About the project

It maps the record, not meteor frequency

Digitized newspaper coverage varies greatly by year and state. Reprints and OCR quality also affect the dots. The timeline's separate coverage line shows newspaper-title counts from available local sweep statistics. Missing years are unknown, not zero. The article counts are available in coverage.json.

Two different kinds of pin

Fall cases use approximate reported localities. Fireball groups use one representative publishing town: the resolved newspaper location closest to the other publishing locations in that group. These points are not impact sites or reconstructed witness positions. Fireball marker size uses the upstream grouped-city count, which is only a proxy for independent reporting. Read the field definitions ↗

Not a complete census

The atlas includes groups with at least two grouped publishing cities, or a human-review flag. Some candidate groups can still contain misclassified or unrelated sky reports. Dates retain the input's precision and uncertainty. The 1770–1964 axis describes the project's scope, not continuous observations. Release selection and limits ↗

Five of seven known falls recovered

The 1870s recovery test found five of seven recorded U.S. falls: 71%, with a wide approximate 95% interval of 36–92%. A full-year name rescan found no relevant reports of the two misses. The project attributes them to corpus coverage, although reports lost or garbled by OCR cannot be ruled out. Five out of five among recoverable cases is not evidence of perfect overall recall. Recovery test and rescan ↗

The blind check changed the method

A sample of 100 low-ranked fall reports found seven specimen destinations the ranker had missed. The rules were revised. Because that sample was used to tune the changes, it cannot independently validate the revised rules. Blind-check results ↗

Human review has limits, too

In a blind excerpt review, the human reader agreed with 22 of 24 decided model judgements; four more were undecidable. That small comparison is not a general accuracy score. Garbled OCR, clipped excerpts and plausible hoaxes remain sources of error. Human check ↗ · Audits and pilot limits ↗

The release includes the 13 curated fall cases, selected pilot fireball groups, source links and coverage statistics. Nested CSV fields contain JSON. Source text excerpts are short and preserve OCR unless a correction is explicitly marked. The release is a snapshot, not a live feed.

Regenerate from the project root with python -m skyinprint.export_public. New sweep statistics are picked up automatically; expanding the event catalogue requires updating the upstream grouped and reviewed events first. Reproduction details ↗

Library of Congress / Chronicling America provides the newspaper scans. American Stories (Dell et al.) provides the OCR corpus. The reference meteorite snapshot comes from NASA's Meteorite Landings dataset and the Meteoritical Bulletin.

Historical newspaper astronomy has predecessors. The Historical Fireballs project, along with the Nininger and Monnig collections, informs this work's source-checking approach. The pilot makes no claim to be the first project to study newspaper fireballs.

Map boundaries: US Atlas 3.0.1, from U.S. Census Bureau cartography. Maps use vendored D3 7.9.0 and topojson-client 3.1.0. Typeface: Barlow, with local system-font fallbacks. Derived data: CC BY 4.0; project code: MIT.

Institutional updates are attributed to the institution, dated, and distinguished from publicly available historical sources. Private correspondence, staff identities, outreach logs and supplied attachments are excluded. Read the public update record ↗