diff --git a/db/client.ts b/db/client.ts
new file mode 100644
index 000000000..d0446b347
--- /dev/null
+++ b/db/client.ts
@@ -0,0 +1,13 @@
+import { Kysely, PostgresDialect } from 'kysely';
+import { Pool, PoolConfig, types } from 'pg';
+import type { Database } from './types';
+
+// pg turns a date into a Date at local midnight, so 2026-07-01 is 04:00Z on a
+// laptop in Eastern time and 00:00Z in Lambda. Keep dates as 'YYYY-MM-DD'.
+types.setTypeParser(types.builtins.DATE, (value) => value);
+
+export function createDb(config: PoolConfig): Kysely<Database> {
+  return new Kysely<Database>({
+    dialect: new PostgresDialect({ pool: new Pool(config) }),
+  });
+}
diff --git a/db/migrations/0001_init_schema.ts b/db/migrations/0001_init_schema.ts
new file mode 100644
index 000000000..02a566a54
--- /dev/null
+++ b/db/migrations/0001_init_schema.ts
@@ -0,0 +1,246 @@
+import { Kysely, sql } from 'kysely';
+
+// Tables, columns, and sources are documented in docs/database-schema.md.
+export async function up(db: Kysely<any>): Promise<void> {
+  await db.schema
+    .createTable('users')
+    .addColumn('id', 'serial', (col) => col.primaryKey())
+    .addColumn('email', 'text', (col) => col.notNull().unique())
+    .addColumn('firstName', 'text', (col) => col.notNull())
+    .addColumn('lastName', 'text', (col) => col.notNull())
+    .addColumn('status', 'text', (col) => col.notNull())
+    .execute();
+
+  await db.schema
+    .createTable('stations')
+    .addColumn('id', 'serial', (col) => col.primaryKey())
+    .addColumn('code', 'text', (col) => col.notNull().unique())
+    .execute();
+
+  await db.insertInto('stations').values({ code: '19737-02' }).execute();
+
+  await db.schema
+    .createTable('imports')
+    .addColumn('id', 'serial', (col) => col.primaryKey())
+    .addColumn('kind', 'text', (col) => col.notNull())
+    .addColumn('fileName', 'text', (col) => col.notNull())
+    .addColumn('s3Key', 'text', (col) => col.notNull())
+    .addColumn('fileHash', 'text', (col) => col.notNull())
+    .addColumn('uploadedBy', 'integer', (col) =>
+      col.notNull().references('users.id'),
+    )
+    .addColumn('uploadedAt', 'timestamptz', (col) =>
+      col.notNull().defaultTo(sql`now()`),
+    )
+    .addColumn('status', 'text', (col) => col.notNull())
+    .addColumn('errors', 'jsonb')
+    .execute();
+
+  await db.schema
+    .createTable('daily_observations')
+    .addColumn('stationId', 'integer', (col) =>
+      col.notNull().references('stations.id'),
+    )
+    .addColumn('date', 'date', (col) => col.notNull())
+    .addColumn('maxTempF', 'smallint')
+    .addColumn('minTempF', 'smallint')
+    .addColumn('avgTempF', 'numeric')
+    .addColumn('normalTempF', 'smallint')
+    .addColumn('precipIn', 'numeric')
+    .addColumn('precipTrace', 'boolean', (col) => col.notNull())
+    .addColumn('snowfallIn', 'numeric')
+    .addColumn('snowDepthIn', 'numeric')
+    .addColumn('sunshineMin', 'smallint')
+    .addColumn('sunshinePossibleMin', 'smallint')
+    .addColumn('sunshinePct', 'smallint')
+    .addColumn('fastestMileMph', 'smallint')
+    .addColumn('fastestMileDir', 'varchar(3)')
+    .addColumn('fastestMileTime', 'time')
+    .addColumn('peakGustKts', 'smallint')
+    .addColumn('peakGustDir', 'varchar(3)')
+    .addColumn('peakGustTime', 'time')
+    .addColumn('avgStationPressureMb', 'numeric')
+    .addColumn('sunrise', 'time')
+    .addColumn('sunset', 'time')
+    .addColumn('observerInitials', 'text')
+    .addColumn('remarks', 'text')
+    .addColumn('importId', 'integer', (col) =>
