diff --git a/enrichers/split_words/manifest.yaml b/enrichers/split_words/manifest.yaml
new file mode 100644
index 0000000..dd26215
--- /dev/null
+++ b/enrichers/split_words/manifest.yaml
@@ -0,0 +1,10 @@
+id: split_words
+name: "Split text into words"
+description: "Tokeniza el texto del nodo (entities.notes con fallback a name) en palabras y crea un nodo Word por cada token. Pensado para probar grouping con volumen alto: cualquier parrafo decente supera el umbral de 50 trivialmente."
+applies_to: [text, Text]
+emits: [Word]
+relations: [WORD_OF]
+params:
+  - { name: max_words, type: int, default: 500 }
+  - { name: min_length, type: int, default: 3, description: "ignora tokens con menos de N caracteres (filtra ruido tipo 'a', 'el', 'de')" }
+  - { name: dedupe, type: bool, default: true, description: "si true, una palabra repetida produce un solo nodo" }
diff --git a/enrichers/split_words/run.py b/enrichers/split_words/run.py
new file mode 100644
index 0000000..9cdaf0a
--- /dev/null
+++ b/enrichers/split_words/run.py
@@ -0,0 +1,325 @@
+#!/usr/bin/env python3
+"""Enricher split_words — tokeniza texto en palabras (regex puro, offline).
+
+Pensado para probar el grouping de issue 0035 con volumen alto: cualquier
+parrafo decente supera el umbral de 50 trivialmente, asi se ve un Group
+cuadrado por flujo.
+
+Wire protocol estandar (issue 0026):
+  - stdin:  JSON con node_id, node_name, metadata, ops_db_path, app_dir,
+            cache_dir, registry_root, params.
+  - stderr: lineas `PROGRESS:<float> <stage>` para feedback de UI.
+  - stdout: una linea JSON al final con resumen.
+  - exit code 0 = ok, !=0 = error.
+
+Lectura del texto (igual que split_sentences):
+  1. `entities.notes`  (panel Note del Inspector).
+  2. node_name         (fallback minimo).
+
+Tokenizacion: split por whitespace + puntuacion, lowercased. Filtra
+tokens con `len < min_length` para evitar ruido (a, el, de, y, o, ...).
+Por defecto deduplica para devolver vocabulario unico (mas util para
+explorar contenido); con `dedupe=false` cada ocurrencia es un nodo
+(util para volumen / stress).
+
+Grouping (issue 0035c): mismo patron que split_sentences y web_search.
+"""
+from __future__ import annotations
+
+import json
+import os
+import re
+import sqlite3
+import sys
+import time
+import uuid
+from datetime import datetime, timezone
+
+
+DEFAULT_GROUP_THRESHOLD = 50
+GROUP_PREVIEW_K         = 10
+
+# Tokenizer: secuencias de letras (incluye acentos espanyoles + apostrofo
+# interno tipo "don't"). Mas robusto que split por espacios para texto
+# real con puntuacion adyacente. Numeros se ignoran — pensado para
+# contenido natural, no datos.
+_TOKEN_RE = re.compile(r"[A-Za-zÁÉÍÓÚÜÑáéíóúüñ][A-Za-zÁÉÍÓÚÜÑáéíóúüñ'-]*")
+
+
+def progress(p: float, stage: str = "") -> None:
+    sys.stderr.write(f"PROGRESS:{p:.2f} {stage}\n")
+    sys.stderr.flush()
+
+
+def log(msg: str) -> None:
+    sys.stderr.write(f"{msg}\n")
+    sys.stderr.flush()
+
+
+def now_iso() -> str:
+    return datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")
+
+
+def now_ms() -> int:
+    return int(time.time() * 1000)
+
+
+def has_group_id_column(conn: sqlite3.Connection) -> bool:
+    try:
+        cur = conn.execute("PRAGMA table_info(entities)")
+        for row in cur:
+            if row[1] == "group_id":
+                return True
+    except sqlite3.Error:
+        pass
+    return False
+
+
+def read_text(ops_db_path: str, node_id: str, node_name: str) -> str:
+    notes = ""
+    try:
+        c = sqlite3.connect(ops_db_path)
+        try:
+            row = c.execute(
+                "SELECT notes FROM entities WHERE id=?", (node_id,)
+            ).fetchone()
+            if row and isinstance(row[0], str):
+                notes = row[0]
+        finally:
+            c.close()
+    except sqlite3.Error:
+        notes = ""
+    if notes and notes.strip():
+        return notes.strip()
+    return (node_name or "").strip()
+
+
+def tokenize(text: str, *, min_length: int, dedupe: bool) -> list[str]:
+    """Devuelve lista de tokens (lower) filtrados por min_length.
