diff --git a/CHANGELOG.md b/CHANGELOG.md
index 19746b41..5ed53a54 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,8 +4,44 @@ All notable changes to `mostlyright`. The format follows [Keep a Changelog](http
 
 ## [Unreleased]
 
+## [1.12.0] — 2026-07-08 — research(source=) pin + Mode 2 window fix
+
+Minor release, dual-SDK (PyPI 1.12.0 + npm 1.12.0 in lockstep).
+
+### Added
+
+- **`research(source=...)` single-source pin (both SDKs).** The singular
+  `source=` kwarg is now functional (was `NotImplementedError` since Phase 10):
+  observation rows are filtered to the requested source's alias-accept set
+  after the multi-source merge and before daily `obs_*` aggregation. Accepts
+  the full Mode 2 vocabulary (`iem`, `iem.archive`, `iem.live`, `awc`,
+  `awc.live`, `ghcnh`, `ghcnh.archive`) — byte-identical alias tables across
+  Python (`mostlyright.mode2._SOURCE_ALIASES`) and TS
+  (`RESEARCH_SOURCE_ALIASES`). Unknown sources raise before any network
+  fetch. `cli_*` settlement columns are unaffected (NWS CLI is independent
+  of the observation pin); the pinned DataFrame carries
+  `df.attrs["source"]`. Inherited v0.1 limitation documented: the
+  AWC > IEM > GHCNh merge runs before the pin filter, so pinning a
+  lower-priority source yields only rows the merge did not supersede.
+  `sources=[...]` (plural, multi-source subset) remains v0.3; its error
+  message no longer claims `source=` is unwired. The `source=None` default
+  path is byte-identical to 1.11.0 (parity gate green). TS `source: null`
+  follows the Phase 21 wire contract (null-as-absent). Follow-up nits: #107.
+
 ### Fixed
 
+- **`mode2.research_by_source` returned whole calendar months instead of the
+  requested window.** The month-granular parquet cache flows through
+  `_fetch_observations_range` unfiltered, so a `KNYC` Jan 6–12 query
+  returned all 865 January rows (a Jan 28–Feb 3 query returned two full
+  months) — silent out-of-window leakage into backtest frames, including
+  the README's own `train = research_by_source(...)` example. Rows are now
+  trimmed to `[from_date, to_date]` by UTC date-part, inclusive, matching
+  the TS port's shipped filter exactly (cross-SDK lockstep); the +1-day
+  fetch extension remains as a fetch-width knob only. Every return shape
+  is fixed (DataFrame, `as_dataframe=False`, wrapper/backends). Callers
+  needing the last LST day's pre-midnight UTC tail should extend
+  `to_date` by one day.
 - **Earnings venue-rule compound-word under-count (D-30).** Hyphenated compounds
   (e.g. `supply-chain` for `chain`) were dropped from BOTH venue counts by an
   inverted guard; they now count on both venues per each issuer's actual rule.
diff --git a/package.json b/package.json
index ff769b68..053c8beb 100644
--- a/package.json
+++ b/package.json
@@ -37,7 +37,7 @@
     {
       "name": "@mostlyrightmd/weather (W1: AWC+CLI only, gzipped ESM)",
       "path": "packages-ts/weather/dist/index.mjs",
-      "limit": "25 KB"
+      "limit": "26 KB"
     },
     {
       "name": "@mostlyrightmd/markets (W1: Kalshi NHIGH/NLOW, gzipped ESM)",
diff --git a/packages-ts/core/package.json b/packages-ts/core/package.json
index 16e1b5e2..df013116 100644
--- a/packages-ts/core/package.json
+++ b/packages-ts/core/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@mostlyrightmd/core",
-  "version": "1.11.0",
+  "version": "1.12.0",
   "description": "TypeScript SDK core for quants, ML pipelines, and AI agents: types, schemas, validators, temporal-safety primitives, and the research() join over weather data + prediction-market settlements. Local-first, no hosted backend.",
   "keywords": [
     "weather",
diff --git a/packages-ts/markets/package.json b/packages-ts/markets/package.json
index 3c229576..dbe5561e 100644
--- a/packages-ts/markets/package.json
+++ b/packages-ts/markets/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@mostlyrightmd/markets",
-  "version": "1.11.0",
+  "version": "1.12.0",
   "description": "Prediction-market data for TypeScript / Node — Kalshi NHIGH/NLOW weather-contract resolvers, Polymarket discovery + settlement, and Kalshi + Polymarket trade history. For quants, backtesting, and ML training pipelines.",
   "keywords": [
     "kalshi",
diff --git a/packages-ts/meta/package.json b/packages-ts/meta/package.json
index 5058a748..0ae058d3 100644
--- a/packages-ts/meta/package.json
+++ b/packages-ts/meta/package.json
@@ -1,6 +1,6 @@
 {
   "name": "mostlyright",
-  "version": "1.11.0",
+  "version": "1.12.0",
   "description": "Public-data SDK for TypeScript — one import for quants, ML pipelines, and AI agents. Adapters ship weather (METAR, ASOS, GHCNh, NWS CLI) and prediction-market settlements (Kalshi NHIGH/NLOW, Polymarket) today; SEC filings, Federal Reserve series, court filings, FDA approvals, and equities are next. Local-first, no hosted backend.",
   "keywords": [
     "weather",
diff --git a/packages-ts/meta/src/mode2.ts b/packages-ts/meta/src/mode2.ts
index be400431..2d950016 100644
--- a/packages-ts/meta/src/mode2.ts
+++ b/packages-ts/meta/src/mode2.ts
@@ -80,6 +80,51 @@ export function isMode2Source(value: unknown): value is Mode2Source {
   return typeof value === "string" && (MODE2_SOURCES as readonly string[]).includes(value);
 }
 
+// ── research({ source }) filter vocabulary — the WIDENED superset ──────────
+// `researchBySource` (Mode 2 dispatch) narrows to the four dotted forms
+// because a live-IEM fetcher gap means bare/live inputs can't be dispatched.
+// `research({ source })` is a FILTER over already-fetched, already-merged
+// rows, so it accepts Python's full vocabulary. The accepted names and
+// alias-accept sets below are byte-identical to Python
+// `mostlyright.mode2._VALID_OBSERVATION_SOURCES` / `_SOURCE_ALIASES`
+// (packages/core/src/mostlyright/mode2.py:50-63).
+
+/**
+ * research()-filter alias table — the Python-parity superset of
+ * {@link SOURCE_ALIASES}. Extends the narrow dispatch table (the four dotted
+ * forms) with the three bare parser tags (`iem`, `awc`, `ghcnh`) rather than
+ * re-declaring the shared dotted rows, so the dotted aliases have one source
+ * of truth. Mirrors Python `_SOURCE_ALIASES` (7 entries) verbatim.
+ */
+export const RESEARCH_SOURCE_ALIASES: ReadonlyMap<string, ReadonlySet<string>> = new Map<
+  string,
+  ReadonlySet<string>
+>([
+  ...SOURCE_ALIASES,
+  ["iem", new Set(["iem", "iem.archive", "iem.live"])],
+  ["awc", new Set(["awc", "awc.live"])],
+  ["ghcnh", new Set(["ghcnh", "ghcnh.archive"])],
+]);
+
+/**
+ * Accepted `research({ source })` names — the KEYS of
+ * {@link RESEARCH_SOURCE_ALIASES}. Byte-identical to Python
+ * `_VALID_OBSERVATION_SOURCES`: `iem`, `iem.archive`, `iem.live`, `awc`,
+ * `awc.live`, `ghcnh`, `ghcnh.archive`.
+ */
+export const RESEARCH_OBSERVATION_SOURCES: ReadonlySet<string> = new Set(
+  RESEARCH_SOURCE_ALIASES.keys(),
+);
+
+/**
+ * Type-guard: true iff `value` is an accepted `research({ source })` name.
+ * WIDER than {@link isMode2Source} — accepts bare parser tags too, matching
+ * Python's `source in _VALID_OBSERVATION_SOURCES` membership check.
+ */
+export function isResearchObservationSource(value: unknown): boolean {
+  return typeof value === "string" && RESEARCH_OBSERVATION_SOURCES.has(value);
+}
+
 /**
  * Throw {@link SourceMismatchError} if any row's `source` field
  * disagrees with the expected source vocabulary. Rows missing the
diff --git a/packages-ts/meta/src/research.ts b/packages-ts/meta/src/research.ts
index c42ab1a6..5543a561 100644
--- a/packages-ts/meta/src/research.ts
+++ b/packages-ts/meta/src/research.ts
@@ -52,6 +52,11 @@ import {
   parseGhcnhPsv,
   parseIemCsv,
 } from "@mostlyrightmd/weather";
+import {
+  RESEARCH_OBSERVATION_SOURCES,
+  RESEARCH_SOURCE_ALIASES,
+  isResearchObservationSource,
+} from "./mode2.js";
 import { type ResearchKwargsExtension, validateResearchKwargs } from "./research.types.js";
 
