Home · Level 1 · Session 1 Ukubona LLC ↗ Automation ↗

🌱 Level 1 of 5 · Session 1 of 5

Level 1 · Foundations

The Existing Process

How Evidence Becomes Policy — and Where It Doesn't

What is the current pipeline for literature review at WHO India, and which steps are consuming time that AI could return to you?

∂

v4 · TMVES compression chain · five phases

The statistical backbone of this curriculum — from raw world-model θᵗ through composite loss, gradient policy, federated correction, to prognosis loop L(θᵗ⁺¹). Skip if equations aren't your thing. Open if you want to see why.

For the curious

▾

You don't need this to use the tools in Sessions 2–4. This section exists for participants who want to understand why the pipeline stages are ordered the way they are — and what a formal decision-theoretic account of evidence synthesis would look like. If that's not you today, keep scrolling.

I · Tensor

θᵗ Temporal Twin

The agent's full world-model at timestamp t. Raw input: what is known, what is assumed, what is missing.

VoI source →

II · Matrix

L₀+Σwᵢ·Lᵢ Composite Loss

Baseline loss L₀ plus attention-weighted document losses. Off-diagonal: relevance wᵢ shapes what evidence enters the synthesis.

Attention layer →

III · Vector

f(σ²,λ,ε) Ex-Ante Gradient

Decision-time policy. σ²=diagnostic variance, λ=irreversibility penalty, ε=noise. Returns η_safe — how much to act now.

E[Δuncert]/cost →

IV · Eigenmode

γ|ε_FGT|² Ex-Post Credit

Federated ground truth correction — arrives after the decision. γ=intervention magnitude. Updates weights w_ijk. Delayed by design.

Learning time →

V · Scalar

L(θᵗ⁺¹) Prognosis Loop

Projects the updated state into the next loss landscape. The loop closes: today's decision reshapes tomorrow's prior.

Action / brief

v4 upgrade — temporal separation: Phase III f(σ²,λ,ε) operates at decision time (ex-ante). Phase IV γ|ε_FGT|² operates at learning time (ex-post, delayed). These are not the same moment. A recommendation made before the federated ground truth arrives is made under FGT=∅ — which means weight updates are frozen and the eigenmode enforces: hold. VoI = E[Δuncertainty] / η_cost now selects which descent step to take next, replacing the older raw η ordering. In evidence synthesis: the highest VoI action is almost always waiting for the decisive information rather than acting on incomplete screening.

Where were you headed?

In our first session, before we looked at a single tool or pipeline diagram, we asked one question: when you set off on a literature review, what is your destination?

The answers were spread across the full length of the evidence pipeline — and every single one was correct. They just weren't all the same destination. That gap is what this curriculum is designed to close.

Click the answer that most closely matched yours.

Session 0 · Your answer

"When you set off on a literature review — what is your destination?"

A database search result

PubMed hits, Scopus exports, a set of titles to read

Phase I · Tensor · θᵗ

A screened list of papers

Titles and abstracts triaged for relevance

Phase II · Matrix · L₀+Σwᵢ·Lᵢ

An extracted dataset

Key findings, numbers, methods pulled from papers

Phase III · Vector · f(σ²,λ,ε)

A synthesis or summary

What the literature says, in aggregate

Phase IV · Eigenmode · γ|ε_FGT|²

A policy brief or recommendation

A decision-ready document for leadership

Phase V · Scalar · L(θᵗ⁺¹)

Where your destination sits in the compression chain

Your boss pointed here: the scalar — a policy brief or recommendation. That's Phase V: L(θᵗ⁺¹), the prognosis loop. Everything else is a step on the way, not the destination. This session maps exactly what happens between where most people stop and where decisions get made.

Why this matters Every answer in that first session was honest. But the further left your destination sat, the more work was being left on the table — work that AI can now do. The pipeline below shows where time disappears, and the rest of this curriculum shows you how to get it back.

The pipeline, honestly described

Every evidence-to-policy cycle at WHO India moves through a recognisable sequence of steps. The process is not broken — it has produced genuine results, from informing PM-JAY design to shaping national NCD guidelines. But it is slow, and the slowness is not random. It concentrates at specific, identifiable points that are exactly where AI tools offer the most leverage.

Here is the pipeline as it typically runs for a health financing or UHC evidence brief:

Primary pipeline · five-phase compression chain · TMVES

I · Tensor

θᵗ

Policy question defined

Database search

II · Matrix

L₀+Σwᵢ·Lᵢ

Title / Abstract screening

III · Vector

f(σ²,λ,ε)

Full-text retrieval

IV · Eigenmode

γ|ε_FGT|²

Data extraction

Synthesis

Brief drafted

V · Scalar

L(θᵗ⁺¹)

Policy decision

Red steps — Database Search (I), Title/Abstract Screening (II), and Data Extraction (IV) — account for 60–70% of total review time. They map directly to Phases I, II, and IV of the TMVES chain. Session 2 maps 20 AI tools against exactly these three steps. The algebra terms are optional reading; the red blocks are not.

