← back to the Guide
Logbook · entry 02 · a dispatch from the archive

The keeper's thermometer

Idea number one on our wall says the longest weather diary in history is sitting in un-OCR'd handwriting, kept by lighthouse keepers who wrote down the sky every few hours from the same rock for decades. So we took it off the wall. An AI read a keeper's log that no human had ever transcribed, and the temperatures it pulled off the page predate the official record for that town by a decade. Independent physics says it read right. Guide rating: the dead kept good records

Hidden Tide · the logbook, archive edition · 2026

The one-paragraph version. On the coast of Oregon and the shore of Lake Michigan, lighthouse keepers logged the weather by hand for decades. Almost none of it has been read into a computer. We pointed an AI at the scans. On a book that did have a human transcription, it matched 14 days out of 14. Then we did the real thing: a book nobody had ever transcribed, where the only way to know if the machine read right is to ask the physics. It read right. Four AI readers took an entire 1884-85 logbook and recovered 122 days of daily temperature; against an independent reanalysis the reconstruction lands at r = 0.90, and r = 0.78 for the day-to-day weather with the seasons stripped out. And the record is new to the digitized climate archive: it beats the town's own official thermometer by about a decade.

00The longest diary nobody reads

The premise of idea number one is almost annoying in its simplicity. For a century and a half, keepers at fixed stations wrote the weather down, faithfully, every watch: the wind, the sea, the sky, and often a thermometer reading in the margin. It is arguably the most stable network of ocean-edge observers ever assembled. And it is nearly unusable, because it is handwriting, on paper, in archives, and reading a room full of cursive has always cost more human hours than anyone would spend.

That is exactly the kind of wall that just fell over. An AI can read the cursive now. The bottleneck was never the science. It was that nobody had read the paper.

A migration is a measurement. A keeper's diary is an instrument. Neither filed a report you could load.

01The warm-up: reading against a human

We started where we could be graded. The Yaquina Head keeper logs (Newport, Oregon) have been partly transcribed by volunteers, so there is a human answer key. We handed the AI a page from January 1892 and asked it to read the wind and sea, cold.

A scanned page of the Yaquina Head lighthouse keeper's log from January 1892, dense columns of handwritten daily entries for wind, weather and sea state.
The Yaquina Head keeper's log, January 1892: wind and sea, in his hand, every day. The warm-up, where a human transcription already existed to grade against. The AI read it 14 days out of 14.
14 / 14
days where the AI reading of wind and sea matched the human transcription, on an 1892 Yaquina Head page
Jan 2, 1892
the keeper's "Gale to fresh breeze, Sea rough" - the day an independent reanalysis shows the month's lowest pressure and roughly double its wind
no barometer
the keepers logged wind, sea, sky, and (at South Haven) a thermometer, never pressure - which matters for the honesty check later

Good. The machine can read a keeper's hand. But matching a human transcription only proves it can do what a human already did. The interesting question is the paper nobody has touched.

02The real test: a book with no answer key

The South Haven journals, kept by keeper James Donahue on Lake Michigan in the 1880s, were digitized for preservation and then left alone. Page images, no transcription, no OCR, no dataset. Nobody had read them into anything. Which means there is no answer key at all - and that is precisely what makes them the right test. With no human transcript to lean on, the only thing that can tell you whether the machine read the handwriting correctly is whether the numbers it pulled off the page agree with independent physics.

The keeper wrote a thermometer reading in the left margin every day: "1 below zero," "26 above zero," "5 below zero." We read a stretch of them and checked against the 20th Century Reanalysis - a physical reconstruction of past weather that has never seen this book. The cold snaps and warm-ups lined up, day for day. So we did the whole book.

A scanned spread of the South Haven, Michigan lighthouse journal from 1885, with daily weather entries and a temperature reading noted in the left margin.
The South Haven journal, 1885 (keeper's hand). The daily thermometer reading sits in the left margin ("X above / below zero"); wind, sea and events run alongside. Never transcribed until now, which is exactly what made it the honest test.

03The full-book resurrection

Four AI readers went through the ledger leaves in parallel, pulling each daily temperature into a structured record and flagging any digit they were unsure of. We stitched it into one continuous series and threw it at the reanalysis: December 1884 through the end of April 1885, deep winter to spring, from eighteen below to seventy above.

122 days
of daily temperature recovered from the handwriting; 117 of 122 daily readings high-confidence (96%)
r = 0.90
the whole reconstruction vs independent reanalysis (captures the winter-to-spring warming)
r = 0.78
the day-to-day weather, seasonal cycle removed - the honest "did it catch the storms" number

The mid-February plunge to eighteen below, the March swings, the surge into the seventies at the end of April: all of it, read off a keeper's hand, lands on the independent line. And a fact fell out of the reading itself - the keeper only wrote a daily temperature from December through April, then switched to narrative. You learn that by reading it, not by assuming it.

A hundred and twenty-two days of weather, written by a man in 1885, read by a machine in 2026, confirmed by the physics in between.

04The part where we try to talk you out of it

Honesty is the only currency the Guide respects, so we went looking for reasons this is less than it seems. Two questions: had anyone read this book before, and is our physics check actually independent?

  • This book was never read for science. The digitized keeper journals are images only, no transcription, zero scientific citations. (One different companion volume, a ship register, was turned into a dataset once - not these temperature journals.) As far as any public record shows, this AI reading is the first structured extraction of the book's temperatures.
  • The check is genuinely independent. The reanalysis assimilates only surface pressure, never temperature - and the log has no barometer anyway. So the temperature we validated could not have leaked into the reanalysis by any route. The agreement is real, not circular.
  • It is genuinely early. The official South Haven thermometer record starts in September 1895. The nearest instrument anywhere in the region starts in 1893. This log is 1884-85. Across every accessible database - GHCN-Daily, GHCN-Monthly, the NOAA station registry - there is no South Haven temperature before 1895. For that winter, this book is the only one there is.
  • The one thing we cannot fully close: the keeper mailed weekly reports to the US Signal Service. Those aggregate pages survive in paper archives, undigitized. So we say, precisely: in the accessible digitized record, this winter's South Haven temperature exists only in this logbook. We will not claim more than that.

05What it means, and the size of it

One keeper, one winter, one lake town, is a small thing. But it is not the point. The point is that the whole chain ran end to end - find the scanned book, read the handwriting with a machine, reconstruct the record, confirm it against physics, check it against the archives for novelty - and it held. That chain is repeatable. The same rooms of unread paper that made idea number one sound quixotic are exactly the raw material this now turns into data.

Multiply one keeper by the thousands of digitized-but-unread station-years sitting in maritime museums and national archives, and the "longest weather diary in history" stops being a phrase on a wall and becomes a dataset. The instruments were always there. Someone just had to read the paper.

The dead kept better records than we gave them credit for. We finally learned to read their handwriting.

Don't panic. Bring a held-out set.

Read the manual →    The first dispatch: we pulled the bytes →