+      col.notNull().references('imports.id'),
+    )
+    .addPrimaryKeyConstraint('daily_observations_pkey', ['stationId', 'date'])
+    .execute();
+
+  await db.schema
+    .createTable('hourly_observations')
+    .addColumn('stationId', 'integer', (col) => col.notNull())
+    .addColumn('date', 'date', (col) => col.notNull())
+    .addColumn('hour', 'smallint', (col) => col.notNull())
+    .addColumn('tempF', 'smallint')
+    .addColumn('precipIn', 'numeric')
+    .addColumn('precipTrace', 'boolean', (col) => col.notNull())
+    .addColumn('windDir', 'varchar(3)')
+    .addColumn('windSpeedMph', 'smallint')
+    .addColumn('sunshineMin', 'smallint')
+    .addColumn('skyCover', 'smallint')
+    .addColumn('visibilityMi', 'numeric')
+    .addColumn('presentWeather', 'text')
+    .addColumn('humidityPct', 'smallint')
+    .addColumn('mountainVis', 'text')
+    .addColumn('remarks', 'text')
+    .addPrimaryKeyConstraint('hourly_observations_pkey', [
+      'stationId',
+      'date',
+      'hour',
+    ])
+    .addForeignKeyConstraint(
+      'hourly_observations_day_fkey',
+      ['stationId', 'date'],
+      'daily_observations',
+      ['stationId', 'date'],
+      (fk) => fk.onDelete('cascade'),
+    )
+    .execute();
+
+  await db.schema
+    .createTable('scheduled_observations')
+    .addColumn('stationId', 'integer', (col) => col.notNull())
+    .addColumn('date', 'date', (col) => col.notNull())
+    .addColumn('obsTime', 'time', (col) => col.notNull())
+    .addColumn('stationPressureIn', 'numeric')
+    .addColumn('attachedThermC', 'numeric')
+    .addColumn('observedBarometerMb', 'numeric')
+    .addColumn('pressureCorrectionMb', 'numeric')
+    .addColumn('dryBulbF', 'numeric')
+    .addColumn('wetBulbF', 'numeric')
+    .addColumn('dewpointF', 'smallint')
+    .addColumn('humidityPct', 'smallint')
+    .addColumn('maxTempF', 'smallint')
+    .addColumn('minTempF', 'smallint')
+    .addColumn('precipIn', 'numeric')
+    .addColumn('precipTrace', 'boolean', (col) => col.notNull())
+    .addColumn('snowfallIn', 'numeric')
+    .addColumn('snowDepthIn', 'numeric')
+    .addColumn('vaporPressureMb', 'numeric')
+    .addPrimaryKeyConstraint('scheduled_observations_pkey', [
+      'stationId',
+      'date',
+      'obsTime',
+    ])
+    .addForeignKeyConstraint(
+      'scheduled_observations_day_fkey',
+      ['stationId', 'date'],
+      'daily_observations',
+      ['stationId', 'date'],
+      (fk) => fk.onDelete('cascade'),
+    )
+    .execute();
+
+  await db.schema
+    .createTable('daily_records')
+    .addColumn('stationId', 'integer', (col) =>
+      col.notNull().references('stations.id'),
+    )
+    .addColumn('month', 'smallint', (col) => col.notNull())
+    .addColumn('day', 'smallint', (col) => col.notNull())
+    .addColumn('highF', 'smallint')
+    .addColumn('highYears', sql`smallint[]`)
+    .addColumn('lowF', 'smallint')
+    .addColumn('lowYears', sql`smallint[]`)
+    .addColumn('precipIn', 'numeric')
+    .addColumn('precipYears', sql`smallint[]`)
+    .addColumn('snowIn', 'numeric')
+    .addColumn('snowYears', sql`smallint[]`)
+    .addColumn('peakGustMph', 'smallint')
+    .addColumn('peakGustDir', 'varchar(3)')
+    .addColumn('peakGustIsEstimated', 'boolean', (col) => col.notNull())
+    .addColumn('peakGustYears', sql`smallint[]`)
+    .addColumn('notes', 'text')
+    .addColumn('importId', 'integer', (col) =>
+      col.notNull().references('imports.id'),
+    )
+    .addPrimaryKeyConstraint('daily_records_pkey', [
+      'stationId',
+      'month',
+      'day',
+    ])
+    .execute();
+
+  await db.schema
