diff --git a/docs/snippets/weather/reading/absent_values.py b/docs/snippets/weather/reading/absent_values.py
new file mode 100644
index 0000000..7e99901
--- /dev/null
+++ b/docs/snippets/weather/reading/absent_values.py
@@ -0,0 +1,23 @@
+from __future__ import annotations
+
+import math
+
+from idfkit.weather import WeatherFile, monthly_means
+
+epw: WeatherFile = ...  # type: ignore[assignment]
+# --8<-- [start:example]
+# A measurement the file says was not taken is absent, not the number the format
+# reserves for it. Absence is nan: no weather file can express one, so it is not
+# a value the column could otherwise have held.
+albedo = epw.hours.albedo
+print(math.isnan(albedo[0]))  # True where the file wrote its missing value
+
+# How many hours of a column are present is carried rather than recomputed.
+print(epw.hours.present_count["albedo"])
+
+# A monthly mean excludes the absent hours from the sum AND from the divisor, so
+# it is the mean of the values that exist. The count says how many that was.
+january = monthly_means(epw, "dry_bulb_temperature")[0]
+print(january.mean)  # nan when the month held no measurement at all
+print(january.count)  # and 0 beside it, which says why
+# --8<-- [end:example]
diff --git a/docs/snippets/weather/reading/read_an_epw.py b/docs/snippets/weather/reading/read_an_epw.py
new file mode 100644
index 0000000..5747051
--- /dev/null
+++ b/docs/snippets/weather/reading/read_an_epw.py
@@ -0,0 +1,21 @@
+from __future__ import annotations
+
+from idfkit.weather import parse_epw
+
+epw_text: str = ...  # type: ignore[assignment]
+# --8<-- [start:example]
+# Reading takes the text you already hold. No network, no filesystem, nothing
+# asynchronous, so it costs a fraction of the retrieval that produced the text.
+epw = parse_epw(epw_text)
+
+print(epw.location.city)  # Chicago Ohare Intl Ap
+print(epw.hours.row_count)  # 8760, taken from the file's declared data period
+
+# Every numeric field is a named column, packed and the same length.
+temperature = epw.hours.dry_bulb_temperature  # array("d"), degrees C
+print(temperature[0])  # the first hour of 1 January
+
+# The hour column runs 1 to 24, and hour 24 is the last hour of its OWN day
+# rather than hour 0 of the next.
+print(epw.hours.hour[23], epw.hours.day[23])  # 24.0 1.0
+# --8<-- [end:example]
diff --git a/docs/weather/reading.md b/docs/weather/reading.md
new file mode 100644
index 0000000..6feb2fd
--- /dev/null
+++ b/docs/weather/reading.md
@@ -0,0 +1,148 @@
+# How to read a weather file
+
+You have the text of an EPW file, retrieved by [downloading a
+station's files](downloads.md) or read off disk. This guide turns it into the
+header records and the hourly table, and computes monthly figures from it
+without treating an unmeasured hour as a number.
+
+Reading is the same operation in both languages and is spelled the same way. The
+retrieval that precedes it is not: Python hands you a path and TypeScript hands
+you text, for the reasons [How to download weather files](downloads.md) sets
+out. None of that reaches a reader whose input comes from the caller, which is
+why this page has one set of instructions rather than two.
+
+{{ parity("weather-file-reading") }}
+
+!!! info "In JavaScript, weather is a separate install"
+    `pip install idfkit` installs weather support and its station index
+    unconditionally. `npm install idfkit` installs neither. Add `@idfkit/weather`
+    by name. The reader lives on the portable surface of that package, not on
+    `@idfkit/weather/node`: it takes a string and touches no disk, so it runs in
+    a browser tab, a worker or an edge runtime as readily as in Node.
+
+## Read the file
+
+=== "Python"
+
+    ```python
+    --8<-- "docs/snippets/weather/reading/read_an_epw.py:example"
+    ```
+
+=== "TypeScript"
+
+    ```ts
+    --8<-- "docs/snippets/js/weather/reading/read_an_epw.ts:example"
+    ```
+
+`parse_epw` / `parseEpw` takes text. Each language also has the source-taking
+form for a path, `load_epw` and `loadEpw`, which reads the bytes, decodes them
+as the weather formats are written, and adds nothing else. In TypeScript that
+one is in `@idfkit/weather/node`, because it is the half that touches a disk.
+
+## Reach a field by its name, never by its position
+
+An EPW row is thirty-five comma-separated fields with no header line, so a
+caller reading one positionally counts to it. The reader does that once and
+gives every field a name: `dry_bulb_temperature` / `dryBulbTemperature`,
+`relative_humidity` / `relativeHumidity`, `wind_speed` / `windSpeed`, and so on
+for every column of the format.
+
+Two fields are text and stay text. The source-and-uncertainty flags are
+obviously so. The present weather codes are not: they hold a nine-digit value
+such as `999999999`, which is nine single-digit observations written side by
+side rather than a number near a billion. Coercing that yields a column of
+meaningless integers, so it is never parsed.
