Essay · Findings

Three hours.

The morning at the point is not a forecast. It is a clock: the land breeze, the sea breeze that ends it, the tide that keeps solar time, the rain that waits for the afternoon. The station on the roof has now measured all of it, and a forecast only has to fill in what the clock leaves open.

27 September 2026 · From the station audit and the ground-truth log

At first light the wind at La Saladita comes off the land. It has been coming off the land all night, from the east-north-east, and it will keep coming until a little after ten. Then it stops, turns, and comes in off the sea. Between the two is the session. It is about three hours long, and it is the same three hours nearly every day of the year.

The wind clock

The weather station at the beach has been logging wind every five minutes since late 2023. Read for the moment the sea breeze turns onshore, the record gives a median of 10:25 in the morning, with the middle half of days between 09:56 and 10:55. By month the median moves inside a narrow band: 11:00 in January, 10:30 in February, 09:40 in March, 09:50 in June, 10:05 in July, 10:40 in August, 10:25 in September, 10:05 in October, 10:10 in November, 10:30 in December. The table has no April or May figure yet. From first light to the onshore turn is a median 3.3 hours. In most months the dawn wind is offshore on 97 to 100 percent of days. Two different anemometers, read independently, put the turn within 25 minutes of each other.

None of that is a forecast. A forecast is a claim about a day that could go several ways. The land breeze at dawn is not a claim. It is a property of a valley that cools overnight next to an ocean that does not, and the record says so with a spread you can set a watch by. The wind clock page lays out the full table.

The morning is not a forecast. It is a clock.

The tide clock

The tide keeps time too, and the reason is the sun. The site’s tide model is a harmonic fit to the Zihuatanejo gauge, 33 constituents over two years and four months of hourly readings, fit error 6.4 cm. It is a mixed tide, form factor 1.19, the once-a-day and twice-a-day bands about equal at 0.27 m and 0.24 m. What gives it a clock is that the twice-a-day band is 58 percent solar: the S2 constituent is 0.114 m against 0.066 m for the lunar M2. The moon sets the range and decides which of the day’s two lows is the lower one; the sun sets the hours. The solar lows sit at 02:34 and 14:34 local every day, and 81 percent of all the model’s lows fall between one and four in the morning or one and four in the afternoon, in every month of the year.

So a low tide during the dawn session almost never happens. In the model year from October 2026 to September 2027 a low falls between 06:00 and 09:00 on six days out of 365, all of them in May, June, July and August. From September through April the count is zero. From October to March the dawn window sits on a high or rising tide on most days: 100 percent of October, 90 of November and December, 87 of January, 86 of February, 71 of March. The model has no nodal correction, so its ranges for 2026–27 are probably 15 to 25 percent low; the timing, which is what this essay leans on, is the part a harmonic fit gets right. The tide clock page has the month-by-month histograms.

The rain clock

Rain here is a night and afternoon thing. The window from 06:00 to noon carries about 8 percent of wet-season rain. The rest arrives after the sea breeze, or in the dark. And the signature a barometer might warn you with, the pressure jump and the temperature drop, arrives with the rain, not ahead of it. The station cannot see a storm coming; it can only tell you one is here. The how it rains page shows the hourly distribution.

The sun clock

The UV index at the station is 3 or below before nine in the morning in every month of the year. It reaches 8 or above only from about 11:15 to 14:30. The session ends before the sun does. The heat and UV page has the curves.

Two mornings

The ground-truth log holds the owner’s words by date, with the station and the hindcast beside them. Two September days show the clock working, and one shows the site getting it wrong on paper while the beach got it right.

Tuesday 15 September 2026, logged six days later: “last tuesday and wednesday were great. tuesday the wind held all day and the shape was perfect.” The station read 0.5 knots at dawn. Between 11:00 and 17:00, when the sea breeze should have been in, it averaged 1.1 knots and never exceeded 3.5; the day’s strongest gust was 4.7 knots. The station audit ranks it the calmest afternoon in its record. The hindcast had 0.8 m at 14 seconds from 207°. No number was logged, only the words, and the words are that the clock skipped its afternoon.