+    .createType('audit_log_action')
+    .asEnum(['insert', 'update', 'delete'])
+    .execute();
+
+  await db.schema
+    .createTable('audit_log')
+    .addColumn('id', 'serial', (col) => col.primaryKey())
+    .addColumn('tableName', 'text', (col) => col.notNull())
+    .addColumn('rowKey', 'jsonb', (col) => col.notNull())
+    .addColumn('action', sql`audit_log_action`, (col) => col.notNull())
+    .addColumn('oldData', 'jsonb')
+    .addColumn('newData', 'jsonb')
+    .addColumn('changedBy', 'integer', (col) => col.references('users.id'))
+    .addColumn('importId', 'integer', (col) => col.references('imports.id'))
+    .addColumn('changedAt', 'timestamptz', (col) =>
+      col.notNull().defaultTo(sql`now()`),
+    )
+    .execute();
+
+  // Trigger arguments are the table's primary key columns, used for rowKey.
+  // Writers identify themselves with set_config('app.user_id' / 'app.import_id',
+  // ..., true) inside their transaction; hourly and scheduled rows have no
+  // importId column, so the row itself can't say which import wrote it.
+  await sql`
+    CREATE FUNCTION audit_row() RETURNS trigger AS $$
+    DECLARE
+      rec jsonb := to_jsonb(CASE WHEN TG_OP = 'DELETE' THEN OLD ELSE NEW END);
+      key jsonb := '{}';
+      col text;
+    BEGIN
+      FOREACH col IN ARRAY TG_ARGV LOOP
+        key := key || jsonb_build_object(col, rec -> col);
+      END LOOP;
+
+      INSERT INTO audit_log ("tableName", "rowKey", action, "oldData", "newData", "changedBy", "importId")
+      VALUES (
+        TG_TABLE_NAME,
+        key,
+        lower(TG_OP)::audit_log_action,
+        CASE WHEN TG_OP <> 'INSERT' THEN to_jsonb(OLD) END,
+        CASE WHEN TG_OP <> 'DELETE' THEN to_jsonb(NEW) END,
+        nullif(current_setting('app.user_id', true), '')::int,
+        nullif(current_setting('app.import_id', true), '')::int
+      );
+      RETURN NULL;
+    END;
+    $$ LANGUAGE plpgsql
+  `.execute(db);
+
+  const audited: [string, string[]][] = [
+    ['daily_observations', ['stationId', 'date']],
+    ['hourly_observations', ['stationId', 'date', 'hour']],
+    ['scheduled_observations', ['stationId', 'date', 'obsTime']],
+    ['daily_records', ['stationId', 'month', 'day']],
+  ];
+  for (const [table, key] of audited) {
+    await sql`
+      CREATE TRIGGER ${sql.id(`${table}_audit`)}
+      AFTER INSERT OR UPDATE OR DELETE ON ${sql.table(table)}
+      FOR EACH ROW EXECUTE FUNCTION audit_row(${sql.raw(
+        key.map((col) => `'${col}'`).join(', '),
+      )})
+    `.execute(db);
+  }
+}
+
+export async function down(db: Kysely<any>): Promise<void> {
+  await db.schema.dropTable('audit_log').execute();
+  await db.schema.dropType('audit_log_action').execute();
+  await db.schema.dropTable('daily_records').execute();
+  await db.schema.dropTable('scheduled_observations').execute();
+  await db.schema.dropTable('hourly_observations').execute();
+  await db.schema.dropTable('daily_observations').execute();
+  await sql`DROP FUNCTION audit_row()`.execute(db);
+  await db.schema.dropTable('imports').execute();
+  await db.schema.dropTable('stations').execute();
+  await db.schema.dropTable('users').execute();
+}
diff --git a/db/types.ts b/db/types.ts
new file mode 100644
index 000000000..a550120a0
--- /dev/null
+++ b/db/types.ts
@@ -0,0 +1,143 @@
+import type { ColumnType, Generated } from 'kysely';
+
+// Table types for Kysely, matching db/migrations. pg returns numeric as a
+// string so no precision is lost, timestamptz as a Date, and date as
+// 'YYYY-MM-DD' when connected through createDb() in db/client.ts.