+
+    Si `dedupe=True`, conserva solo la primera aparicion (preserva orden).
+    """
+    seen: set[str] = set()
+    out: list[str] = []
+    for m in _TOKEN_RE.finditer(text):
+        tok = m.group(0).lower()
+        if len(tok) < min_length:
+            continue
+        if dedupe:
+            if tok in seen:
+                continue
+            seen.add(tok)
+        out.append(tok)
+    return out
+
+
+def insert_word(conn: sqlite3.Connection, *, word: str, rank: int,
+                  batch_id: str, group_id: str | None,
+                  has_group_col: bool) -> str:
+    ts = now_iso()
+    new_id = f"Word_{now_ms()}_{rank}"
+    meta = {
+        "word":     word,
+        "rank":     rank,
+        "batch_id": batch_id,
+    }
+    meta_json = json.dumps(meta, ensure_ascii=False)
+    if has_group_col:
+        conn.execute(
+            "INSERT INTO entities (id, name, type_ref, source, metadata, "
+            " group_id, created_at, updated_at) "
+            "VALUES (?, ?, 'Word', 'enricher:split_words', ?, ?, ?, ?)",
+            (new_id, word, meta_json, group_id, ts, ts),
+        )
+    else:
+        conn.execute(
+            "INSERT INTO entities (id, name, type_ref, source, metadata, "
+            " created_at, updated_at) "
+            "VALUES (?, ?, 'Word', 'enricher:split_words', ?, ?, ?)",
+            (new_id, word, meta_json, ts, ts),
+        )
+    return new_id
+
+
+def insert_group_entity(conn: sqlite3.Connection, *, source_node_id: str,
+                         source_node_name: str, count: int,
+                         batch_id: str) -> str:
+    ts = now_iso()
+    new_id = f"Group_{now_ms()}_{abs(hash(source_node_id + batch_id)) % 100000}"
+    name = f"split_words: {source_node_name} ({count})"
+    meta = {
+        "enricher":       "split_words",
+        "count":          count,
+        "batch_id":       batch_id,
+        "source_node_id": source_node_id,
+    }
+    meta_json = json.dumps(meta, ensure_ascii=False)
+    conn.execute(
+        "INSERT INTO entities (id, name, type_ref, source, metadata, "
+        " created_at, updated_at) "
+        "VALUES (?, ?, 'Group', 'enricher:split_words', ?, ?, ?)",
+        (new_id, name, meta_json, ts, ts),
+    )
+    return new_id
+
+
+_REL_COUNTER = 0
+
+
+def insert_relation(conn: sqlite3.Connection, from_id: str, to_id: str,
+                    name: str) -> bool:
+    global _REL_COUNTER
+    cur = conn.execute(
+        "SELECT 1 FROM relations WHERE from_entity=? AND to_entity=? "
+        "AND name=? LIMIT 1",
+        (from_id, to_id, name),
+    )
+    if cur.fetchone():
+        return False
+    ts = now_iso()
+    _REL_COUNTER += 1
+    rel_id = f"rel_{now_ms()}_{_REL_COUNTER}_{name.lower()}"
+    conn.execute(
+        "INSERT INTO relations (id, name, from_entity, to_entity, "
+        " created_at, updated_at) VALUES (?, ?, ?, ?, ?, ?)",
+        (rel_id, name, from_id, to_id, ts, ts),
+    )
+    return True
+
+
+def main() -> int:
+    raw = sys.stdin.read()
+    try:
+        ctx = json.loads(raw)
+    except Exception as e:
+        log(f"stdin not valid JSON: {e}")
+        return 2
+
+    node_id     = ctx.get("node_id") or ""
+    node_name   = (ctx.get("node_name") or "").strip()
+    ops_db_path = ctx.get("ops_db_path") or ""
+    params      = ctx.get("params") or {}
+    max_words   = int(params.get("max_words", 500))
+    min_length  = int(params.get("min_length", 3))
+    dedupe_raw  = params.get("dedupe", True)
+    if isinstance(dedupe_raw, str):
+        dedupe = dedupe_raw.lower() not in ("false", "0", "no", "")
+    else:
+        dedupe = bool(dedupe_raw)
+
+    if not node_id or not ops_db_path:
+        log("missing node_id / ops_db_path")
+        return 2
+
+    ops_db_path = ops_db_path.replace("\\", "/")
+    app_dir_raw = (ctx.get("app_dir") or "").replace("\\", "/")
+    if not os.path.isabs(ops_db_path):
+        if app_dir_raw and os.path.isdir(app_dir_raw):
+            cand = os.path.normpath(os.path.join(app_dir_raw, ops_db_path))
+            if os.path.exists(cand):
+                ops_db_path = cand
+        if not os.path.isabs(ops_db_path):
+            ops_db_path = os.path.abspath(ops_db_path)
+
+    if not os.path.exists(ops_db_path):
+        log(f"ops_db_path no existe: {ops_db_path}")
+        print(json.dumps({"error": "ops_db not found",
+                          "ops_db_path": ops_db_path,
+                          "entities_added": 0, "relations_added": 0}))
+        return 7
+
+    progress(0.10, "reading")
+    text = read_text(ops_db_path, node_id, node_name)
+    # min_length aplica a tokens, no al texto entrante. Para texto entrante
+    # exigimos algo razonable: al menos un token posible.