 // Re-export PairsRow so callers can `import { research, type PairsRow } from "mostlyright"`.
@@ -109,9 +114,15 @@ export interface ResearchOptions extends ResearchKwargsExtension {
    *  StationOverrideWarning via `onWarning?`; output row carries
    *  `settlementMismatch: true`. Only valid with `contract` selector. */
   stationOverride?: string;
-  /** Mode 1 source subset — dedupe within. Mutually exclusive with `source`. */
+  /** Mode 1 multi-source SUBSET selector — dedupe within. Validated in v0.2;
+   *  its dedupe-within-subset wiring lands in v0.3 (passing it throws).
+   *  Mutually exclusive with `source`. */
   sources?: ReadonlyArray<string>;
-  /** Mode 2 single-source pin — error on mismatch. Mutually exclusive with `sources`. */
+  /** Mode 2 single-source pin. FUNCTIONAL: filters the merged observation
+   *  rows to the requested source's alias set before daily obs_* aggregation
+   *  (cli_* unaffected). Accepts any name in `RESEARCH_OBSERVATION_SOURCES`
+   *  (the four dotted forms plus bare `iem`/`awc`/`ghcnh`). Mutually
+   *  exclusive with `sources`. */
   source?: string;
   /** Attach per-issuer trade timeseries via @mostlyrightmd/markets/trades.
    *  Requires `contract` or `contracts`. */
@@ -785,7 +796,11 @@ async function fetchGhcnhWithCache(
  * in `[fromDate, toDate]`. Each row carries:
  *  - `cli_*` populated from IEM CLI (final preferred per `mergeClimate`).
  *  - `obs_*` daily aggregates over the 3-source merged observations
- *    (AWC > IEM > GHCNh per `mergeObservations`).
+ *    (AWC > IEM > GHCNh per `mergeObservations`). When `opts.source` is set
+ *    (Mode 2 single-source pin), obs rows are filtered to that source's
+ *    alias set AFTER the merge — so pinning a lower-priority source yields
+ *    only the rows the merge did not supersede (see the inline note);
+ *    `cli_*` is unaffected.
  *  - `fcst_*` unconditionally null (Mode 1).
  *  - `market_close_utc` formatted `YYYY-MM-DDTHH:MM:SSZ`.
  *
@@ -854,18 +869,35 @@ export async function research(
     if (hasContracts) names.push("contracts");
     throw new Error(`research(): selectors are mutually exclusive; got ${JSON.stringify(names)}`);
   }
-  if (opts.sources !== undefined && opts.source !== undefined) {
+  if (opts.sources !== undefined && opts.source != null) {
     throw new Error("research(): sources and source are mutually exclusive");
   }
-  // Iter-1 codex HIGH: sources / source validation is shipped in Phase 10
-  // v0.2 but the data-selection wiring lands in v0.3. Without this guard
-  // the station-path runs the full multi-source merge regardless — silent
-  // data-selection corruption.
-  if (opts.sources !== undefined || opts.source !== undefined) {
+  // `sources` (plural) — the Mode 1 multi-source SUBSET selector — is
+  // validated in Phase 10 v0.2 but its dedupe-within-subset wiring lands in
+  // v0.3; passing it throws. NOTE: `source` (singular) is now FUNCTIONAL
+  // (single-source pin, wired below) — only the multi-source subset remains
+  // deferred.
+  if (opts.sources !== undefined) {
+    throw new Error(
+      "research(): the `sources` multi-source subset selector is validated in " +
+        "Phase 10 v0.2 but its dedupe-within-subset wiring lands in v0.3. For " +
+        "single-source pinning today pass `source=` (functional now), or use " +
+        "`researchBySource(station, source, ...)` from @mostlyrightmd/meta.",
+    );
+  }
+  // `source` (singular) — Mode 2 single-source pin. Validate against the
+  // widened Mode 2 vocabulary BEFORE any network fetch so invalid input
+  // never burns API quota. Mirrors Python `research_by_source`'s ValueError
+  // guard (packages/core/src/mostlyright/mode2.py:112-115); the accepted
+  // names are byte-identical across SDKs (RESEARCH_OBSERVATION_SOURCES).
+  // Presence is nullish (`!= null`), NOT `!== undefined`: per the Phase 21
+  // wire contract (research.types.ts `present()`), a JSON-round-tripped
+  // Python `source=None` arrives as `null` and must behave as ABSENT.
+  if (opts.source != null && !isResearchObservationSource(opts.source)) {
     throw new Error(
-      "research(): sources / source validation surface is shipped in Phase 10 v0.2 " +
-        "but the data-selection wiring lands in v0.3. For Mode 2 single-source pinning " +
-        "today, use `researchBySource(station, source, ...)` from @mostlyrightmd/meta.",
+      `Mode 2 source must be one of ${JSON.stringify(
+        [...RESEARCH_OBSERVATION_SOURCES].sort(),
+      )}; got ${JSON.stringify(opts.source)}`,
     );
   }
   if (opts.stationOverride !== undefined && !hasContract) {
@@ -977,12 +1009,32 @@ export async function research(
   const sorted = sortByObservedAtThenSource(combinedRaw);
   const merged = mergeObservations(sorted);
 
+  // ── Mode 2 single-source pin (opts.source) ────────────────────────────
+  // When `source` is set, keep only merged observation rows whose
+  // parser-emitted `source` tag is in the alias-accept set for the requested
+  // source — applied BEFORE daily obs_* aggregation. The cli_* settlement
+  // columns are UNAFFECTED (CLI is independent of the observation pin).
+  // Mirrors Python `mode2.research_by_source`'s post-merge source filter.
+  //
+  // Inherited limitation (Python mode2.py:85-93 parity): the AWC > IEM >
+  // GHCNh merge above runs on the full multi-source set, so pinning e.g.
+  // `iem` returns only the IEM rows AWC did NOT supersede on a colliding
+  // (station_code, observed_at, observation_type) key — not the full
+  // upstream IEM coverage. Callers needing pre-merge isolation should call
+  // the per-source fetchers in @mostlyrightmd/weather directly.
+  let observationsForBuild: ReadonlyArray<Observation> = merged;
+  if (opts.source != null) {
+    const accept = RESEARCH_SOURCE_ALIASES.get(opts.source);
+    // `accept` is guaranteed defined — `opts.source` validated above.
+    observationsForBuild = merged.filter((r) => accept?.has(r.source) ?? false);
+  }
+
   const observationsByDate: Record<string, PairsObservationLike[]> = {};
   // dates is guaranteed non-empty by buildDateList contract (throws on
   // fromDate > toDate; both validated above).
   const dateLo = dates[0] ?? "";
   const dateHi = dates[dates.length - 1] ?? "";
-  for (const obs of merged) {
+  for (const obs of observationsForBuild) {
     const settleDate = observedSettlementDate(obs.observed_at, resolved.code);
     if (settleDate === null) continue;
     if (settleDate < dateLo || settleDate > dateHi) continue;
diff --git a/packages-ts/meta/tests/compose.test.ts b/packages-ts/meta/tests/compose.test.ts
index 56ee536f..44b93665 100644
--- a/packages-ts/meta/tests/compose.test.ts
+++ b/packages-ts/meta/tests/compose.test.ts
@@ -239,10 +239,15 @@ describe("research() Phase 10 signature validation", () => {
     ).rejects.toThrow(/v0.3/);
   });
 
