diff --git a/backend/app.py b/backend/app.py
index 155246a..1d18327 100644
--- a/backend/app.py
+++ b/backend/app.py
@@ -271,5 +271,8 @@ app.include_router(review_router)
 from .publish import router as publish_router
 app.include_router(publish_router)
 
+from .inference import router as inference_router
+app.include_router(inference_router)
+
 # expose for tests
 __all__ = ["app", "__version__"]
diff --git a/backend/inference.py b/backend/inference.py
new file mode 100644
index 0000000..0022ec0
--- /dev/null
+++ b/backend/inference.py
@@ -0,0 +1,364 @@
+"""Inference (S7): prompt revisions (REST), Call 1 vision classification, Call 2
+content prose, test-inference, inference-off server description, inference_log.
+Cites METADATA-35..115, SETTINGS-30/45/90.
+"""
+
+import base64
+import hashlib
+import json
+import time
+import urllib.error
+import urllib.request
+from typing import Any
+
+from fastapi import APIRouter
+from fastapi.responses import JSONResponse
+
+from .db import get_connection
+from .crypto import decrypt
+
+router = APIRouter(prefix="/api/v1")
+
+CALL1_SEED = (
+    "Classify the slab's species, wood categories, edge type, figure, and grade from the "
+    "supplied photos, choosing only from the supplied synced taxonomy lists, and return the "
+    "result fields as JSON: species_id, wood_category_ids, edge_type_term_id, figure_term_ids, "
+    "grade_term_id, character, inclusions, voids, checks. Report a confidence (0-1) for each "
+    "classified field. Map free-text observations to the supplied feature-tag list only. "
+    "Invent no category or term."
+)
+CALL2_SEED = (
+    "Write the long product description for the slab from the supplied photos and listing "
+    "facts. The name and short description are already built, so do not repeat them. Use the "
+    "brand voice and geo context given. The dimensions in the fact block are exact: use those "
+    "strings and invent no numbers. Write prose a premium lumber buyer would read."
+)
+_SEEDS = {"call1": CALL1_SEED, "call2": CALL2_SEED}
+_THRESHOLD = 0.7
+
+
+def _error(code: str, message: str, detail: Any = None) -> dict:
+    return {"error": {"code": code, "message": message, "detail": detail}}
+
+
+def _now() -> str:
+    from datetime import datetime, timezone
+
+    return datetime.now(timezone.utc).replace(microsecond=0, tzinfo=None).isoformat()
+
+
+def _hash(text: str) -> str:
+    return hashlib.sha256(text.encode("utf-8")).hexdigest()
+
+
+def _seed_prompts(conn) -> None:
+    for name, seed in _SEEDS.items():
+        if conn.execute("SELECT 1 FROM prompts WHERE name=?", (name,)).fetchone():
+            continue
+        cur = conn.execute(
+            "INSERT INTO prompts(name, active_revision_id, created_at, updated_at) VALUES (?,?,?,?)",
+            (name, None, _now(), _now()),
+        )
+        pid = cur.lastrowid
+        cur2 = conn.execute(
+            "INSERT INTO prompt_revisions(prompt_id, content, content_hash, created_at, created_by) VALUES (?,?,?,?,?)",
+            (pid, seed, _hash(seed), _now(), "seed"),
+        )
+        conn.execute("UPDATE prompts SET active_revision_id=? WHERE id=?", (cur2.lastrowid, pid))
+
+
+def _active_prompt(conn, call: str) -> tuple[int, str] | None:
+    row = conn.execute(
+        "SELECT p.id, pr.id, pr.content FROM prompts p LEFT JOIN prompt_revisions pr ON p.active_revision_id=pr.id WHERE p.name=?",
+        (call,),
+    ).fetchone()
+    if row is None or row[2] is None:
+        return None
+    return row[1], row[2]
+
+
+# ---------- Prompt revision REST (METADATA-106) ----------
+@router.post("/prompts/{call}/revisions", status_code=201)
+async def append_revision(call: str, payload: dict[str, Any]) -> dict:
+    if call not in _SEEDS:
+        return JSONResponse(status_code=404, content=_error("not_found", "unknown call"))
+    content = payload.get("content")
+    if not content:
+        return JSONResponse(status_code=400, content=_error("validation", "content required"))
+    conn = get_connection()
+    _seed_prompts(conn)
+    prow = conn.execute("SELECT id FROM prompts WHERE name=?", (call,)).fetchone()
+    cur = conn.execute(
+        "INSERT INTO prompt_revisions(prompt_id, content, content_hash, created_at, created_by) VALUES (?,?,?,?,?)",
+        (prow[0], content, _hash(content), _now(), "api"),
+    )
+    conn.execute("UPDATE prompts SET active_revision_id=?, updated_at=? WHERE id=?", (cur.lastrowid, _now(), prow[0]))
+    conn.commit()
+    return {"id": cur.lastrowid, "prompt": call, "content_hash": _hash(content)}