+type Numeric = ColumnType<string, number | string, number | string>;
+type Json = ColumnType<unknown, string, string>;
+
+export interface UsersTable {
+  id: Generated<number>;
+  email: string;
+  firstName: string;
+  lastName: string;
+  status: string;
+}
+
+export interface StationsTable {
+  id: Generated<number>;
+  code: string;
+}
+
+export interface ImportsTable {
+  id: Generated<number>;
+  kind: string;
+  fileName: string;
+  s3Key: string;
+  fileHash: string;
+  uploadedBy: number;
+  uploadedAt: ColumnType<Date, Date | string | undefined, Date | string>;
+  status: string;
+  errors: Json | null;
+}
+
+export interface DailyObservationsTable {
+  stationId: number;
+  date: string;
+  maxTempF: number | null;
+  minTempF: number | null;
+  avgTempF: Numeric | null;
+  normalTempF: number | null;
+  precipIn: Numeric | null;
+  precipTrace: boolean;
+  snowfallIn: Numeric | null;
+  snowDepthIn: Numeric | null;
+  sunshineMin: number | null;
+  sunshinePossibleMin: number | null;
+  sunshinePct: number | null;
+  fastestMileMph: number | null;
+  fastestMileDir: string | null;
+  fastestMileTime: string | null;
+  peakGustKts: number | null;
+  peakGustDir: string | null;
+  peakGustTime: string | null;
+  avgStationPressureMb: Numeric | null;
+  sunrise: string | null;
+  sunset: string | null;
+  observerInitials: string | null;
+  remarks: string | null;
+  importId: number;
+}
+
+export interface HourlyObservationsTable {
+  stationId: number;
+  date: string;
+  hour: number;
+  tempF: number | null;
+  precipIn: Numeric | null;
+  precipTrace: boolean;
+  windDir: string | null;
+  windSpeedMph: number | null;
+  sunshineMin: number | null;
+  skyCover: number | null;
+  visibilityMi: Numeric | null;
+  presentWeather: string | null;
+  humidityPct: number | null;
+  mountainVis: string | null;
+  remarks: string | null;
+}
+
+export interface ScheduledObservationsTable {
+  stationId: number;
+  date: string;
+  obsTime: string;
+  stationPressureIn: Numeric | null;
+  attachedThermC: Numeric | null;
+  observedBarometerMb: Numeric | null;
+  pressureCorrectionMb: Numeric | null;
+  dryBulbF: Numeric | null;
+  wetBulbF: Numeric | null;
+  dewpointF: number | null;
+  humidityPct: number | null;
+  maxTempF: number | null;
+  minTempF: number | null;
+  precipIn: Numeric | null;
+  precipTrace: boolean;
+  snowfallIn: Numeric | null;
+  snowDepthIn: Numeric | null;
+  vaporPressureMb: Numeric | null;
+}
+
+export interface DailyRecordsTable {
+  stationId: number;
+  month: number;
+  day: number;
+  highF: number | null;
+  highYears: number[] | null;
+  lowF: number | null;
+  lowYears: number[] | null;
+  precipIn: Numeric | null;
+  precipYears: number[] | null;
+  snowIn: Numeric | null;
+  snowYears: number[] | null;
+  peakGustMph: number | null;
+  peakGustDir: string | null;
+  peakGustIsEstimated: boolean;
+  peakGustYears: number[] | null;
+  notes: string | null;
+  importId: number;
+}
+
+// Written only by the audit_row() trigger.
