flip examples
Real, browsable flip notebooks — research with custody, grading, and corroboration discipline. No human typed the flip commands in any of these; each was built by an agent inside an ordinary working conversation. Every page below is rendered from the notebook itself.
nj-schools — did NJ enrollment actually dip?
Four NJ DOE fall-enrollment files under custody, graded, totals computed two independent ways. The dip was real but modest — and the decline since 2023-24 is now bigger than the pandemic dip.
nj-enrollment-forward — what happens next?
The forward-looking half: dated forecasts with probabilities, resolution criteria, and annulment conditions — the two-object rule (claims carry grades, forecasts carry probabilities) in practice.
rag-hallucination-lit — a small literature review
Papers captured as PDFs with hashes, graded, and read into claims — the lit-review kind's collection contract checked by the doctor from day one.
okf-origins — where did the format come from?
One natural-language reporting assignment becomes a traceable origin account, a six-source gap map, and a bounded roadmap verdict. The public history starts at a snapshot; planning is visible, a schedule is not.
ten-thousand-steps — a claim teardown
Where does the 10,000-steps-a-day target come from? A 1965 pedometer's trade name, not clinical evidence. The folk claim kept as data with its holder, the origin claim probed against the captured text, and the number superseded by what two cohort studies actually measured.
retraction-watch — the null that ships
Five citation classics audited for retractions and corrections. Four land honest nulls — absence claims carrying exactly the surfaces checked, each armed with a reopen trigger — and one (Séralini 2012) has the whole documented arc: retracted, republished, and a one-directional metadata trail.
ev-charger-watch — a question under pursuit
Is US public EV charging getting more reliable? A standing question walked end-to-end: evidence that answers narrower questions than asked, a zero-yield probe with its cause, a sharpened re-pose, a dormant sub-question, a banded answer with armed reopen triggers, and the refresh written as a commission contract.
autonomous-research-loop — the big one
A case study rather than a browsable notebook. Two private notebooks built by an unattended loop that ran 67 hours across 43 sessions — 380 captured sources, 82 claims, 47 questions closed, against 19/8/6 in the largest notebook above. Twelve verbatim excerpts with every redaction marked: a question that reopened 50 minutes after it was answered, two claims deliberately left unconfirmed, an honest null, and a session timeline no person produces.