+    if len(text.strip()) < min_length:
+        msg = (f"texto demasiado corto ({len(text)} chars). Escribe el "
+               f"contenido en el panel Note del nodo (doble click) o "
+               f"pon un name mas largo.")
+        log(msg)
+        print(json.dumps({"error": msg, "entities_added": 0,
+                          "relations_added": 0}))
+        return 2
+
+    progress(0.30, "tokenizing")
+    words = tokenize(text, min_length=min_length, dedupe=dedupe)
+    if max_words > 0:
+        words = words[:max_words]
+
+    if not words:
+        msg = (f"sin tokens tras filtrar (texto de {len(text)} chars, "
+               f"min_length={min_length}, dedupe={dedupe})")
+        log(msg)
+        print(json.dumps({"error": msg, "entities_added": 0,
+                          "relations_added": 0}))
+        return 2
+
+    progress(0.55, "writing")
+    conn = sqlite3.connect(ops_db_path)
+    conn.execute("PRAGMA foreign_keys=OFF")
+    entities_added = 0
+    relations_added = 0
+    group_id: str | None = None
+    batch_id = uuid.uuid4().hex
+    try:
+        has_group_col = has_group_id_column(conn)
+        n_total = len(words)
+        threshold = DEFAULT_GROUP_THRESHOLD
+
+        if n_total >= threshold and has_group_col:
+            group_id = insert_group_entity(
+                conn,
+                source_node_id=node_id,
+                source_node_name=node_name or "(text)",
+                count=n_total,
+                batch_id=batch_id,
+            )
+            entities_added += 1
+            if insert_relation(conn, group_id, node_id, "WORD_OF"):
+                relations_added += 1
+            preview = words[:GROUP_PREVIEW_K]
+            grouped = words[GROUP_PREVIEW_K:]
+        else:
+            preview = words
+            grouped = []
+
+        for i, w in enumerate(preview):
+            wid = insert_word(
+                conn, word=w, rank=i + 1, batch_id=batch_id,
+                group_id=None, has_group_col=has_group_col,
+            )
+            entities_added += 1
+            if insert_relation(conn, wid, node_id, "WORD_OF"):
+                relations_added += 1
+
+        for j, w in enumerate(grouped):
+            rank = GROUP_PREVIEW_K + j + 1
+            wid = insert_word(
+                conn, word=w, rank=rank, batch_id=batch_id,
+                group_id=group_id, has_group_col=has_group_col,
+            )
+            entities_added += 1
+            if insert_relation(conn, wid, node_id, "WORD_OF"):
+                relations_added += 1
+            if grouped and j % 50 == 0:
+                progress(0.55 + 0.40 * (j / max(1, len(grouped))), "writing")
+
+        conn.commit()
+    finally:
+        conn.close()
+
+    progress(1.0, "done")
+    print(json.dumps({
+        "words":           len(words),
+        "entities_added":  entities_added,
+        "relations_added": relations_added,
+        "batch_id":        batch_id,
+        "group_id":        group_id or "",
+        "grouped":         bool(group_id),
+        "deduped":         dedupe,
+    }, ensure_ascii=False))
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/examples/types.yaml b/examples/types.yaml
index c326081..3d66260 100644
--- a/examples/types.yaml
+++ b/examples/types.yaml
@@ -134,6 +134,18 @@ entities:
       - { name: rank,     type: int }
       - { name: batch_id, type: string }
 
+  # Word — token emitido por split_words. Pensado para probar grouping
+  # con volumen alto: cualquier parrafo decente supera el umbral de 50
+  # tokens trivialmente. Color amarillo para distinguir de Sentence.
