is affordable. The question is whether anyone’s budget has a line item for questioning the questions the rest of the budget is built to answer.

Who Funds It

#

The honest assessment.

The Indian government is not the right first funder. Government funding cycles run 2-3 years for approval alone. The project would need to conform to existing program frameworks, which means it would be shaped by exactly the institutional categories the skeptic is designed to challenge. Government funding is essential for scale. It is counterproductive for a pilot whose purpose is to demonstrate that the government’s existing categories are insufficient.

The Tata Trusts are the most natural fit in the Indian philanthropic landscape. History of funding public health innovation with institutional patience. Experience with projects that challenge conventional approaches. Comfort with ambiguity and long time horizons. The risk: the Trusts fund many things, and this pilot competes for attention with initiatives that have more immediate, measurable outcomes. The skeptic architecture produces outputs that are, by design, uncomfortable. Whether a philanthropic institution funds the production of its own discomfort is an open question.

The Gates Foundation’s India office is a possibility if positioned as maternal health equity research, which it genuinely is. The risk: the Foundation’s theory of change emphasizes scalable, measurable interventions. The skeptic architecture’s output is a challenge to the concept of measurability itself. The Foundation would need to fund a project that questions the epistemological framework the Foundation uses to evaluate all its other projects.

ICMR with a multilateral partner (WHO India, UNICEF) is viable if positioned as health systems research. ICMR has funded unconventional research. The multilateral partner provides legitimacy and institutional cover. The risk: multilateral timelines and reporting requirements may constrain the pilot’s ability to challenge the very frameworks the multilaterals use.

A consortium of Indian research universities (IIT Bombay’s CTARA, IIPH Hyderabad, Azim Premji University, possibly JNU’s Centre for the Study of Social Systems) pooling existing grants and institutional resources. This is the most intellectually natural home. The risk: academic institutions move slowly, interdisciplinary collaboration is difficult to coordinate, and the pilot requires field deployment capacity that most universities do not have.

The Anthropic Institute or a comparable AI safety research body could fund the skeptic component specifically. The skeptic is, in a precise technical sense, an alignment tool: it aligns the system’s categories with reality rather than with the training data’s assumptions about reality. If AI safety research is serious about ensuring that AI systems operate on accurate representations of the world, the skeptic is a safety intervention. The risk: AI safety institutions are oriented toward frontier model risks, not toward the epistemological infrastructure of domain-specific deployments.

The realistic path is a consortium. No single funder covers the full scope. The epistemic AI and ISHI are funded as health systems research through ICMR or a health-focused philanthropy. The skeptic is funded as epistemological research through an academic consortium or an AI safety body. The human layer is funded through a community health organization with existing relationships in the pilot geography. The integration is funded by whoever has the institutional patience to hold the pieces together.

This is messy. It is also how most consequential things get built in India. The clean, single-funder model belongs to countries with different institutional architectures. India’s architecture is coalitional, and the pilot’s funding will be too.

What Actually Stands in the Way

#

The money is not the hard part. ₹17 crore is findable.

The hard parts, in order of difficulty:

Institutional tolerance for uncomfortable outputs. The skeptic produces findings that tell institutions their categories are wrong. The intent layer produces findings about why the categories are wrong, which implicates the institutions that set them. The retroduction layer produces evidence that the gaps in knowledge are structured by institutional incentives. Every component of this architecture produces outputs that are inconvenient to the institutions that would need to act on them. Funding the architecture is easy. Acting on its findings is hard. Funding it and then not acting on its findings is worse than not funding it, because it produces the appearance of epistemological rigor without the substance.

Data access. Indian health data is fragmented across systems that do not communicate. HMIS, RCH portal, Ayushman Bharat, state-level registries, each operates on its own infrastructure with its own access protocols. The DISHA bill, if enacted, would provide a regulatory framework for health data governance, but as of now the framework is uncertain. ISHI requires integrated data across sources. Getting that integration requires relationships, permissions, and patience that no budget line can purchase.

Community trust. The pilot requires sustained engagement with communities whose experience of being studied is, historically, that researchers arrive, extract data, publish papers, and leave. The communities in which the pilot would operate have been the subjects of research for decades. They have seen the programs come and go. They have reason to be skeptical of another project that promises to see them more clearly. Earning their participation requires time, presence, and the kind of reciprocal relationship that research timelines and budgets do not naturally support.

Philosophical depth without tokenism. The seven traditions cannot be reduced to seven modules. Each has internal diversity, contested interpretations, and scholars who would reasonably object to their tradition being operationalized as a test applied by a machine learning model. Engaging each tradition honestly requires sustained intellectual partnership with scholars who may not share the project’s assumptions. This is not a consulting engagement. It is a collaboration, and collaborations require mutual respect, shared governance, and the willingness to be changed by the encounter.

