What gets measured, and how the system learns from it.
Yesterday's zoom ended with a question: where does the feed's brain get its inputs? Here. Every scroll, pause, save and forward becomes a small factual event in one first-party stream. That stream feeds two learners — the feed (what should SHE see tomorrow?) and the content strategy (what should the agents MAKE next?) — and it does so under privacy rules that are built into the data's shape, not promised in a policy document.
What actually gets recorded.
Six kinds of fact, all first-party, all tied to a pseudonym rather than a phone number. Nothing else — no contacts, no location trail, no third-party trackers.
Seen
This card appeared on her screen, in this position, on this morning.
Lingered
How long a card held her — the quiet signal that something landed.
Skipped
Scrolled past fast. As informative as a like, and far more common.
Saved / liked
Deliberate, reversible choices — stored as state she can see and undo.
Forwarded
She sent it to someone on WhatsApp. The north-star event — the whole system optimises for this.
Chosen at sign-up
Language, region, the gods she cares about — declared taste, kept separate from observed taste.
The loop — yesterday's behaviour, tomorrow's feed.
Follow it clockwise from the phone. Events flow into one append-only stream; overnight, the stream is boiled down into features (her taste points shift, what she skipped counts gently against); the feed brain from the last zoom reads those features and leaves a new ready-list for her morning; she reacts, and the loop closes. One full turn takes a day — by design.
She scrolls, forwards two cards
Each action becomes an event: seen, lingered, skipped, forwarded.
Events land in one stream
Append-only, pseudonymous, exactly-once. The system's only source of behavioural truth.
The stream becomes understanding
Her interest points move on the taste map; per-card counters update (what forwards, what gets skipped); the AI reads aggregates and rewrites her one-paragraph profile.
A new list is waiting
The feed brain re-picks her ~50 cards with the fresh understanding. Tomorrow is measurably different from today.
The same events teach the makers.
The loop above tunes who sees what. A second reader of the same stream tunes what gets made at all.
Forward-rates by recipe
Which deity × language × style combinations actually get forwarded — not which single card went big. The system amplifies recipes, not one-hit dishes.
Straight into the briefs
The Creative Head agent reads those aggregates (and the gaps — zero English Shiva cards, say) before writing the next round of Content Briefs. Measured demand becomes next week's production plan.
Analytics beside, never in between
Dashboards and experiments read the same stream as a sibling. The recommender never consumes data "through" an analytics vendor — one hop fewer to break, and no third party in the loop.
What the loop deliberately does not do.
No real-time adaptation
The feed never rearranges itself mid-scroll to chase your attention. Learning happens overnight — that single choice removes the doom-scroll mechanic at the architecture level.
No model reads raw events
The AI only ever sees aggregates and summaries, offline. Nothing a user does is piped into a live model context.
No loop for minors
A known-child account gets the contextual feed — festivals and quality only, behavioural inputs zeroed. Not a softer version of the loop; the absence of it, as India's DPDP law requires.
No identity in the lake
Events carry a pseudonym; the mapping to a real account lives in one guarded table. Erasure deletes the mapping and the state, and the remaining events point at nobody.