-  it("source alone raises v0.3 deferred (iter-1 codex HIGH)", async () => {
+  it("source alone is now FUNCTIONAL — an unknown source is rejected before any fetch", async () => {
+    // `source=` (single-source pin) is wired in this PR; it no longer defers
+    // to v0.3. Validation of an unknown source fires BEFORE any network call
+    // (mirrors Python `research_by_source`'s ValueError guard), so this stays
+    // network-free. The functional filter path over mocked fetchers is
+    // covered in research.test.ts.
     await expect(
-      research("KNYC", "2025-01-06", "2025-01-12", { source: "iem.archive" }),
-    ).rejects.toThrow(/researchBySource/);
+      research("KNYC", "2025-01-06", "2025-01-12", { source: "not-a-real-source" }),
+    ).rejects.toThrow(/Mode 2 source must be one of/);
   });
 
   it("stationOverride requires contract", async () => {
diff --git a/packages-ts/meta/tests/mode2.test.ts b/packages-ts/meta/tests/mode2.test.ts
index 93774ee4..7d3fb61d 100644
--- a/packages-ts/meta/tests/mode2.test.ts
+++ b/packages-ts/meta/tests/mode2.test.ts
@@ -246,6 +246,26 @@ describe("researchBySource — iem.archive dispatch", () => {
     });
     expect(rows).toEqual([]);
   });
+
+  it("window boundary: keeps rows ON fromDate + ON toDate; drops fromDate-1 + toDate+1", async () => {
+    // The window filter is inclusive on both ends (observed_at.slice(0,10)
+    // >= fromDate && <= toDate). Pin the exact boundary rows.
+    const csv = iemCsv([
+      makeIemRow({ valid: "2024-06-09 12:51" }), // fromDate-1 → DROP
+      makeIemRow({ valid: "2024-06-10 12:51" }), // fromDate   → KEEP
+      makeIemRow({ valid: "2024-06-20 12:51" }), // toDate     → KEEP
+      makeIemRow({ valid: "2024-06-21 12:51" }), // toDate+1   → DROP
+    ]);
+    installFetchMock({ iemAsosCsv: csv });
+    const rows = await researchBySource("NYC", "iem.archive", "2024-06-10", "2024-06-20", {
+      iemPolitenessMs: 0,
+    });
+    const days = new Set(rows.map((r) => r.observed_at.slice(0, 10)));
+    expect(days.has("2024-06-10")).toBe(true); // ON fromDate kept
+    expect(days.has("2024-06-20")).toBe(true); // ON toDate kept
+    expect(days.has("2024-06-09")).toBe(false); // fromDate-1 dropped
+    expect(days.has("2024-06-21")).toBe(false); // toDate+1 dropped
+  });
 });
 
 describe("researchBySource — awc.live dispatch", () => {
@@ -315,6 +335,26 @@ describe("researchBySource — awc.live dispatch", () => {
     const outDates = rows.map((r) => r.observed_at.slice(0, 10));
     expect(outDates).not.toContain(outOfWindowFuture);
   });
+
+  it("window boundary: keeps rows ON fromDate + ON toDate; drops fromDate-1 + toDate+1", async () => {
+    // Multi-day window pins fromDate and toDate as DISTINCT boundaries (the
+    // sibling test above uses a degenerate single-day window). The AWC branch
+    // filters on observed_at.slice(0,10) inclusive on both ends. The mock
+    // returns records verbatim regardless of `hours`, so fixed dates are safe.
+    const awc: AwcMockMetar[] = [
+      { icaoId: "KNYC", obsTime: epochOfUtc("2024-06-09T12:00:00Z"), temp: 20 }, // fromDate-1 DROP
+      { icaoId: "KNYC", obsTime: epochOfUtc("2024-06-10T12:00:00Z"), temp: 21 }, // fromDate   KEEP
+      { icaoId: "KNYC", obsTime: epochOfUtc("2024-06-20T12:00:00Z"), temp: 22 }, // toDate     KEEP
+      { icaoId: "KNYC", obsTime: epochOfUtc("2024-06-21T12:00:00Z"), temp: 23 }, // toDate+1   DROP
+    ];
+    installFetchMock({ awc });
+    const rows = await researchBySource("NYC", "awc.live", "2024-06-10", "2024-06-20");
+    const days = new Set(rows.map((r) => r.observed_at.slice(0, 10)));
+    expect(days.has("2024-06-10")).toBe(true); // ON fromDate kept
+    expect(days.has("2024-06-20")).toBe(true); // ON toDate kept
+    expect(days.has("2024-06-09")).toBe(false); // fromDate-1 dropped
+    expect(days.has("2024-06-21")).toBe(false); // toDate+1 dropped
+  });
 });
 