Sustainability beyond the pilot. Eighteen months produces evidence. It does not produce an institution. If the pilot works, what happens next? Who maintains the systems? Who updates the corpora? Who ensures the skeptic’s independence when the institutional pressures to smooth its outputs begin, as they inevitably will? The pilot budget does not include the cost of building the institutional home that would need to exist for the architecture to survive beyond the pilot period.

The Arithmetic That Never Works

#

I wonder sometimes whether the reason this kind of architecture has never been built is not that it is expensive or technically difficult but that it requires an institutional posture that institutions are not designed to hold.

The posture is: we believe our current categories are insufficient, we are willing to fund the discovery of their insufficiency, and we will act on the findings even when the findings implicate our own decision-making. No institution in history has sustained this posture voluntarily. Regulatory bodies are forced into it by crises. Scientific communities are forced into it by paradigm shifts. Governments are forced into it by consequences too visible to ignore.

The argument for building the architecture before the consequences is the same argument the series has been making since Part 74: the cheapest time to interrogate an objective function is before it runs. The most expensive time is after the consequences have compounded.

This arithmetic has never once, in the history of institutional decision-making, been sufficient to produce the investment.

But the arithmetic has never been wrong either.

The Notebook and the Budget

#

The man has a page with a number circled on it. ₹17 crore. Two underlines.

He knows the number is not the obstacle. He has watched enough institutional budgets to know that the money exists for things institutions want to fund. The obstacle is that this project asks institutions to fund the examination of their own assumptions, and institutions are not built to want that.

He also knows something from thirty-three years in healthcare that the budget does not capture. He knows that the women in the PHCs in the pilot districts are already living inside the stratum gap. They are already experiencing the consequences of categories that do not fit their lives. They do not need a pilot to tell them this. They need the systems that serve them to know it.

The pilot is not for them. The pilot is for the institutions. It is the instrument by which institutional knowledge is forced to confront its own insufficiency, documented rigorously enough that the confrontation cannot be administratively absorbed.

Whether any institution will commission this confrontation voluntarily, before the consequences force it, is the question the notebook keeps asking and the budget cannot answer.

He turns the page. On the other side, facing inward, is a question he wrote down three weeks ago: “If the system could flag the moment when its own categories stop fitting a person’s life, would anyone build it? And if they built it, would anyone act on the flag?”

He does not know. He is fifty-three. The number is circled. The question is unanswered.

The tire still has a slow leak.

This is the fifth and final essay in The Insufficient, a sub-series of The Approximate Mind. The series examined what lies beneath the empirical record AI systems are built to search: the skeptic that questions categories, the traditions that provide seven ways to doubt, the intent upstream of the specification, and the retroductive method for working backward from outcomes to undocumented mechanisms. This essay provides the practical assessment: what the architecture costs, who might fund it, and what stands in the way. The answer to the last question is not money or technology. It is institutional willingness to fund the examination of institutional assumptions. The architecture is buildable. The question is whether anyone commissions it before the consequences make the commissioning unavoidable.

References

#

Health Systems Research in India

Reddy, K. Srinath, et al. “Towards Achievement of Universal Health Care in India by 2020: A Call to Action.” The Lancet, vol. 377, no. 9767, 2011, pp. 760-768.

Rao, Mohan, et al. “Human Resources for Health in India.” The Lancet, vol. 377, no. 9765, 2011, pp. 587-598.

Maternal Health in India

Montgomery, Ann L., et al. “Maternal Mortality in India: Causes and Healthcare Service Use Based on a Nationally Representative Survey.” PLOS ONE, vol. 9, no. 1, 2014.

Registrar General of India. Special Bulletin on Maternal Mortality in India. Sample Registration System, various years.

Data Governance and Health Information Systems

Rajan, S. Irudaya, and K.S. James. “Third National Family Health Survey in India: Issues, Problems and Prospects.” Economic and Political Weekly, vol. 43, no. 48, 2008.

Institutional Design and Public Interest Technology

Mazzucato, Mariana. Mission Economy: A Moonshot Guide to Changing Capitalism. Harper Business, 2021.

Jasanoff, Sheila. The Ethics of Invention: Technology and the Human Future. W.W. Norton, 2016.

Power, Michael. The Audit Society: Rituals of Verification. Oxford University Press, 1997.

Indian Philanthropy and Health Innovation

Sheth, Arpan, et al. “India Philanthropy Report.” Bain & Company, various years.

Critical Realism Applied

Pawson, Ray, and Nick Tilley. Realistic Evaluation. SAGE Publications, 1997.

Danermark, Berth, et al. Explaining Society: Critical Realism in the Social Sciences. Routledge, 2002.