+export interface AuditLogTable {
+  id: Generated<number>;
+  tableName: string;
+  rowKey: unknown;
+  action: 'insert' | 'update' | 'delete';
+  oldData: unknown;
+  newData: unknown;
+  changedBy: number | null;
+  importId: number | null;
+  changedAt: Generated<Date>;
+}
+
+export interface Database {
+  users: UsersTable;
+  stations: StationsTable;
+  imports: ImportsTable;
+  daily_observations: DailyObservationsTable;
+  hourly_observations: HourlyObservationsTable;
+  scheduled_observations: ScheduledObservationsTable;
+  daily_records: DailyRecordsTable;
+  audit_log: AuditLogTable;
+}
diff --git a/docs/database-schema.md b/docs/database-schema.md
new file mode 100644
index 000000000..e94144be9
--- /dev/null
+++ b/docs/database-schema.md
@@ -0,0 +1,196 @@
+# Database schema
+
+Draft for [#3](https://github.com/Code-4-Community/bho/issues/3). PostgreSQL on AWS RDS. Created by `db/migrations/0001_init_schema.ts`; Kysely table types are in `db/types.ts`. Connect with `createDb()` in `db/client.ts`, which returns `date` columns as `'YYYY-MM-DD'` strings rather than a `Date` at local midnight.
+
+Sources:
+
+- **Daily sheet**: [DAILY - 07. July2026.xlsx](https://docs.google.com/spreadsheets/d/1cWloMpYHZDGkUFulmeW0g1bY3J97Ts4b/edit), tab `Sheet1`. One 60-row block per day. Cell references are for the July 1 block; add 60 rows for each later day.
+- **Historical sheet**: [BHO_WeatherRecords_Master_TestData.xlsx](https://docs.google.com/spreadsheets/d/1WI8CxTJ7edRajo3JHZX0WmTklvCqRy5x/edit), tab `All Daily Records`, rows 4-369.
+
+Column names are camelCase, so Kysely queries use the same names as the TypeScript types. Raw SQL has to quote them (`"maxTempF"`).
+
+## Relations
+
+```mermaid
+erDiagram
+  users ||--o{ imports : uploads
+  users ||--o{ audit_log : changes
+  imports ||--o{ audit_log : produces
+  imports ||--o{ daily_observations : creates
+  imports ||--o{ daily_records : creates
+  stations ||--o{ daily_observations : records
+  stations ||--o{ daily_records : holds
+  daily_observations ||--|{ hourly_observations : has
+  daily_observations ||--|{ scheduled_observations : has
+```
+
+`daily_records` joins to observations on station, month, and day, not a foreign key.
+
+## users
+
+Source: the app. `main` no longer has the scaffold's `User` entity, so this schema creates the table, with the same columns.
+
+| Column    | Type   | Null | Note                                     |
+| --------- | ------ | ---- | ---------------------------------------- |
+| id        | serial | N    | PK                                       |
+| email     | text   | N    | Unique                                   |
+| firstName | text   | N    |                                          |
+| lastName  | text   | N    |                                          |
+| status    | text   | N    | `ADMIN` or `STANDARD`; text, not an enum |
+
+## stations
+
+Source: the app. Starts with one row, `19737-02`; the client plans to add more. A table rather than an enum, so a new station is a new row, not a migration.
+
+| Column | Type   | Null | Note               |
+| ------ | ------ | ---- | ------------------ |
+| id     | serial | N    | PK                 |
+| code   | text   | N    | Unique, `19737-02` |
+
+## imports
+
+Source: the app, one row per uploaded spreadsheet. The uploaded file stays in S3 as the raw source of truth. Rows are never updated or deleted. Each upload, including a re-upload of the same file, adds a new row.