 describe("researchBySource — ghcnh.archive dispatch", () => {
diff --git a/packages-ts/meta/tests/research.test.ts b/packages-ts/meta/tests/research.test.ts
index 383edd54..2466a097 100644
--- a/packages-ts/meta/tests/research.test.ts
+++ b/packages-ts/meta/tests/research.test.ts
@@ -360,3 +360,191 @@ describe("research() — TS-W1 Wave 6 (AWC + CLI)", () => {
     expect(rows[0]?.obs_high_f).not.toBeNull();
   });
 });
+
+// ---------------------------------------------------------------------------
+// Mode 2 single-source pin — research({ source })
+// ---------------------------------------------------------------------------
+//
+// `source=` filters the merged observation rows to the requested source's
+// alias set BEFORE daily obs_* aggregation (cli_* is unaffected). The fixture
+// below carries one AWC row (40°C → 104°F, the MAX), one IEM CSV row
+// (tmpf=75 → 75°F, parsed as both METAR + SPECI = 2 rows), and one GHCNh row
+// (20°C → 68°F, the MIN). Distinct observed_at values keep all four rows past
+// `mergeObservations`, so pinning provably drops the non-matching sources.
+
+const PIN_DAY = "2025-06-10";
+const PIN_NOW = new Date("2025-06-10T20:00:00Z"); // inside AWC's 168h window
+
+const IEM_HEADER =
+  "station,valid,tmpf,dwpf,drct,sknt,gust,alti,mslp,vsby,skyc1,skyl1,skyc2,skyl2,skyc3,skyl3,skyc4,skyl4,wxcodes,p01i,snowdepth,peak_wind_gust,peak_wind_drct,peak_wind_time,metar";
+
+function makeIemRow(overrides: Record<string, string> = {}): string {
+  const defaults: Record<string, string> = {
+    station: "NYC",
+    valid: `${PIN_DAY} 12:51`,
+    tmpf: "75",
+    dwpf: "55",
+    drct: "180",
+    sknt: "8",
+    gust: "",
+    alti: "29.92",
+    mslp: "1013.2",
+    vsby: "10",
+    skyc1: "FEW",
+    skyl1: "2500",
+    skyc2: "",
+    skyl2: "",
+    skyc3: "",
+    skyl3: "",
+    skyc4: "",
+    skyl4: "",
+    wxcodes: "",
+    p01i: "",
+    snowdepth: "",
+    peak_wind_gust: "",
+    peak_wind_drct: "",
+    peak_wind_time: "",
+    metar: "KNYC 101251Z 18008KT 10SM FEW025 24/13 A2992 RMK AO2",
+    ...overrides,
+  };
+  return IEM_HEADER.split(",")
+    .map((c) => defaults[c] ?? "")
+    .join(",");
+}
+
+function iemCsv(rows: string[]): string {
+  return `${[IEM_HEADER, ...rows].join("\n")}\n`;
+}
+
+const GHCNH_HEADER =
+  "temperature_Source_Station_ID|DATE|temperature|temperature_Quality_Code|dew_point_temperature|dew_point_temperature_Quality_Code";
+
+function ghcnhPsv(date: string, tempC: string): string {
+  return `${GHCNH_HEADER}\nICAO-KNYC|${date}|${tempC}|0|20.0|0\n`;
+}
+
+// Mixed-source fixture: AWC (104°F, max) + IEM (75°F ×2) + GHCNh (68°F, min).
+function installMixedSourceMock() {
+  return installFetchMock({
+    awc: [
+      {
+        icaoId: "KNYC",
+        obsTime: epochOfUtc(`${PIN_DAY}T18:00:00Z`),
+        temp: 40, // 104°F — MAX; excluded whenever the pin isn't awc
+        dewp: 10,
+        wspd: 5,
+      },
+    ],
+    iemAsosCsv: iemCsv([makeIemRow({ valid: `${PIN_DAY} 12:51`, tmpf: "75" })]),
+    ghcnhPsv: ghcnhPsv(`${PIN_DAY}T06:00:00Z`, "20.0"), // 68°F — MIN
+    ghcnhStatus: 200,
+  });
+}
+
+describe("research() — Mode 2 single-source pin (opts.source)", () => {
+  beforeEach(() => vi.restoreAllMocks());
+  afterEach(() => vi.restoreAllMocks());
+
+  it("source:'iem' computes obs_* from IEM rows only (AWC max temp excluded)", async () => {
+    installMixedSourceMock();
+    const rows = await research("NYC", PIN_DAY, PIN_DAY, {
+      source: "iem",
+      now: PIN_NOW,
+      cache: null,
+    });
+    expect(rows).toHaveLength(1);
+    const row = rows[0];
+    // IEM contributes 2 rows (METAR + SPECI off the same CSV), both 75°F.
+    expect(row?.obs_count).toBe(2);
+    expect(row?.obs_high_f).toBe(75); // NOT 104 (AWC) → AWC provably excluded
+    expect(row?.obs_low_f).toBe(75); // NOT 68 (GHCNh) → GHCNh provably excluded
+    // cli_* is independent of the obs pin (no CLI fixture → null).
+    expect(row?.cli_high_f).toBeNull();
+  });
+
+  it("source:'awc.live' accepts bare 'awc'-tagged rows via the alias set", async () => {
+    installMixedSourceMock();
+    const rows = await research("NYC", PIN_DAY, PIN_DAY, {
+      source: "awc.live",
+      now: PIN_NOW,
+      cache: null,
+    });
+    expect(rows).toHaveLength(1);
+    const row = rows[0];
+    // Only the single bare-'awc' row survives ('awc' ∈ alias set of 'awc.live').
+    expect(row?.obs_count).toBe(1);
+    expect(row?.obs_high_f).toBe(104);
+    expect(row?.obs_low_f).toBe(104);
+  });
+
+  it("unknown source throws BEFORE any fetch", async () => {
+    const fetchSpy = installMixedSourceMock();
+    await expect(
+      research("NYC", PIN_DAY, PIN_DAY, { source: "banana", now: PIN_NOW, cache: null }),
+    ).rejects.toThrow(/Mode 2 source must be one of/);
+    expect(fetchSpy).not.toHaveBeenCalled();
+  });
+
+  it("source + sources together still throws the mutual-exclusion error", async () => {
+    const fetchSpy = installFetchMock({});
+    await expect(
+      research("NYC", PIN_DAY, PIN_DAY, {
+        source: "iem",
+        sources: ["awc"],
+        now: PIN_NOW,
+        cache: null,
+      }),
+    ).rejects.toThrow(/mutually exclusive/);
+    expect(fetchSpy).not.toHaveBeenCalled();
+  });
+
+  it("sources:['iem'] alone still throws — rewritten message (no longer claims `source` is unwired)", async () => {
+    const fetchSpy = installFetchMock({});
+    let err: unknown = null;
+    try {
+      await research("NYC", PIN_DAY, PIN_DAY, { sources: ["iem"], now: PIN_NOW, cache: null });
+    } catch (e) {
+      err = e;
+    }
+    expect(err).toBeInstanceOf(Error);
+    const msg = (err as Error).message;
+    expect(msg).toMatch(/sources/);
+    expect(msg).toMatch(/v0\.3/);
+    // The rewritten message points at the now-functional `source=`.
+    expect(msg).toMatch(/functional now/);
+    expect(fetchSpy).not.toHaveBeenCalled();
+  });
+
+  it("no `source` → full merge unchanged (golden: AWC max + all 4 rows present)", async () => {
+    installMixedSourceMock();
+    const rows = await research("NYC", PIN_DAY, PIN_DAY, { now: PIN_NOW, cache: null });
+    expect(rows).toHaveLength(1);
+    const row = rows[0];
+    // Golden for the unpinned path: full 3-source merge — AWC 104°F max,
+    // GHCNh 68°F min, all four merged rows counted, mean over all four.
+    expect(row?.obs_count).toBe(4);
+    expect(row?.obs_high_f).toBe(104);
+    expect(row?.obs_low_f).toBe(68);
+    expect(row?.obs_mean_f).toBe(80.5);
+  });
+
+  it("source: null (JSON round-trip of Python None) behaves as ABSENT — full merge, no throw", async () => {
+    // Phase 21 wire contract (research.types.ts `present()`): a Python
+    // `source=None` crosses the JSON wire as `null` and must be treated as
+    // not-provided, identically to omitting the key. Codex P2 regression
+    // guard: `!== undefined` presence tests would throw the unknown-source
+    // error here.
+    installMixedSourceMock();
+    const rows = await research("NYC", PIN_DAY, PIN_DAY, {
+      source: null as unknown as string,
+      now: PIN_NOW,
+      cache: null,
+    });
+    expect(rows).toHaveLength(1);
+    const row = rows[0];
+    expect(row?.obs_count).toBe(4);
+    expect(row?.obs_high_f).toBe(104);
+    expect(row?.obs_low_f).toBe(68);
+    expect(row?.obs_mean_f).toBe(80.5);
+  });
+});
diff --git a/packages-ts/weather/package.json b/packages-ts/weather/package.json
index 85244385..eae22685 100644
--- a/packages-ts/weather/package.json
+++ b/packages-ts/weather/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@mostlyrightmd/weather",
-  "version": "1.11.0",
+  "version": "1.12.0",
   "description": "Weather data for TypeScript / Node — live METAR (AWC), ASOS archive (IEM), historical observations (GHCNh), and NWS climate text products (CLI). For quants, ML training pipelines, and weather-bot agents. Direct public-API access, no hosted backend.",
   "keywords": [
     "weather",
diff --git a/packages/core/pyproject.toml b/packages/core/pyproject.toml
index 6812e940..3096be2f 100644
--- a/packages/core/pyproject.toml
+++ b/packages/core/pyproject.toml
@@ -1,6 +1,6 @@
 [project]
 name = "mostlyrightmd"
-version = "1.11.0"
+version = "1.12.0"
 description = "Python SDK for quants, ML engineers, and AI agents — one interface to public data. Adapters ship weather + prediction-market settlements (Kalshi NHIGH/NLOW, Polymarket) today; SEC filings, Federal Reserve series, court filings, FDA approvals, and equities are next. Schema-versioned, leakage-free, local-first. Imports as `mostlyright`."
 readme = "README.md"
 license = "MIT"
diff --git a/packages/core/src/mostlyright/mode2.py b/packages/core/src/mostlyright/mode2.py
index 5111d827..e4f3b4e8 100644
--- a/packages/core/src/mostlyright/mode2.py
+++ b/packages/core/src/mostlyright/mode2.py
@@ -63,6 +63,22 @@
 }
 