Session 2 tool map →

∂

Deep dive: the v4 compression loss chain

          θᵗ → L₀+Σwᵢ·Lᵢ → f(σ²,λ,ε) → γ|ε_FGT|² → L(θᵗ⁺¹) · five phases · ex-ante / ex-post separation
        

Five phases · algebra

▾

θᵗ → L₀+Σwᵢ·Lᵢ → f(σ²,λ,ε) → γ|ε_FGT|² → L(θᵗ⁺¹)

Five layers of AI in evidence work

Before mapping tools to tasks, it helps to know what kind of AI you're dealing with. The taxonomy below — drawn from how AI actually enters health systems — maps each layer to its role in a literature review pipeline. Read it against the pipeline above: World AI and Perception AI operate at the search end (Phases I–II); Agentic AI handles extraction (Phase III); Generative AI produces the synthesis (Phase IV); Embodied AI audits the whole chain for bias, equity, and compliance (closing the loop into Phase V).

AI Layer Taxonomy · Evidence Synthesis

From embodied wisdom to the world's raw data — five layers, five roles

▾

Layer	Role in Evidence Synthesis · v4 Phase
Embodied AI	Audits the entire workflow — bias detection, equity checks, PRISMA compliance · Phase V · Scalar · L(θᵗ⁺¹)
Generative AI	Synthesizes findings into reports, policy briefs, and insights · Phase IV · Eigenmode · γ\|ε_FGT\|²
Agentic AI	Automates screening, extraction, and repetitive tasks · Phase III · Vector · f(σ²,λ,ε)
Perception AI	Ingests specific literature across journals, languages, and grey sources · Phase II · Matrix · L₀+Σwᵢ·Lᵢ
World AI	Understands general science across all domains · Phase I · Tensor · θᵗ

No standard classification exists — this taxonomy is inspired by how AI actually enters health system procurement and workflow. Full reference →

The red steps account for roughly 60–70% of total review time in a standard health economics evidence review. They are also the steps most likely to be done inconsistently under time pressure, which means the quality of the final brief is often determined not by the synthesis but by what got dropped during screening.

The four bottlenecks specific to WHO India

Generic reviews of "how literature review works" are not what this session is for. The friction that WHO India financing and economics staff experience is specific. Here are the four most consequential pressure points:

The grey literature gap

A significant share of the evidence most relevant to Indian health financing — NSSO household surveys, state NHA accounts, NITI Aayog working papers, HTAIn assessments — never reaches PubMed or Cochrane. Standard database searches systematically miss it. Screening 200 journal abstracts while missing the 2023 HTAIn report on PM-JAY equity outcomes is not a search strategy; it is a selection bias.

The WEIRD-data bias

The indexed literature over-represents high-income country health systems. A search for evidence on community health insurance and catastrophic expenditure returns studies from the Netherlands, Canada, and South Korea before it surfaces comparable work from Kerala, Andhra Pradesh, or Rwanda. Without active filtering, AI tools inherit this bias — and can amplify it by defaulting to the most-cited results.

The clinical-economics confusion

Most literature review tools — and many default AI prompts — are designed around clinical questions. Searching for "impact of health insurance on health outcomes India" returns RCT literature on specific diseases rather than economic evaluations of insurance coverage, financial risk protection, or utilisation equity. The question framing determines the evidence retrieved. This is addressed directly in Session 3.

The brief-to-decision gap

Even well-conducted reviews regularly fail to reach decision-makers in usable form. A 45-page systematic review of health financing in LMICs does not serve a Joint Secretary at MoHFW under a two-day deadline. The translation step — from evidence synthesis to a one-page brief calibrated to budget constraints, political context, and equity commitments — is where most reviews stall. This is the subject of Session 5.

What AI can and cannot fix

It is worth being precise about where AI tools intervene in this pipeline — and where they do not.

AI accelerates: database search construction, title and abstract screening, full-text data extraction, structuring a synthesis across multiple studies, drafting a first version of a brief. A task that takes two researchers three weeks of screening can often be reduced to a focused afternoon of prompt refinement, AI-assisted screening, and human verification.

AI does not replace: the judgment call about which question matters, the contextual knowledge that a particular Indian state has no functional secondary care infrastructure, the equity lens that asks who benefits and who is excluded, the political reading of what a Ministry will act on. These remain yours. This curriculum is about using AI to protect your time for the parts that only you can do.

A 2023 analysis of India's National Health Programme guidelines found that systematic review evidence was formally cited in fewer than half of NHP guideline documents reviewed — not because the evidence did not exist, but because the pipeline from production to decision was too slow and too fragmented to be useful at the moment of decision.

That gap is the reason this curriculum exists. The tools in the next session are not there to make literature review more academic — they are there to close the distance between evidence and the decision that needs it, before the window closes.

Audit your own process

🧪 Self-audit — where does your review time actually go?

For a recent or ongoing evidence brief, estimate how much of your total time falls at each stage. This takes 90 seconds and generates a personalised read on where AI will give you the most back.

Defining and scoping the question Clarifying the policy context, agreeing on the research question, setting inclusion criteria

Database search & grey literature hunt Building search strings, running searches, tracking down reports and government documents

Title / Abstract screening Reading through results to determine relevance — typically the single largest time sink

Data extraction Pulling numbers, methods, and findings from included papers into a usable format

Writing the synthesis and brief Turning evidence into a usable document for decision-makers

🎯 Key takeaway

The existing process is not the enemy — it is the baseline. The three steps that consume the most time — database search, title/abstract screening, and data extraction — correspond to Phases I–III of the TMVES chain: θᵗ → L₀+Σwᵢ·Lᵢ → f(σ²,λ,ε). These are precisely the steps Session 2 maps to 20 AI tools. The goal of this curriculum is not to replace your judgment but to compress the mechanical steps so that more of your working week is spent on the parts that require it: defining the right question, applying the equity lens, and translating findings into something a decision-maker can act on — Phase V.