+
+| Column     | Type        | Null | Note                                            |
+| ---------- | ----------- | ---- | ----------------------------------------------- |
+| id         | serial      | N    | PK                                              |
+| kind       | text        | N    | `daily` or `historical`                         |
+| fileName   | text        | N    |                                                 |
+| s3Key      | text        | N    | Key of the uploaded file in S3                  |
+| fileHash   | text        | N    | SHA-256 of the file, to catch duplicate uploads |
+| uploadedBy | int         | N    | FK users                                        |
+| uploadedAt | timestamptz | N    | Defaults to now                                 |
+| status     | text        | N    | `succeeded` or `failed`; text, not an enum      |
+| errors     | jsonb       | Y    | Flagged rows and the reason for each            |
+
+Ideas for later:
+
+- `rowCount` (rows imported): not sure what the existing parser looks like
+- `partial`? Do we want to allow for this, feels simpler to force a re-import and log that import failed
+
+## daily_observations
+
+Source: daily sheet, "Summary of Day" and nearby rows. PK `(stationId, date)`.
+
+| Column               | Type       | Null | Source                                      |
+| -------------------- | ---------- | ---- | ------------------------------------------- |
+| stationId            | int        | N    | FK stations; the station the upload is for  |
+| date                 | date       | N    | J1 "Wednesday, July 1, 2026"                |
+| maxTempF             | smallint   | Y    | A44 "24HR Max"                              |
+| minTempF             | smallint   | Y    | B44 "24HR Min"                              |
+| avgTempF             | numeric    | Y    | D46 "Average Temperature"; 81.5             |
+| normalTempF          | smallint   | Y    | D47 "Normal Temperature"                    |
+| precipIn             | numeric    | Y    | C44 "24HR Precip"                           |
+| precipTrace          | boolean    | N    | Our flag for when C44 is `T` (July 5: C284) |
+| snowfallIn           | numeric    | Y    | E44 "Snowfall"                              |
+| snowDepthIn          | numeric    | Y    | F44 "Dpth 0700"                             |
+| sunshineMin          | smallint   | Y    | G44 "Total Min."                            |
+| sunshinePossibleMin  | smallint   | Y    | F29 "Poss 914" (possible sunshine minutes)  |
+| sunshinePct          | smallint   | Y    | H44 "% poss." (0.77 becomes 77)             |
+| fastestMileMph       | smallint   | Y    | I44                                         |
+| fastestMileDir       | varchar(3) | Y    | J44                                         |
+| fastestMileTime      | time       | Y    | K44 "0245E" (24-hour, EST)                  |
+| peakGustKts          | smallint   | Y    | L44 "20 WSW @ 0854E" (knots)                |
+| peakGustDir          | varchar(3) | Y    | L44                                         |
+| peakGustTime         | time       | Y    | L44                                         |
+| avgStationPressureMb | numeric    | Y    | E56 "Station Pressure" AVG                  |
+| sunrise              | time       | Y    | K29                                         |
+| sunset               | time       | Y    | K30                                         |
+| observerInitials     | text       | Y    | H59 "MFD/AKJ"                               |
+| remarks              | text       | Y    | G47-G49 "Additional Notes"                  |
+| importId             | int        | N    | FK imports; the last upload that wrote it   |
+
+Every row comes from an upload. Staff re-upload the month's sheet several times a day, and the most recent upload overwrites each day it contains. Weather values are nullable: an empty cell means no data and is stored as null, never 0. The same applies to `daily_records`.
+
+Not stored: departure and degree days (rows 48-50). They can be computed.
+
+## hourly_observations
+
+Source: daily sheet, rows 4-27 ("00-01" to "23-24"). PK `(stationId, date, hour)`. Deleting a day deletes its hours.