+  - name: Word
+    color: "#FCD34D"
+    icon: ti-letter-w
+    principal_field: word
+    fields:
+      - { name: word,     type: string, required: true }
+      - { name: rank,     type: int }
+      - { name: batch_id, type: string }
+
   # Nodo grupo — cuadrado (regla de forma). Issue 0035: contenedor para
   # agrupar resultados de enrichers cuando exceden el umbral. Los hijos
   # son entidades reales con `group_id` apuntando al Group.
diff --git a/tests/test_split_words.py b/tests/test_split_words.py
new file mode 100644
index 0000000..f883e00
--- /dev/null
+++ b/tests/test_split_words.py
@@ -0,0 +1,154 @@
+"""Tests del enricher split_words (offline, regex puro).
+
+Mismo patron que test_split_sentences.py: nodo text con `notes` largo,
+verificamos tokens, fallback a name, threshold/grouping y dedupe.
+"""
+from __future__ import annotations
+
+from conftest import (
+    base_ctx, list_entities, list_relations, make_node, run_enricher,
+)
+
+
+# Texto con suficientes palabras unicas para superar threshold (50)
+# si dedupe=true: cuento ~85 unicas a ojo. Sin dedupe (count=true)
+# todas las ocurrencias dan ~140+ tokens.
+LONG_TEXT = (
+    "Las estrellas brillan suavemente sobre el horizonte mientras "
+    "la marea retrocede dejando huellas mojadas en la arena fina. "
+    "Caminamos lentamente conversando sobre proyectos antiguos, "
+    "ideas frescas, libros leidos durante el invierno pasado, "
+    "viajes pendientes hacia tierras lejanas con culturas vibrantes. "
+    "Recordamos infancias compartidas, amigos perdidos, victorias "
+    "modestas, fracasos instructivos. Cada palabra dibuja un mapa "
+    "diferente del territorio interno que habitamos. Los nombres de "
+    "ciudades antiguas resuenan: Estambul, Marrakech, Kioto, Lisboa, "
+    "Praga, Budapest, Cuzco, Cartagena. Tambien tecnologia: servidores, "
+    "bases datos, redes neuronales, modelos linguisticos, sistemas "
+    "distribuidos, criptografia moderna. La conversacion fluye sin "
+    "esfuerzo aparente entre dominios completamente distintos."
+)
+
+
+def test_split_words_creates_word_nodes(ops_db, app_dir, registry_root):
+    """Texto corto < threshold genera Words sueltos sin Group."""
+    short_text = "uno dos tres cuatro cinco seis siete ocho nueve diez."
+    make_node(ops_db, node_id="t1", name="short",
+              type_ref="text", notes=short_text)
+    ctx = base_ctx(ops_db=ops_db, app_dir=app_dir, registry_root=registry_root,
+                   node_id="t1", node_name="short", node_type="text")
+    rc, out, err = run_enricher("split_words", ctx)
+    assert rc == 0, err
+    assert out["grouped"] is False
+    assert out["entities_added"] == out["words"]
+    assert out["words"] >= 5  # filtra <3 chars: dos, tres, cuatro, ...
+
+    words = list_entities(ops_db, type_ref="Word")
+    assert len(words) == out["words"]
+
+
+def test_split_words_above_threshold_creates_group(ops_db, app_dir,
+                                                     registry_root):
+    """Texto largo (≥50 tokens unicos) → Group + 10 sueltos + resto."""
+    make_node(ops_db, node_id="t1", name="largo",
+              type_ref="text", notes=LONG_TEXT)
+    ctx = base_ctx(ops_db=ops_db, app_dir=app_dir, registry_root=registry_root,
+                   node_id="t1", node_name="largo", node_type="text")
+    rc, out, err = run_enricher("split_words", ctx, timeout=60)
+    assert rc == 0, err
+    assert out["words"] >= 50
+    assert out["grouped"] is True
+    assert out["group_id"]
+    # Group + words = entities_added.
+    assert out["entities_added"] == out["words"] + 1
+
+    # Hay 10 Words sueltos (group_id NULL) y resto agrupados.
