method

the lanes

the corpus was produced by four parallel agent research lanes — company history (69 events, 66 sources), founder writing and early posting, coverage and sentiment, and the unfinished idea (85 cataloged sources) — each writing a lane report against one shared scratch source catalog before synthesis. the pattern comes from the stripe-history formula: bounded collections, typed sources, and a review ledger. roam’s version drops the live-updating machinery (the company is quiet) and deepens the qualitative lanes: what the founder said, what the discourse made of it, and what the idea still owes.

what the formula proved this run: the source catalog is the real shared artifact — lane disagreements resolve as catalog entries, not arguments. where it needed help: a shared id convention upfront (lanes minted overlapping names that had to be reconciled), an explicit per-claim confidence field from the start, and a declared owner for deduplication so two lanes don’t archive the same thread under different ids.

the evidence rules

1. every event, quote, and claim carries at least one source id.
2. dates keep their precision — a year, a month, or a day — and approximate dates are labeled rather than sharpened.
3. confidence stays honest: confirmed, reported, and inferred are distinct claims, not stages of the same claim.
4. sources are typed by role — primary, interview, reporting, archive, community — so a claim’s foundation is visible before it is read.
5. sentiment labels describe the coverage, not the truth of it; the same source can be evidence for both a hype wave and a backlash.

known gaps

1. roam’s founding date is reported across 2017–2019 depending on whether you count the prototype, the company, or the public beta — this dossier keeps the spread rather than picking one.
2. the white paper’s original roam-hosted url link-rots; the readable copy is a community mirror, and both are cited.
3. no public current usage or revenue numbers exist; arr figures are community-reported and labeled as such.
4. the twitter corpus comes through threadreader archives and the atlas.fm creator index — complete threads survive, but standalone tweets and engagement counts are partial.
5. the reddit corpus is thin by design: the subreddit’s history is contested and largely unarchived, so moderation claims lean on conor’s own threads and contemporaneous coverage.
6. trial length and believer-plan pricing drifted over time; the catalog keeps the prices each source actually recorded.
7. several sources are future-dated relative to the events they describe — karpathy’s 2026 gist, the roam-tools changelog — and are treated as retrieved primary material, not normalized backward.

corrections

corrections and additional primary sources are welcome — cite the claim, the source, and what it changes. the catalog is designed to be audited, not just read.