Saturday 26 September 2026, logged the same morning: “the waves this morning were really good. clean, good shape”, “like a 7 or 8 out of 10.” Station wind 0.3 to 0.8 knots from six to eight, 1.7 knots at nine, south-west at 5 knots from ten. On the clock to the minute. The hindcast had 1.4 m at 7.85 seconds from 243° with a 0.78 m, 13.2-second train from 207° underneath it; the engine said 3 out of 10, and the log records that as the third owner-rated morning the mixed-train rule has under-scored. The log’s tide field said “dropping-mid to low, 06–10h”. That was a bug. The hourly tide label had been parsing local time as UTC, so it ran six hours early. The model has a low at 03:03, a high at 08:49 and a low at 15:16. The rated window was on a rising-to-high tide, which is what the tide clock says a September dawn almost always is.

Everything a forecast can add sits inside those three hours.

What is left for a forecast

Take the clock as given and see what remains. The wind at dawn is known. The tide phase at dawn is known for the month. The rain is probably not coming until afternoon, and the UV will not bite until the session is over. What the clock cannot say is what the ocean sends: the size, the period, the direction, whether there is one train or two, and whether a storm is close enough to change any of that. That is the forecast, and it lives entirely inside the three hours. A swell model is worth having for exactly that reason.

What the models cannot see, the instrument on the roof can. The sea-breeze turn is a local circulation the coarse grid smooths over; when the site point was corrected in September, the model’s error on temperature and rain fell and its error on gusts rose, from 6.55 to 8.03 knots, because the model does not know how calm a sheltered morning here is. The nocturnal rain is convective and arrives without warning. The tide keeps solar time that the model never looks at. Each of those is a measured baseline now.

So the honest forecast for the point is not a model’s number. It is the measured baseline, plus the model’s deviation from it, stated as a deviation. Dawn offshore, turn at ten twenty, tide rising, rain after noon, and here is what the swell model adds. The site has started building its pages that way: the weather page’s “right now” comes from the station rather than the model cell, and the forecast is scored against the beach, not against itself. Whether a given morning is a 7 or a 3 is still the owner’s call, logged by date, and the engine has been wrong three times this month. The clock has not.

Sources & method

Wind, rain and UV clocks. The station audit of the Playa Saladita weather station (five-minute readings, data/ground_truth/ambient_station_raw/), now being written up as The wind clock, How it rains and Heat and UV: onshore turn median 10:25 (IQR 10:00–10:55) with the monthly medians quoted; first light to onshore 3.3 h; dawn offshore 97–100% of days in most months; two anemometers within 25 minutes; 06:00–12:00 carries ~8% of wet-season rain; storm signature arrives with the rain; UV ≤3 before 09:00, ≥8 only ~11:15–14:30. Those pages may post after this essay.

Tide clock. functions/api/_findings_tide_clock.js, generated 27 September 2026 by scripts/analyze_tide_clock.mjs from the site’s harmonic model (UHSLC Zihuatanejo 687, 2011–13, 33 constituents, RMS 0.064 m, no nodal correction): form factor 1.195 (diurnal band 0.274 m, semidiurnal 0.241 m; S2 0.1137 m, M2 0.066 m, solar share of the semidiurnal band 57.9%), S2 lows 02:34 / 14:34; low at dawn 1.6% of days (6 of 365) in Oct 2026–Sep 2027, 0% Sep–Apr in that window (0% Sep–Mar and 3.3% in Apr in Oct 2025–Sep 2026); high-tide dawn Oct 100%, Nov 90%, Dec 90.3%, Jan 87.1%, Feb 85.7%, Mar 71%. Detail: The tide clock.

The two mornings. data/ground_truth/observations.jsonl, rows dated 2026-09-15 (logged 09-21) and 2026-09-26 (logged 09-26, correction logged 09-27: tide low 03:03, high 08:49, low 15:16; the hourly label ran 6 h early). Station wind for 09-15 from the September raw archive (11:00–17:00 mean 1.12 kt, max 3.5 kt, 72 readings; the “calmest afternoon” ranking is the audit’s); day gust 4.7 kt from forecast_ledger.jsonl. Model gust error 6.55 → 8.03 kt from the 19 September correction commit; see The accidental audit.

What this essay does not claim. No surf outcome beyond the two logged ratings; no April or May wind figure; the tide model’s amplitudes are of medium confidence and only its timing is used here.

Cite as: “Three hours”, Addie Conner, La Saladita Field Guide, lasaladita.com/findings/three-hours/, 27 September 2026.

Continue reading