 
+def _in_query_window(observed_at: Any, from_date: str, to_date: str) -> bool:
+    """True iff ``observed_at``'s UTC date-part is in ``[from_date, to_date]``.
+
+    Compares the ISO date-part slice (``observed_at[:10]``) inclusively at
+    both ends — a pure string comparison, no timezone math. The slice
+    handles both emitted forms (``"2025-01-01T00:51:00Z"`` and
+    ``"2025-01-06T12:00:00+00:00"``) identically. Rows with missing/empty
+    ``observed_at`` return False: they cannot be placed in any window.
+    Must match the TS port (``packages-ts/meta/src/mode2.ts``) for
+    cross-SDK lockstep.
+    """
+    if not observed_at:
+        return False
+    return from_date <= observed_at[:10] <= to_date
+
+
 def research_by_source(
     station: str,
     source: str,
@@ -102,7 +118,13 @@ def research_by_source(
 
     Returns:
         DataFrame (or list[dict]) with one row per observation matching
-        the requested source.
+        the requested source, bounded to ``[from_date, to_date]`` by UTC
+        date-part (``observed_at[:10]``), inclusive at both ends. The fetch
+        extends one calendar day past ``to_date`` to capture the last LST
+        day's pre-midnight UTC tail, but those extension rows are trimmed
+        out of the result. Callers doing LST settlement math who need that
+        pre-midnight UTC tail of the final LST day should pass a ``to_date``
+        one day later than their settlement window.
 
     Raises:
         ValueError: ``source`` not in the supported set.
@@ -131,8 +153,10 @@ def research_by_source(
     )
 
     info = _resolve_station(station)
-    # Mirror research(): extend the fetch by one day so the last LST
-    # day's pre-midnight UTC tail observations are captured.
+    # Mirror research(): extend the fetch by one day so the last LST day's
+    # pre-midnight UTC tail observations are captured. The returned rows
+    # are trimmed back to the strict [from_date, to_date] window below —
+    # the extension is only a fetch-width knob, not a result-width one.
     extended_to = (_date.fromisoformat(to_date) + timedelta(days=1)).isoformat()
 
     awc_rows: list[dict[str, Any]] | None = None
@@ -152,8 +176,22 @@ def research_by_source(
     # mostlyright' canonical vocab is dotted; the alias table bridges
     # both at the boundary without rewriting the per-row source —
     # downstream Validator sees the truthful parser-emitted tag).
+    #
+    # Trim to the strict [from_date, to_date] window on the same pass.
+    # `_fetch_observations_range` returns whole cached calendar months
+    # (month-granular parquet — research.py extends cached months verbatim)
+    # plus the +1-day fetch extension above, so raw_obs leaks rows outside
+    # the caller's window. This single filter fixes EVERY return shape
+    # (DataFrame, as_dataframe=False list, wrapper/backends) because it runs
+    # before any of them. Rows with missing/empty observed_at are dropped
+    # by `_in_query_window` (they cannot be placed in any window).
     accepted_sources = _SOURCE_ALIASES.get(source, {source})
-    filtered = [r for r in raw_obs if r.get("source") in accepted_sources]
+    filtered = [
+        r
+        for r in raw_obs
+        if r.get("source") in accepted_sources
+        and _in_query_window(r.get("observed_at"), from_date, to_date)
+    ]
 
     if not as_dataframe:
         return filtered
diff --git a/packages/core/src/mostlyright/research.py b/packages/core/src/mostlyright/research.py
index 53e7bd60..b073b424 100644
--- a/packages/core/src/mostlyright/research.py
+++ b/packages/core/src/mostlyright/research.py
@@ -1612,6 +1612,24 @@ def research(
             :data:`mostlyright.snapshot._STATION_TZ`. Passed through to
             settlement-date math; rarely needed for the US registry
             (all entries are covered).
+        source: Pin observations to a single named source (Mode 2). One of
+            :data:`mostlyright.mode2._VALID_OBSERVATION_SOURCES` (bare
+            ``"iem"``/``"awc"``/``"ghcnh"`` or the dotted canonical forms);
+            validated early, so an unknown value raises ``ValueError``
+            before any fetch. When set, observation rows are filtered to the
+            requested source's alias-accept set before aggregation, and
+            ``df.attrs["source"]`` records the pin. The ``cli_*`` columns are
+            UNAFFECTED — NWS CLI is the settlement source, independent of the
+            observation-source pin. **v0.1.0 limitation:** the Mode 1 merge
+            policy (AWC > IEM > GHCNh on key collision) runs BEFORE the pin
+            filter, so pinning ``"iem"`` over a window where AWC also
+            reported yields only the IEM rows AWC did not supersede — not the
+            full upstream IEM coverage. ``source=None`` (default) leaves the
+            path byte-identical to the v0.14.1 parity baseline.
+        sources: Multi-source subset selector (plural). Validated at the
+            mutual-exclusion boundary but the data-selection wiring lands in
+            v0.3 — passing it raises ``NotImplementedError``. Mutually
+            exclusive with ``source=``.
 
     Returns:
         DataFrame (or ``list[dict]`` when ``as_dataframe=False``) with **one
@@ -1704,22 +1722,38 @@ def research(
     if sources is not None and source is not None:
         raise ValueError("research(): sources= and source= are mutually exclusive")
 
-    # Iter-1 codex HIGH: sources= / source= are validated at the
-    # mutual-exclusion boundary but the actual data-selection wiring lands
-    # in v0.3. The station-path silently runs the full multi-source merge
-    # regardless of these kwargs, which would be silent data-selection
-    # corruption. Surface a clear NotImplementedError pointing callers at
-    # `mostlyright.mode2.research_by_source` (the Mode-2-pin path) until
-    # the kwargs are wired into the station-path dispatch.
-    if sources is not None or source is not None:
+    # `sources=` (plural, multi-source subset) is validated at the
+    # mutual-exclusion boundary but its data-selection wiring lands in
+    # v0.3. The singular `source=` pin IS wired below (a post-fetch
+    # single-source row subset), so only the plural form still raises.
+    if sources is not None:
         raise NotImplementedError(
-            "research(): sources= and source= validation surface is shipped in "
-            "Phase 10 v0.2 but the data-selection wiring lands in v0.3. For "
-            "Mode 2 single-source pinning today, use "
-            "`mostlyright.mode2.research_by_source(station, source, ...)` "
-            "directly. Mode 1 multi-source subset (sources=[...]) ships in v0.3."
+            "research(): the sources= multi-source subset selector is validated "
+            "(mutual exclusion) but its data-selection wiring lands in v0.3. Pin "
+            "a single observation source with source=... today; the multi-source "
+            "subset (sources=[...]) is the v0.3 deliverable."
         )
 