+
+
+@router.post("/prompts/{call}/activate")
+async def activate_revision(call: str, payload: dict[str, Any]) -> dict:
+    if call not in _SEEDS:
+        return JSONResponse(status_code=404, content=_error("not_found", "unknown call"))
+    rev_id = payload.get("revision_id")
+    conn = get_connection()
+    _seed_prompts(conn)
+    prow = conn.execute("SELECT id FROM prompts WHERE name=?", (call,)).fetchone()
+    row = conn.execute(
+        "SELECT id FROM prompt_revisions WHERE id=? AND prompt_id=?", (rev_id, prow[0])
+    ).fetchone()
+    if row is None:
+        return JSONResponse(status_code=404, content=_error("not_found", "revision not found"))
+    conn.execute("UPDATE prompts SET active_revision_id=?, updated_at=? WHERE id=?", (rev_id, _now(), prow[0]))
+    conn.commit()
+    return {"ok": True, "active_revision_id": rev_id}
+
+
+@router.get("/prompts/{call}")
+async def get_prompt(call: str) -> dict:
+    if call not in _SEEDS:
+        return JSONResponse(status_code=404, content=_error("not_found", "unknown call"))
+    conn = get_connection()
+    _seed_prompts(conn)
+    prow = conn.execute("SELECT * FROM prompts WHERE name=?", (call,)).fetchone()
+    revs = conn.execute(
+        "SELECT id, content_hash, created_at, created_by FROM prompt_revisions WHERE prompt_id=? ORDER BY id",
+        (prow[0],),
+    ).fetchall()
+    active = conn.execute(
+        "SELECT content FROM prompt_revisions WHERE id=?", (prow["active_revision_id"],)
+    ).fetchone()
+    return {
+        "name": call,
+        "active_revision_id": prow["active_revision_id"],
+        "active_content": active[0] if active else None,
+        "revisions": [dict(r) for r in revs],
+    }
+
+
+# ---------- Inference client (OpenAI-compatible, stdlib) ----------
+class InferenceClient:
+    def __init__(self, base_url: str, api_key: str, model: str, timeout: float = 60.0):
+        self.base_url = base_url.rstrip("/")
+        self.api_key = api_key
+        self.model = model
+        self.timeout = timeout
+
+    def chat(self, messages: list[dict]) -> str:
+        url = f"{self.base_url}/chat/completions"
+        body = json.dumps({"model": self.model, "messages": messages}).encode("utf-8")
+        req = urllib.request.Request(
+            url, data=body,
+            headers={"Authorization": f"Bearer {self.api_key}", "Content-Type": "application/json"},
+            method="POST",
+        )
+        try:
+            with urllib.request.urlopen(req, timeout=self.timeout) as resp:
+                data = json.loads(resp.read().decode("utf-8"))
+        except urllib.error.HTTPError as e:
+            raise InferenceError(f"inference HTTP {e.code}") from e
+        except urllib.error.URLError as e:
+            raise InferenceConnectionError(f"inference unreachable: {e.reason}") from e
+        return data["choices"][0]["message"]["content"]
+
+
+class InferenceError(Exception):
+    pass
+
+
+class InferenceConnectionError(InferenceError):
+    pass
+
+
+def _inference_creds(conn) -> tuple[str, str, str]:
+    b = conn.execute("SELECT value_enc FROM settings WHERE key='inference_base_url'").fetchone()
+    k = conn.execute("SELECT value_enc FROM settings WHERE key='inference_api_key'").fetchone()
+    m = conn.execute("SELECT value FROM settings WHERE key='inference_model'").fetchone()
+    return (decrypt(b[0]) if b and b[0] else ""), (decrypt(k[0]) if k and k[0] else ""), (m[0] if m and m[0] else "")
+
+
+def _enabled(conn, key: str) -> bool:
+    row = conn.execute("SELECT value FROM settings WHERE key=?", (key,)).fetchone()
+    return bool(row and row[0] in ("true", "True", "1"))
+
+
+@router.post("/slabs/{slab_id}/inference/call1")
+async def call1(slab_id: str) -> dict:
+    conn = get_connection()
+    row = conn.execute("SELECT * FROM slabs WHERE id=? AND deleted_at IS NULL", (slab_id,)).fetchone()
+    if row is None:
+        return JSONResponse(status_code=404, content=_error("not_found", "Slab not found"))
+    if not _enabled(conn, "inference_enabled"):
+        return JSONResponse(status_code=409, content=_error("inference_disabled", "Inference is off"))
+    base, key, model = _inference_creds(conn)
+    if not base or not key:
+        return JSONResponse(status_code=409, content=_error("inference_unconfigured", "Inference is not configured"))
+    client = InferenceClient(base, key, model)
+    act = _active_prompt(conn, "call1")
+    prompt_text = act[1] if act else CALL1_SEED
+    # Build vision message with processed photos (1024px) + taxonomy lists.