+
+| Column         | Type       | Null | Source                                           |
+| -------------- | ---------- | ---- | ------------------------------------------------ |
+| stationId      | int        | N    | FK daily_observations, with `date`               |
+| date           | date       | N    | FK daily_observations, with `stationId`          |
+| hour           | smallint   | N    | A, start hour ("00-01" is 0)                     |
+| tempF          | smallint   | Y    | B "Temp (°F)"                                    |
+| precipIn       | numeric    | Y    | C "Precip"                                       |
+| precipTrace    | boolean    | N    | Our flag for when C is `T`                       |
+| windDir        | varchar(3) | Y    | D "Dir."                                         |
+| windSpeedMph   | smallint   | Y    | E "Spd (mph)"                                    |
+| sunshineMin    | smallint   | Y    | F "Sunshine (min)"                               |
+| skyCover       | smallint   | Y    | G "Sky Cover (0-8)"                              |
+| visibilityMi   | numeric    | Y    | H "Lwst VSBL (mls)"; "1/8" mile stored as 0.125  |
+| presentWeather | text       | Y    | I "Prsnt WX"; codes, "RW-" is light rain showers |
+| humidityPct    | smallint   | Y    | J "Rel Hum."                                     |
+| mountainVis    | text       | Y    | K "Mnts VSBL" ("1@1.5")                          |
+| remarks        | text       | Y    | L "Notes"                                        |
+
+G, H, and I are only recorded at a few hours each day (06-07, 09-10, and 12-13 on July 21). An empty cell there means nobody observed that hour; `0` in I means they observed and there was no weather. Codes in I: `R` rain, `W` showers, `L` drizzle, `F` fog, `-` light, `+` heavy, so `R-F` (I370, July 7) is light rain and fog. These are the standard US codes; BHO has not confirmed them.
+
+## scheduled_observations
+
+Source: daily sheet, "Scheduled Observations", rows 34-38. PK `(stationId, date, obsTime)`. Deleting a day deletes its readings.
+
+| Column               | Type     | Null | Source                                                |
+| -------------------- | -------- | ---- | ----------------------------------------------------- |
+| stationId            | int      | N    | FK daily_observations, with `date`                    |
+| date                 | date     | N    | FK daily_observations, with `stationId`               |
+| obsTime              | time     | N    | A (700, 800, 1000, 1300, 1900); F52 1200 on some days |
+| stationPressureIn    | numeric  | Y    | B, 0700 and 1900 only                                 |
+| attachedThermC       | numeric  | Y    | C53 "Attached Therm."; F53 at 1200                    |
+| observedBarometerMb  | numeric  | Y    | C54 "Observed Barometer"; F54 at 1200                 |
+| pressureCorrectionMb | numeric  | Y    | C55 "Total Correction"; F55 at 1200                   |
+| dryBulbF             | numeric  | Y    | C                                                     |
+| wetBulbF             | numeric  | Y    | D                                                     |
+| dewpointF            | smallint | Y    | E                                                     |
+| humidityPct          | smallint | Y    | F (0.8 becomes 80)                                    |
+| maxTempF             | smallint | Y    | G, 0700 only, 24 hours ending 0700                    |
+| minTempF             | smallint | Y    | H, 0700 only                                          |
+| precipIn             | numeric  | Y    | I, 0700 only                                          |
+| precipTrace          | boolean  | N    | Our flag for when I is `T`                            |
+| snowfallIn           | numeric  | Y    | J, 0700 only                                          |
+| snowDepthIn          | numeric  | Y    | K, 0700 only                                          |
+| vaporPressureMb      | numeric  | Y    | L                                                     |
+
+## daily_records
+
+Source: historical sheet. One sheet row becomes one row. PK `(stationId, month, day)`. Each element's value is null when the sheet says "None", and its years are null too.