+    # Mode 2 single-source pin. Validate source= EARLY — before any network
+    # or cache work — against Mode 2's source vocabulary, then compute the
+    # alias-accept set used to subset the raw observation rows after the
+    # fetch. Reuse mode2's tables rather than duplicating them. The import
+    # is function-local (not module-level): mode2 already imports research
+    # function-locally, so a top-level `from mostlyright.mode2 import ...`
+    # here would add a research->mode2 module edge and make the coupling
+    # bidirectional; a call-time import keeps it one-directional and is
+    # provably non-circular (mode2 has no top-level research import).
+    accepted_sources: frozenset[str] | None = None
+    if source is not None:
+        from mostlyright.mode2 import _SOURCE_ALIASES, _VALID_OBSERVATION_SOURCES
+
+        if source not in _VALID_OBSERVATION_SOURCES:
+            raise ValueError(
+                f"research(): source must be one of "
+                f"{sorted(_VALID_OBSERVATION_SOURCES)}; got {source!r}"
+            )
+        accepted_sources = _SOURCE_ALIASES.get(source, frozenset({source}))
+
     # `station_override=` only makes sense when paired with `contract=`.
     if station_override is not None and not _has_contract:
         raise ValueError(
@@ -1846,6 +1880,25 @@ def research(
         extended_to,
         prefetched_awc_rows=awc_rows,
     )
+
+    # Mode 2 single-source pin (source=): keep only rows whose parser-
+    # emitted `source` is in the requested alias-accept set. Applied as a
+    # pure post-fetch row subset BEFORE the settlement-date bucketing below,
+    # which already trims the window — so the pinned frame keeps the same
+    # columns/shape as unpinned research(). PARITY FIREWALL: this runs only
+    # when source is not None (accepted_sources stays None otherwise), so
+    # the default path is byte-identical. cli_* (NWS CLI settlement) columns
+    # are unaffected — CLI is the settlement source, independent of the
+    # observation-source pin.
+    #
+    # v0.1.0 limitation (inherited from Mode 2): _fetch_observations_range
+    # applies the Mode 1 merge policy (AWC > IEM > GHCNh on key collision)
+    # BEFORE this filter, so pinning "iem" over a window where AWC also
+    # reported yields only the IEM rows AWC did NOT supersede — not the full
+    # upstream IEM coverage. A pre-merge source-isolated path is v0.2 work.
+    if accepted_sources is not None:
+        raw_obs = [r for r in raw_obs if r.get("source") in accepted_sources]
+
     raw_climate = _fetch_climate_range(info, from_date, to_date)
 
     # Phase 3.4: opt-in QC. Runs the QCEngine + IEM-vs-GHCNh crosscheck
@@ -1912,6 +1965,16 @@ def research(
         with contextlib.suppress(AttributeError):
             result.attrs["qc"] = qc_summary
 
+    # Mode 2 provenance contract: stamp the requested source on the
+    # returned DataFrame when pinned (mirrors mode2.research_by_source's
+    # df.attrs["source"]). No attrs change when source is None — the
+    # parity firewall keeps the default path untouched.
+    if source is not None and as_dataframe:
+        import contextlib
+
+        with contextlib.suppress(AttributeError):
+            result.attrs["source"] = source
+
     # Phase 6 W3-T2: when as_dataframe=False the caller wants raw list[dict] —
     # backend/return_type kwargs do not apply. Same when backend kwarg is the
     # default (pandas + dataframe) — return unchanged for zero-overhead.
diff --git a/packages/core/tests/test_compose.py b/packages/core/tests/test_compose.py
index 97fc49c6..9e5ced2b 100644
--- a/packages/core/tests/test_compose.py
+++ b/packages/core/tests/test_compose.py
@@ -289,27 +289,34 @@ def test_sources_and_source_mutually_exclusive(self):
 
     def test_sources_only_raises_notimplemented_v03(self):
         """Iter-1 codex HIGH: sources= silently bypassed the data-fetch
-        wiring. v0.2 now raises NotImplementedError directing callers at
-        researchBySource for Mode-2 today + v0.3 for sources=[]."""
+        wiring. The plural multi-source subset stays deferred to v0.3;
+        the rewritten message no longer redirects source= at
+        research_by_source (source= is now wired — see Fix B)."""
         from mostlyright import research
 
-        with pytest.raises(NotImplementedError, match=r"v0\.3"):
+        with pytest.raises(NotImplementedError, match=r"v0\.3") as exc:
             research(
                 station="NYC",
                 from_date="2025-01-06",
                 to_date="2025-01-12",
                 sources=["iem.archive"],
             )
-
-    def test_source_only_raises_notimplemented_v03(self):
+        assert "research_by_source" not in str(exc.value)
+
+    def test_source_only_is_wired_not_notimplemented(self):
+        """Fix B: source= is now wired (post-fetch single-source pin), so it
+        no longer raises NotImplementedError. It validates its vocabulary
+        early — an unknown source raises ValueError before any fetch; a
+        valid source runs the join. Full pin behaviour lives in
+        test_research_source_pin.py."""
         from mostlyright import research
 
-        with pytest.raises(NotImplementedError, match="research_by_source"):
+        with pytest.raises(ValueError, match="source must be one of"):
             research(
                 station="NYC",
                 from_date="2025-01-06",
                 to_date="2025-01-12",
-                source="iem.archive",
+                source="bogus.source",
             )
 
