Ukubona · WHO India Evidence Synthesis · Layer IV
Generative AI · Phase IV · Eigenmode

Generative AI

V4 Phase IV · Eigenmode γ|ε_FGT|² synthesis · reports · policy briefs · insights

The layer your stakeholders actually see — transforming extracted evidence into synthesis narratives, policy briefs, and decision-ready insights. Think ChatGPT. Think more carefully than that.

What it is

Language generation from structured evidence inputs

Generative AI synthesises findings into human-readable outputs — reports, policy briefs, executive summaries, and narrative interpretations of data. It operates at Phase IV (Eigenmode, γ|ε_FGT|²), where the evidence has already been extracted and structured by Agentic AI, and now needs to be compressed into argument.

The shorthand "think: ChatGPT" in the original table is both accurate and insufficient. ChatGPT is an instance of Generative AI. But Generative AI in evidence synthesis is not chatting — it is drafting policy-grade documents from curated inputs, with specific output formats, citation requirements, and audience constraints.

Note on the "think: ChatGPT" shorthand — The reference is to the generation capability, not the product. In practice, Claude, Gemini, and domain-fine-tuned models often outperform ChatGPT on structured synthesis tasks. The product is less important than the prompt engineering and the quality of the structured input it receives.
Outputs it produces

What Phase IV actually delivers

Policy Brief
4–8 page decision-ready synthesis for ministry or WHO leadership
Evidence Table Narrative
Prose interpretation of extraction tables, contextualised for setting
GRADE Summary
Structured certainty-of-evidence statements across outcomes
Executive Summary
1-page synthesis for non-technical stakeholders
Discussion Section
Contextualisation of findings against India-specific health system constraints
Recommendation Draft
Provisional recommendations with evidence grading, for human revision
WHO India · Practical example

After agentic extraction of 140 papers on CHW cost-effectiveness, Generative AI produces a first draft of a WHO-formatted policy brief — including regional variation summary, ICER comparison table narrative, India-specific context paragraph referencing ASHA programme data, and tiered recommendations for the NHM technical committee. Human economists revise and validate before submission.

Critical distinction

Synthesis from evidence versus generation from memory

The failure mode of Generative AI in evidence work is using it as a substitute for evidence — asking it to generate findings it was not given. A well-configured synthesis pipeline feeds the model extracted data and instructs it to synthesise only from that input. An ill-configured one asks it to summarise a research area from its training data, which may be outdated, biased toward English-language high-income-country literature, and impossible to verify.

The Eigenmode algebra (γ|ε_FGT|²) reflects this: the output is a transformation of structured evidence inputs, not a free generation. The γ coefficient is a weighting term — not all evidence is synthesised equally.

Audience calibration

The same evidence, four different registers

One of Generative AI's most practically valuable capabilities for WHO India is register switching: taking the same extracted evidence and producing a technical annex for economists, an accessible summary for programme managers, a bullet-point briefing for political advisors, and a plain-language version for community consultation — without the human team rewriting each from scratch.