Forwarded Documents, Not Portals
Draft, deploy-gated. The full narrative publishes with the Lab launch; the sections below are content-free by construction.
The obvious way to pull a family’s schedule into software is to log into the school portal, the camp portal, the sports league portal, and scrape them. We chose the opposite: a parent forwards us one document, and we propose work from it. No credential custody, no brittle scrapers, no auto-reading an inbox.
Why portals lose
Portal scraping looks impressive and ages badly:
- Credential custody. Scraping means holding someone’s login. That is a liability you do not want for a family’s school account.
- Brittleness. Portals change markup constantly. A scraper is a maintenance treadmill that breaks at the worst time, the week before the deadline.
- Over-reach. Logging into an inbox or portal means seeing everything, not just the one flyer the parent cared about.
Why forwarding wins
Forwarding inverts every one of those problems. The user hands us exactly one document, on purpose. We never hold a credential, never auto-read an inbox, and never see anything the parent did not choose to send. The input is explicit, scoped, and consented by the act of forwarding.
From that single document, a read-only runner extracts proposals, dates, deadlines, who each item is for, and routes them through the same approval spine as everything else. The runner reads; the parent decides.
What we ingest, and what we never touch
The Lab makes the boundary visible. We ingest the forwarded document and derive a closed set of structured fields from it. We never touch the rest of an inbox, never store credentials, and never expose the raw document on any public surface. The public projection shows the coarse source-type mix, school, camp, sports, medical, plus counts and confidence bands, never the document itself.
Shared-source learning without leaking households
The same flyer often lands in many families’ inboxes, the district-wide closure notice, the league schedule. When the same public document is forwarded by multiple households, the system can learn the shape of that source once and apply it broadly. That learning is about the source, not the household: no family’s data is pooled, compared, or leaked to another. Shared-source learning improves extraction quality while each household’s contents stay isolated.