     def test_station_override_requires_contract(self):
diff --git a/packages/core/tests/test_mode2.py b/packages/core/tests/test_mode2.py
index 6b87cadd..c12fa1bf 100644
--- a/packages/core/tests/test_mode2.py
+++ b/packages/core/tests/test_mode2.py
@@ -180,3 +180,108 @@ def test_assert_source_identity_no_source_column_is_noop():
 def test_assert_source_identity_empty_df():
     df = pd.DataFrame({"source": []})
     assert_source_identity(df, "iem.archive")
+
+
+# ----------------------------------------------------------------------
+# Fix A — window trim. `_fetch_observations_range` returns whole cached
+# calendar months (month-granular parquet) plus the +1-day fetch
+# extension, so raw_obs leaks rows outside the caller's window. Mode 2
+# must trim to [from_date, to_date] inclusive by UTC date-part. The
+# pre-fix tests above only ever fed in-window rows, so they never caught
+# this — these feed out-of-window rows on purpose.
+# ----------------------------------------------------------------------
+def _patch_fetch(monkeypatch, tmp_path, rows: list[dict]) -> None:
+    """Stub the research-module fetch seam mode2 imports function-locally."""
+    import importlib
+
+    monkeypatch.setenv("MOSTLYRIGHT_CACHE_DIR", str(tmp_path))
+    research_module = importlib.import_module("mostlyright.research")
+    monkeypatch.setattr(research_module, "_fetch_observations_range", lambda *a, **kw: rows)
+    monkeypatch.setattr(research_module, "_all_caches_warm", lambda *a, **kw: True)
+
+
+def test_window_trim_dataframe_full_month(tmp_path, monkeypatch) -> None:
+    """A full month of rows is trimmed to just the queried [from, to] window.
+
+    Boundary rows ON from_date (Jan 6) and ON to_date (Jan 12) are kept;
+    the row on to_date+1 (Jan 13, the +1-day fetch extension) is dropped.
+    """
+    rows = [
+        {
+            "station_code": "KNYC",
+            "observed_at": f"2025-01-{d:02d}T12:00:00Z",
+            "source": "iem",
+            "temp_c": 5.0,
+        }
+        for d in (1, 3, 5, 6, 10, 12, 13, 20, 31)
+    ]
+    _patch_fetch(monkeypatch, tmp_path, rows)
+    df = research_by_source("KNYC", "iem", "2025-01-06", "2025-01-12")
+    assert set(df["observed_at"]) == {
+        "2025-01-06T12:00:00Z",  # ON from_date — kept
+        "2025-01-10T12:00:00Z",
+        "2025-01-12T12:00:00Z",  # ON to_date — kept
+    }
+
+
+def test_window_trim_list_full_month(tmp_path, monkeypatch) -> None:
+    """The same trim applies on the as_dataframe=False raw-rows path."""
+    rows = [
+        {
+            "station_code": "KNYC",
+            "observed_at": f"2025-01-{d:02d}T12:00:00Z",
+            "source": "iem",
+            "temp_c": 5.0,
+        }
+        for d in (1, 3, 5, 6, 10, 12, 13, 20, 31)
+    ]
+    _patch_fetch(monkeypatch, tmp_path, rows)
+    out = research_by_source("KNYC", "iem", "2025-01-06", "2025-01-12", as_dataframe=False)
+    assert isinstance(out, list)
+    assert {r["observed_at"] for r in out} == {
+        "2025-01-06T12:00:00Z",
+        "2025-01-10T12:00:00Z",
+        "2025-01-12T12:00:00Z",
+    }
+
+
+def test_window_trim_month_straddling(tmp_path, monkeypatch) -> None:
+    """A Jan 28 to Feb 3 window keeps in-window rows from BOTH months only."""
+    rows = [
+        {"station_code": "KNYC", "observed_at": ts, "source": "iem", "temp_c": 5.0}
+        for ts in (
+            "2025-01-27T12:00:00Z",  # before window — dropped
+            "2025-01-28T12:00:00Z",  # ON from_date — kept
+            "2025-01-31T12:00:00Z",  # kept
+            "2025-02-01T12:00:00Z",  # kept
+            "2025-02-03T12:00:00Z",  # ON to_date — kept
+            "2025-02-04T12:00:00Z",  # to_date+1 extension — dropped
+        )
+    ]
+    _patch_fetch(monkeypatch, tmp_path, rows)
+    df = research_by_source("KNYC", "iem", "2025-01-28", "2025-02-03")
+    assert set(df["observed_at"]) == {
+        "2025-01-28T12:00:00Z",
+        "2025-01-31T12:00:00Z",
+        "2025-02-01T12:00:00Z",
+        "2025-02-03T12:00:00Z",
+    }
+
+
+def test_window_trim_mixed_timestamp_formats_and_missing(tmp_path, monkeypatch) -> None:
+    """Both the "Z" and "+00:00" forms trim correctly via the [:10] slice;
+    rows with missing/empty observed_at are dropped (cannot be placed)."""
+    rows = [
+        {"station_code": "KNYC", "observed_at": "2025-01-06T00:51:00Z", "source": "iem"},
+        {"station_code": "KNYC", "observed_at": "2025-01-08T12:00:00+00:00", "source": "iem"},
+        {"station_code": "KNYC", "observed_at": "2025-01-13T00:00:00+00:00", "source": "iem"},
+        {"station_code": "KNYC", "observed_at": "2025-01-05T23:59:00Z", "source": "iem"},
+        {"station_code": "KNYC", "observed_at": "", "source": "iem"},
+        {"station_code": "KNYC", "source": "iem"},  # no observed_at key
+    ]
+    _patch_fetch(monkeypatch, tmp_path, rows)
+    df = research_by_source("KNYC", "iem", "2025-01-06", "2025-01-12")
+    assert set(df["observed_at"]) == {
+        "2025-01-06T00:51:00Z",  # ON from_date, Z form — kept
+        "2025-01-08T12:00:00+00:00",  # in-window, +00:00 form — kept
+    }
diff --git a/packages/core/tests/test_research_source_pin.py b/packages/core/tests/test_research_source_pin.py
new file mode 100644
index 00000000..232db6fa
--- /dev/null
+++ b/packages/core/tests/test_research_source_pin.py
@@ -0,0 +1,190 @@
+"""Fix B — ``research(source=...)`` single-source pinning.
+
+Wires the singular ``source=`` pin (previously a ``NotImplementedError``
+stub) as a pure post-fetch row subset applied only when ``source`` is not
+None. The plural ``sources=`` multi-source subset stays deferred to v0.3.
+
+The fetcher is monkeypatched (the research module's function-local seam,
+per the established ``test_qc_wired.py`` pattern) so no test hits the
+network. Observation rows carry ``temp_f`` because the pairs aggregation
+(``_obs_aggregates``) reads that field for ``obs_high_f``/``obs_low_f``.
+"""
+
+from __future__ import annotations
+
+import importlib
+
+import pandas as pd
+import pytest
+
+
+def _load():
+    return importlib.import_module("mostlyright.research")
+
+
+def _patch_fetchers(research_module, monkeypatch, tmp_path, obs) -> None:
+    monkeypatch.setenv("MOSTLYRIGHT_CACHE_DIR", str(tmp_path))
+    monkeypatch.setattr(research_module, "_fetch_observations_range", lambda *a, **kw: list(obs))
+    monkeypatch.setattr(research_module, "_fetch_climate_range", lambda *a, **kw: [])
+    monkeypatch.setattr(research_module, "_all_caches_warm", lambda *a, **kw: True)
+
+
+def test_source_pin_filters_obs_aggregates(tmp_path, monkeypatch) -> None:
+    """With source="iem", obs_* reflect ONLY iem rows.
+
+    All four rows bucket into the Jan 8 settlement date for KNYC (EST →
+    the mid-day UTC stamps share the same local date). The AWC row carries
+    the window max (99F) and the GHCNh row the min (20F); both must be
+    excluded so the pinned Jan 8 row shows the iem-only 45/40 spread.
+    """
+    research_module = _load()
+    obs = [
+        {
+            "station_code": "KNYC",
+            "observed_at": "2025-01-08T15:00:00Z",
+            "source": "iem",
+            "temp_f": 40.0,
+        },
+        {
+            "station_code": "KNYC",
+            "observed_at": "2025-01-08T18:00:00Z",
+            "source": "iem",
+            "temp_f": 45.0,
+        },
+        {
+            "station_code": "KNYC",
+            "observed_at": "2025-01-08T16:00:00Z",
+            "source": "awc",
+            "temp_f": 99.0,
+        },
+        {
+            "station_code": "KNYC",
+            "observed_at": "2025-01-08T17:00:00Z",
+            "source": "ghcnh",
+            "temp_f": 20.0,
+        },
+    ]
+    _patch_fetchers(research_module, monkeypatch, tmp_path, obs)
+
+    df = research_module.research("KNYC", "2025-01-06", "2025-01-12", source="iem")
+    row = df.loc[pd.Timestamp("2025-01-08")]
+    assert row["obs_high_f"] == 45.0
+    assert row["obs_low_f"] == 40.0
+    assert row["obs_count"] == 2
+    assert df.attrs["source"] == "iem"
+
+
+def test_source_pin_alias_awc_live_accepts_bare_awc(tmp_path, monkeypatch) -> None:
+    """source="awc.live" accepts alias-tagged bare "awc" rows (alias table)."""
+    research_module = _load()
+    obs = [
+        {
+            "station_code": "KNYC",
+            "observed_at": "2025-01-08T15:00:00Z",
+            "source": "awc",
+            "temp_f": 55.0,
+        },
+        {
+            "station_code": "KNYC",
+            "observed_at": "2025-01-08T16:00:00Z",
+            "source": "iem",
+            "temp_f": 99.0,
+        },
+    ]
+    _patch_fetchers(research_module, monkeypatch, tmp_path, obs)
+
+    df = research_module.research("KNYC", "2025-01-06", "2025-01-12", source="awc.live")
+    row = df.loc[pd.Timestamp("2025-01-08")]
+    assert row["obs_high_f"] == 55.0
+    assert row["obs_low_f"] == 55.0
+    assert row["obs_count"] == 1
+    assert df.attrs["source"] == "awc.live"
+
+
+def test_unknown_source_raises_before_fetch(tmp_path, monkeypatch) -> None:
+    """An unknown source raises ValueError BEFORE any fetch/cache work."""
+    research_module = _load()
+    monkeypatch.setenv("MOSTLYRIGHT_CACHE_DIR", str(tmp_path))
+
+    called = {"fetch": False}
+
+    def _boom(*a, **kw):
+        called["fetch"] = True
+        raise AssertionError("fetch must not run for an invalid source")
+
+    monkeypatch.setattr(research_module, "_fetch_observations_range", _boom)
+    monkeypatch.setattr(research_module, "_all_caches_warm", lambda *a, **kw: True)
+
+    with pytest.raises(ValueError, match="source must be one of"):
+        research_module.research("KNYC", "2025-01-06", "2025-01-12", source="bogus.source")
+    assert called["fetch"] is False
+
+
+def test_source_and_sources_together_raises_typeerror() -> None:
+    """The existing mutual-exclusion guard is untouched (source= + sources=)."""
+    research_module = _load()
+    with pytest.raises(TypeError, match="mutually exclusive"):
+        research_module.research("KNYC", "2025-01-06", "2025-01-12", source="iem", sources=["iem"])
+
+
+def test_sources_plural_still_not_implemented() -> None:
+    """sources=[...] alone stays NotImplementedError with the rewritten
+    message: points at v0.3 for the multi-source subset only, and no
+    longer redirects source= at research_by_source."""
+    research_module = _load()
+    with pytest.raises(NotImplementedError, match=r"v0\.3") as exc:
+        research_module.research("KNYC", "2025-01-06", "2025-01-12", sources=["iem"])
+    assert "research_by_source" not in str(exc.value)
+
+
+def test_source_none_identical_to_unpinned(tmp_path, monkeypatch) -> None:
+    """source=None is a no-op: byte-identical frame + no attrs["source"]."""
+    research_module = _load()
+    obs = [
+        {
+            "station_code": "KNYC",
+            "observed_at": "2025-01-08T15:00:00Z",
+            "source": "iem",
+            "temp_f": 40.0,
+        },
+        {
+            "station_code": "KNYC",
+            "observed_at": "2025-01-08T16:00:00Z",
+            "source": "awc",
+            "temp_f": 50.0,
+        },
+    ]
+    _patch_fetchers(research_module, monkeypatch, tmp_path, obs)
+
+    df_default = research_module.research("KNYC", "2025-01-06", "2025-01-12")
+    df_none = research_module.research("KNYC", "2025-01-06", "2025-01-12", source=None)
+    pd.testing.assert_frame_equal(df_default, df_none)
+    assert "source" not in df_default.attrs
+
+
+def test_source_pin_noop_when_all_rows_already_match(tmp_path, monkeypatch) -> None:
+    """Pinning "iem" over an all-iem window yields the same obs frame as
+    unpinned — the filter is a pure subset, not a perturbation."""
+    research_module = _load()
+    obs = [
+        {
+            "station_code": "KNYC",
+            "observed_at": "2025-01-08T15:00:00Z",
+            "source": "iem",
+            "temp_f": 40.0,
+        },
+        {
+            "station_code": "KNYC",
+            "observed_at": "2025-01-08T18:00:00Z",
+            "source": "iem",
+            "temp_f": 45.0,
+        },
+    ]
+    _patch_fetchers(research_module, monkeypatch, tmp_path, obs)
+
+    df_pinned = research_module.research("KNYC", "2025-01-06", "2025-01-12", source="iem")
+    df_unpinned = research_module.research("KNYC", "2025-01-06", "2025-01-12")
+    pd.testing.assert_frame_equal(
+        df_pinned[["obs_high_f", "obs_low_f", "obs_count"]],
+        df_unpinned[["obs_high_f", "obs_low_f", "obs_count"]],
+    )
diff --git a/packages/markets/pyproject.toml b/packages/markets/pyproject.toml
index 85a2670c..b56ee0db 100644
--- a/packages/markets/pyproject.toml
+++ b/packages/markets/pyproject.toml
@@ -1,6 +1,6 @@
 [project]
 name = "mostlyrightmd-markets"
-version = "1.11.0"
+version = "1.12.0"
 description = "Prediction-market data for Python — Kalshi NHIGH/NLOW weather-contract resolvers, Polymarket discovery + settlement, and Kalshi + Polymarket trade history. For quants, backtesting, and ML training pipelines. Imports as `mostlyright.markets`."
 readme = "README.md"
 license = "MIT"
diff --git a/packages/weather/pyproject.toml b/packages/weather/pyproject.toml
index 52581f3e..2189c1a8 100644
--- a/packages/weather/pyproject.toml
+++ b/packages/weather/pyproject.toml
@@ -1,6 +1,6 @@
 [project]
 name = "mostlyrightmd-weather"
-version = "1.11.0"
+version = "1.12.0"
 description = "Weather data for Python — live METAR (AWC), ASOS archive (IEM), historical observations (GHCNh), and NWS climate text products (CLI). For quants, ML training pipelines, and weather-bot agents. Direct public-API access, no hosted backend. Imports as `mostlyright.weather`."
 readme = "README.md"
 license = "MIT"
diff --git a/uv.lock b/uv.lock
index 10768b27..44414400 100644
--- a/uv.lock
+++ b/uv.lock
@@ -1628,7 +1628,7 @@ wheels = [
 