+    import sqlite3
+    cn = sqlite3.connect(ops_db)
+    n_loose = cn.execute(
+        "SELECT count(*) FROM entities WHERE type_ref='Word' "
+        "AND group_id IS NULL"
+    ).fetchone()[0]
+    n_grouped = cn.execute(
+        "SELECT count(*) FROM entities WHERE type_ref='Word' "
+        "AND group_id = ?", (out["group_id"],)
+    ).fetchone()[0]
+    cn.close()
+    assert n_loose == 10
+    assert n_loose + n_grouped == out["words"]
+
+
+def test_split_words_dedupe_default_true(ops_db, app_dir, registry_root):
+    """Por defecto dedupe=true: 'casa casa casa' produce 1 Word."""
+    make_node(ops_db, node_id="t1", name="dup",
+              type_ref="text", notes="casa casa casa perro perro gato.")
+    ctx = base_ctx(ops_db=ops_db, app_dir=app_dir, registry_root=registry_root,
+                   node_id="t1", node_name="dup", node_type="text")
+    rc, out, err = run_enricher("split_words", ctx)
+    assert rc == 0, err
+    assert out["deduped"] is True
+    # casa, perro, gato (todas ≥3 chars).
+    assert out["words"] == 3
+
+
+def test_split_words_dedupe_false(ops_db, app_dir, registry_root):
+    """Con dedupe=false cada ocurrencia es un nodo Word."""
+    make_node(ops_db, node_id="t1", name="dup",
+              type_ref="text", notes="casa casa casa perro perro gato.")
+    ctx = base_ctx(ops_db=ops_db, app_dir=app_dir, registry_root=registry_root,
+                   node_id="t1", node_name="dup", node_type="text",
+                   params={"dedupe": False})
+    rc, out, err = run_enricher("split_words", ctx)
+    assert rc == 0, err
+    assert out["deduped"] is False
+    # 3 casa + 2 perro + 1 gato = 6 tokens.
+    assert out["words"] == 6
+
+
+def test_split_words_min_length_filters(ops_db, app_dir, registry_root):
+    """min_length filtra tokens cortos. Default 3."""
+    make_node(ops_db, node_id="t1", name="cortos",
+              type_ref="text",
+              notes="a el de la y o un casa perro elefante.")
+    ctx = base_ctx(ops_db=ops_db, app_dir=app_dir, registry_root=registry_root,
+                   node_id="t1", node_name="cortos", node_type="text")
+    rc, out, err = run_enricher("split_words", ctx)
+    assert rc == 0, err
+    # >=3 chars: casa, perro, elefante. (un=2, de=2, la=2 quedan fuera).
+    assert out["words"] == 3
+
+
+def test_split_words_uses_notes_priority(ops_db, app_dir, registry_root):
+    """Lee `entities.notes` por encima de node_name."""
+    make_node(ops_db, node_id="t1", name="ignorame",
+              type_ref="text",
+              notes="estos cinco tokens deberian ganar.")
+    ctx = base_ctx(ops_db=ops_db, app_dir=app_dir, registry_root=registry_root,
+                   node_id="t1", node_name="ignorame", node_type="text")
+    rc, out, err = run_enricher("split_words", ctx)
+    assert rc == 0, err
+    # estos, cinco, tokens, deberian, ganar (todos ≥3).
+    assert out["words"] == 5
+
+
+def test_split_words_no_text_fails(ops_db, app_dir, registry_root):
+    """Sin notes y name corto → exit 2."""
+    make_node(ops_db, node_id="t1", name="x", type_ref="text", metadata={})
+    ctx = base_ctx(ops_db=ops_db, app_dir=app_dir, registry_root=registry_root,
+                   node_id="t1", node_name="x", node_type="text")
+    rc, out, err = run_enricher("split_words", ctx)
+    assert rc == 2
+    assert out is not None
+    assert "demasiado corto" in (out.get("error") or "")
+
+
+def test_split_words_max_words_truncates(ops_db, app_dir, registry_root):
+    """max_words limita el output."""
+    make_node(ops_db, node_id="t1", name="largo",
+              type_ref="text", notes=LONG_TEXT)
+    ctx = base_ctx(ops_db=ops_db, app_dir=app_dir, registry_root=registry_root,
+                   node_id="t1", node_name="largo", node_type="text",
+                   params={"max_words": 12})
+    rc, out, err = run_enricher("split_words", ctx)
+    assert rc == 0, err
+    assert out["words"] == 12