India's digital infrastructure can capture a tap in 22 languages. It cannot yet tell whether the person understood what they tapped. These two demos attack that gap from two different layers.
Fixing that needs two things: a standard way to record comprehension, and a fast, multilingual way to deliver explanations at population scale. Each demo below isolates one of these, so each can be judged on its own terms.
A simulated loan-app eKYC consent. Toggle between today's flow and the comprehension layer, in English or Hindi, with the checkpoint read aloud by Sarvam's Bulbul voice. Nothing here requires a server-side model at demo time — it cannot fail in a room.
Mode “Today's flow” → tick the box → tap I Agree. The consent artefact appears on the right. The comprehension slot stays empty: the tap was captured; comprehension was not.
Mode “With comprehension layer” → हिं → I Agree. A grade-6 explanation appears — play the orange button to hear it. Answer the teach-back question wrong on purpose. Watch it re-explain more simply.
Pass the check. Both artefacts appear side by side — open “show machine record” to see the comprehension-artefact/v0.1 schema: language, modality, attempts, probe outcomes, timing.
DPI standardised permission with the consent artefact. This demo shows its missing twin — a verifiable, auditable record that the user understood a consequential action. Fail the check twice and the flow pauses, recommends assisted help, and writes no consent at all. The record is honest in both directions. This is the piece that can become a standard.
A checkpoint only works if the explanation arrives instantly, in the user's language, by voice. These demos show the architecture that makes that affordable at a billion-user scale: pre-generated narration served from a Redis cache (Tier 1), live model generation only on cache miss (Tier 2).
Open the eKYC duplex page, pick a language, and ask aloud — “kya share hoga?”, “wapas le sakta hoon?”. Speech is transcribed, matched to an intent, and answered from pre-built audio chunks with near-zero wait.
The tier-compare page runs the same request both ways: Tier 1 cached narration vs Tier 2 live generation. The latency difference on screen IS the business case for preprocessing.
Every cached chunk is exactly the kind of explanation Path A's checkpoint needs. Path B is the engine; Path A is what the engine is for.
Comprehension at population scale cannot add per-transaction AI cost. The four-tier cascade means AI runs only when it matters: the "million journeys" are generated once offline and cached at the edge; materiality routing reserves full live reasoning for high-stakes, irreversible moments.
Frontier models are overwhelmingly English-trained; Indian languages are a decimal fraction of their corpus. The current architecture is honest about this: Claude reasons in English, where the digital procedural lexicon is native; Sarvam localises and voices in the user's language, in culturally native Indic voices. Comprehension requires both intelligence and intimacy — which is exactly why the comprehension layer is the part a sovereign stack should own next.
In the full architecture, the record sits on top of the delivery engine.