Trust

The cheapest check

Every row on every lane links to its source — a GitHub repository, a lab’s own page, a filing on EDGAR. The fastest way to audit this section is to click one and see whether the row said the truth about it.

Known limits

• Ranks are real; star counts are not. Every count on these pages is a sample count — distinct accounts seen in GitHub’s public event stream, which carries only a small fraction of real starring. For composition: the stream holds roughly one starring event per 89 pushes in this window. Measured against real star counts on Sep 11, 2026 (the latest settled day with a snapshot pair), the stream carried about one starring event for every 3 stars actually gained across 8,105 watched repositories. Relative ranking survives sampling; absolute counts do not, which is why every count is labeled as in-sample and compared only within its own day.
• Fresh days are incomplete on purpose. A newly published hour holds a fraction of its final events and fills in over days. Fresh days are ranked within themselves (which holds up), then re-ingested once settled — settled through Sep 11, 2026 as of this generation.
• The star-history series is 27 days old. GitHub withdrew the API that let anyone reconstruct star history, so it exists only going forward: 9,365 repositories snapshotted daily since Aug 20, 2026. Trend lines appear here when there is enough series to draw one honestly — about two weeks.
• The attention floor of 5 is inherited, not measured. It comes from the monthly instrument’s published floor and will be re-settled by measuring a week of arrivals against it.
• The money lane’s readings are interpretations. Each deal filing’s plain-English card is written by a language model from the filing text alone. The filed facts (codes, dates, the document itself) are separate on every card, and the filings the model set aside are listed with its one-line reading of each — the cut is auditable, not asserted.
• The classifier is a model and it errs. AI-tooling calls are made by a language model on each arrival. The dimmed rows on What to try exist so its cuts stay visible; repositories with no description at all are dropped as presumed scratch work rather than guessed at.

The collectors, and how often they run

8 collectors run in five chains. Each chain writes a heartbeat as its steps complete, whatever they found, so the record shows the automation rather than asserting it — and because every lane is idempotent, a run that finds nothing is a successful run, not a silent failure. The list below is declared rather than read back from the heartbeats, so a collector that stopped writing shows up as missing instead of quietly disappearing from it.

This page is a snapshot, not a live readout. The schedule below is a standing rule and does not go stale, but the completion times were read when the page was generated, on Sep 15, 2026, and refreshing will not move them. The collectors keep running regardless; the page catches up the next time it is generated.