 [[package]]
 name = "mostlyrightmd"
-version = "1.11.0"
+version = "1.12.0"
 source = { editable = "packages/core" }
 dependencies = [
     { name = "httpx" },
@@ -1672,7 +1672,7 @@ provides-extras = ["parquet", "research", "polars"]
 
 [[package]]
 name = "mostlyrightmd-markets"
-version = "1.11.0"
+version = "1.12.0"
 source = { editable = "packages/markets" }
 dependencies = [
     { name = "httpx" },
@@ -1726,7 +1726,7 @@ provides-extras = ["parquet", "polymarket", "trades", "earnings", "polars"]
 
 [[package]]
 name = "mostlyrightmd-weather"
-version = "1.11.0"
+version = "1.12.0"
 source = { editable = "packages/weather" }
 dependencies = [
     { name = "filelock" },
@@ -1764,6 +1764,7 @@ polars = [
 satellite = [
     { name = "boto3" },
     { name = "eumdac" },
+    { name = "fsspec" },
     { name = "gcsfs" },
     { name = "h5netcdf" },
     { name = "h5py" },
@@ -1781,7 +1782,8 @@ requires-dist = [
     { name = "eumdac", marker = "extra == 'satellite'", specifier = ">=3.1,<4.0" },
     { name = "faster-whisper", marker = "extra == 'earnings'", specifier = ">=1.0,<2.0" },
     { name = "filelock", specifier = ">=3.12" },
-    { name = "gcsfs", marker = "extra == 'satellite'", specifier = ">=2024.0" },
+    { name = "fsspec", marker = "extra == 'satellite'", specifier = "==2026.6.0" },
+    { name = "gcsfs", marker = "extra == 'satellite'", specifier = "==2026.6.0" },
     { name = "h5netcdf", marker = "extra == 'satellite'", specifier = ">=1.3" },
     { name = "h5py", marker = "extra == 'satellite'", specifier = ">=3.0" },
     { name = "httpx", specifier = ">=0.27" },
@@ -1798,7 +1800,7 @@ requires-dist = [
     { name = "polars", marker = "extra == 'polars'", specifier = ">=1.0,<2.0" },
     { name = "pyarrow", specifier = ">=17.0,<24.0" },
     { name = "pyarrow", marker = "extra == 'polars'", specifier = ">=17.0,<24.0" },
-    { name = "s3fs", marker = "extra == 'satellite'", specifier = ">=2024.0" },
+    { name = "s3fs", marker = "extra == 'satellite'", specifier = "==2026.6.0" },
     { name = "scikit-learn", marker = "extra == 'nwp'", specifier = ">=1.3,<2.0" },
     { name = "tzdata", marker = "sys_platform == 'win32'" },
     { name = "xarray", marker = "extra == 'nwp'", specifier = ">=2024.0" },