+
+!!! warning "Hour 24 is the last hour of its own day"
+    The `hour` column runs 1 to 24, not 0 to 23. Hour 24 belongs to the day its
+    row names, not to midnight of the next one. A chart that assumes otherwise is
+    off by one for every day of the year and looks right until somebody checks a
+    single value against the file.
+
+## A measurement that was not taken is not a number
+
+EPW has no blanks. A station that never measured a quantity writes a reserved
+value in the slot, so a file that never recorded its albedo carries `999.000` in
+that column for all 8,760 hours. Averaged naively, that is a plausible-looking
+number that is wrong by three orders of magnitude.
+
+Both libraries map those values to an absent value instead, per field and per
+value, from the reserved-value table published in the EnergyPlus Weather File
+Data Dictionary and committed once for both languages.
+
+=== "Python"
+
+    ```python
+    --8<-- "docs/snippets/weather/reading/absent_values.py:example"
+    ```
+
+=== "TypeScript"
+
+    ```ts
+    --8<-- "docs/snippets/js/weather/reading/absent_values.ts:example"
+    ```
+
+Absence is `nan` in Python and `NaN` in TypeScript. That is chosen rather than
+convenient: no physical quantity in this format can be nan and no EPW can
+express one, so it is not a value the column could otherwise have held. It is
+therefore distinguishable from a real zero, which matters because zero solar
+radiation at night is a measurement, and it propagates loudly through arithmetic
+instead of quietly biasing a result.
+
+!!! warning "A reserved value is not the same as a large number"
+    Ceiling height reserves three values: `99999` for a measurement that was not
+    made, `77777` for an unlimited ceiling, and `88888` for a cirroform ceiling.
+    The last two are observations. Across the sampled corpus that column holds
+    `77777` in 55% of its rows and `99999` in none, so a reader applying the rule
+    "a large number means missing" blanks more than half of it and reports a real
+    sky condition as unmeasured. The table is per field and per value for exactly
+    this reason.
+
+## Compute a monthly figure
+
+`monthly_means` / `monthlyMeans` returns twelve entries for one column, January
+first, each carrying the mean and the count of hours that went into it.
+
+The count is part of the answer rather than something to derive from it. Without
+it you cannot tell a mean over 744 hours from a mean over three, and both look
+like numbers.
+
+The mean excludes absent hours from the sum **and from the divisor**, so it is
+the mean of the values that exist rather than their sum spread over hours that do
+not. A month in which every hour is absent gives an absent mean with a count of
+zero, which is distinguishable from a month whose mean is genuinely zero.
+
+## What the reader does not do
+
+- **Write an EPW.** Reading only.
+- **Interpret the design conditions record.** It is kept as text. The DDY member
+  of the same archive is the supported path to design conditions; see [How to
+  apply design days](design-days.md).
+- **Read the climate summary.** The `.stat` member of an archive is a report
+  written for a person, in around forty section shapes whose set varies with the
+  station's climate. Parsing it would be brittle in a way this reader is not, and
+  every figure it carries is derivable from the EPW itself. It is used as an
+  oracle in the shared conformance corpus, where a brittle parse is somebody's
+  offline maintenance task rather than something in a caller's path, and that is
+  the only place it is read.
+- **Interpolate, gap-fill or repair.** An absent value is reported as absent.
+  What to do about it depends on what you are computing, which makes it your
+  judgement rather than the reader's.
+- **Guess at what it cannot represent.** A file whose rows disagree with its own
+  declared data period fails, naming the disagreement. So does a row with the
+  wrong field count, naming the row. Nothing partial is ever returned.
+
+## How the two languages are held to the same answer
+
+Both readers are checked against the climate summary the EnergyPlus Weather
+Converter produced from the same archive, in the shared conformance corpus rather
+than in either library's own tests. That summary is not written by either library
+and ships in the same archive as the file it describes, so the expectation and the
+input cannot drift apart.
+
+The check compares monthly means of dry bulb, dew point, relative humidity and
+wind speed over six stations spanning the climate range, and separately asserts
+that the reserved values in a seventh file read as absent. Beyond the four fields
+the summary aggregates, the two libraries are held only to agreeing with each
+other, which is weaker; the corpus says so rather than implying the whole table
+is checked against something external.
diff --git a/mkdocs.yml b/mkdocs.yml
index ab39059..2c0c93f 100644
--- a/mkdocs.yml
+++ b/mkdocs.yml
@@ -78,6 +78,7 @@ nav:
       - weather/index.md
       - Search for weather stations: weather/station-search.md
       - Download weather files: weather/downloads.md
+      - Read a weather file: weather/reading.md
       - Apply design days: weather/design-days.md
       - Geocode addresses: weather/geocoding.md
       - Warm the weather cache for an offline run: how-to/warm-the-weather-cache.md