+
+| Column              | Type       | Null | Source                                             |
+| ------------------- | ---------- | ---- | -------------------------------------------------- |
+| stationId           | int        | N    | FK stations; the station the upload is for         |
+| month               | smallint   | N    | "Month #"                                          |
+| day                 | smallint   | N    | "Day"                                              |
+| highF               | smallint   | Y    | "High °F"                                          |
+| highYears           | smallint[] | Y    | "High Year(s)"; "2014, 1977" becomes `{2014,1977}` |
+| lowF                | smallint   | Y    | "Low °F"                                           |
+| lowYears            | smallint[] | Y    | "Low Year(s)"                                      |
+| precipIn            | numeric    | Y    | "Precip in"                                        |
+| precipYears         | smallint[] | Y    | "Precip Year(s)"                                   |
+| snowIn              | numeric    | Y    | "Snow in"; null for "None"                         |
+| snowYears           | smallint[] | Y    | "Snow Year(s)"; null when snow is "None"           |
+| peakGustMph         | smallint   | Y    | "Peak Gust mph"                                    |
+| peakGustDir         | varchar(3) | Y    | "Wind Dir"                                         |
+| peakGustIsEstimated | boolean    | N    | "Estimated?"                                       |
+| peakGustYears       | smallint[] | Y    | "Gust Year(s)"                                     |
+| notes               | text       | Y    | "Notes"; internal to staff                         |
+| importId            | int        | N    | FK imports; the last upload that wrote it          |
+
+## audit_log
+
+Source: a Postgres trigger, one row per inserted, updated, or deleted row in `daily_observations`, `hourly_observations`, `scheduled_observations`, and `daily_records`.
+
+| Column    | Type        | Null | Note                                           |
+| --------- | ----------- | ---- | ---------------------------------------------- |
+| id        | serial      | N    | PK                                             |
+| tableName | text        | N    |                                                |
+| rowKey    | jsonb       | N    | PK of the affected row                         |
+| action    | enum        | N    | `insert`, `update`, or `delete`                |
+| oldData   | jsonb       | Y    | Whole row before; null on insert               |
+| newData   | jsonb       | Y    | Whole row after; null on delete                |
+| changedBy | int         | Y    | FK users; null for system writes               |
+| importId  | int         | Y    | FK imports; set when an import made the change |
+| changedAt | timestamptz | N    | Defaults to now                                |
+
+One SQL function is attached to the four data tables, so no code path can skip the log. Writers set `app.user_id` and `app.import_id` with `set_config(..., true)` inside their transaction, and the trigger copies them into `changedBy` and `importId`. Hourly and scheduled rows have no `importId` column, so the row alone can't say which import wrote it.
diff --git a/package.json b/package.json
index c9dd38d2d..e30b19f63 100644
--- a/package.json
+++ b/package.json
@@ -40,6 +40,7 @@
     "global": "^4.4.0",
     "jsonwebtoken": "^9.0.3",
     "jwks-rsa": "^4.1.0",
+    "kysely": "^0.27.0",
     "listr2": "11",
     "nodemailer": "^9.0.3",
     "passport": "^0.7.0",
diff --git a/yarn.lock b/yarn.lock
index d7ac9b68a..4a1687870 100644
--- a/yarn.lock
+++ b/yarn.lock
@@ -9392,6 +9392,11 @@ keyv@^4.5.3:
   dependencies:
     json-buffer "3.0.1"
 
+kysely@^0.27.0:
+  version "0.27.6"
+  resolved "https://registry.yarnpkg.com/kysely/-/kysely-0.27.6.tgz#b2a908e160e510f9945586ca9de1616c187a2a8e"
+  integrity sha512-FIyV/64EkKhJmjgC0g2hygpBv5RNWVPyNCqSAD7eTCv6eFWNIi4PN1UvdSJGicN/o35bnevgis4Y0UDC0qi8jQ==
+
 language-subtag-registry@^0.3.20:
   version "0.3.23"
   resolved "https://registry.yarnpkg.com/language-subtag-registry/-/language-subtag-registry-0.3.23.tgz#23529e04d9e3b74679d70142df3fd2eb6ec572e7"