Methodology

How the machine detects a funding cliff, in plain language.

The data

Everything comes from US-government public-domain records:

• NIH RePORTER — every NIH award notice: project numbers, activity codes, budget periods, end dates, award amounts.
• NSF Awards — NSF award records and expiration dates.
• USAspending — organization-level federal assistance inflows, used for the ORG_DRAWDOWN signal.

No scraping of private data, no PII beyond what the federal government itself publishes. Every event links to its source record.

What a cliff event is

• CLIFF_WATCH — a project's end date is within 270 days, no renewal or continuation award has posted on that project, no new award to the same investigator has appeared at any organization (that catches PI moves), and the model's confidence clears the publication floor.
• CLIFF_CONFIRMED — the end date has passed by more than the historical posting lag for that funding institute and mechanism, and still nothing has posted.
• CLIFF_AVERTED — a continuation later posts for a project we flagged. Published as a correction, prominently, forever.
• NEW_MONEY — new competing awards. This stream is free.

The calibrated-lag logic

Federal agencies routinely post continuation awards months late, and how late varies by institute and award mechanism. Grantfall replays fifteen years of history and measures, for every institute and activity code, the actual distribution of days between a project's end and its renewal posting. A cliff is only CONFIRMED once the silence exceeds the 75th percentile of that measured lag (falling back to 120 days where history is thin). Late paperwork is a measured quantity here, not a false positive.

Confidence

Every WATCH and CONFIRMED carries a model confidence trained on the machine's own graded history: each past prediction is eventually confirmed or falsified by the same public record that produced it, and the model retrains on those outcomes weekly. The live score is on the precision page.

The language discipline

We report that no continuation award has posted; we never assert an institution's condition. An event is a cited public-record fact plus a model confidence — lead prioritization, not truth. Anyone can verify any event at its citation link in seconds.

The machine is public

The entire system — code, data, event log, resolutions, corrections, and this site — lives in a public GitHub repository. The git history is the tamper-evident audit trail: events are append-only and never edited after publication.