+    photos = conn.execute(
+        "SELECT processed_blob FROM slab_photos WHERE slab_id=? AND processed_blob IS NOT NULL ORDER BY seq LIMIT 8",
+        (slab_id,),
+    ).fetchall()
+    image_parts = []
+    for p in photos:
+        image_parts.append({
+            "type": "image_url",
+            "image_url": {"url": f"data:image/png;base64,{base64.b64encode(p[0]).decode()}"},
+        })
+    tax = _taxonomy_lists(conn)
+    text = (f"{prompt_text}\n\nSYNCED TAXONOMY (only legal choices):\n{json.dumps(tax)}\n\n"
+            f"SKU={row['sku']} length_in={row['length_in']} thickness_in={row['thickness_in']}")
+    messages = [{"role": "user", "content": image_parts + [{"type": "text", "text": text}]}]
+    t0 = time.time()
+    try:
+        raw = client.chat(messages)
+    except InferenceConnectionError as e:
+        return JSONResponse(status_code=503, content=_error("inference_unreachable", str(e), {"recoverable": True}))
+    except InferenceError as e:
+        return JSONResponse(status_code=502, content=_error("inference_failed", str(e), {"recoverable": True}))
+    latency = int((time.time() - t0) * 1000)
+    parsed = _parse_json(raw)
+    _log_call(conn, slab_id, 1, act[0] if act else None, model, messages, raw, latency)
+    # Apply confidence gate (METADATA-45).
+    suggestions = _gate(parsed)
+    return {"ok": True, "suggestions": suggestions}
+
+
+def _taxonomy_lists(conn) -> dict:
+    out = {"species_leaves": [], "edge_terms": [], "figure_terms": [], "grade_terms": [],
+           "thickness_terms": [], "moisture_terms": [], "feat_tags": []}
+    for r in conn.execute("SELECT id,name FROM woo_categories WHERE parent_id IS NULL").fetchall():
+        out["species_leaves"].append({"id": r[0], "name": r[1]})
+    for attr, key in (("Edge Type", "edge_terms"), ("Figure", "figure_terms"), ("Grade", "grade_terms"),
+                      ("Thickness", "thickness_terms"), ("Moisture", "moisture_terms")):
+        rows = conn.execute(
+            "SELECT t.id,t.name FROM woo_attribute_terms t JOIN woo_attributes a ON t.attribute_id=a.id WHERE a.name=?",
+            (attr,),
+        ).fetchall()
+        out[key] = [{"id": r[0], "name": r[1]} for r in rows]
+    for r in conn.execute("SELECT id,name FROM woo_tags WHERE slug LIKE 'feat-%'").fetchall():
+        out["feat_tags"].append({"id": r[0], "name": r[1]})
+    return out
+
+
+def _parse_json(raw: str) -> dict:
+    try:
+        s = raw.strip()
+        if s.startswith("```"):
+            s = s.split("```")[1]
+            if s.startswith("json"):
+                s = s[4:]
+        return json.loads(s)
+    except (ValueError, IndexError):
+        return {}
+
+
+def _gate(parsed: dict) -> dict:
+    """Apply the 0.7 confidence gate; below-threshold fields are omitted (METADATA-45)."""
+    def keep(field: str, conf_key: str) -> Any:
+        conf = parsed.get(conf_key)
+        if isinstance(conf, (int, float)) and conf >= _THRESHOLD and field in parsed:
+            return parsed[field]
+        return None
+
+    out: dict[str, Any] = {}
+    if v := keep("species_id", "species_confidence"):
+        out["species_id"] = v
+    if v := keep("wood_category_ids", "wood_category_confidence"):
+        out["wood_category_ids"] = v
+    if v := keep("edge_type_term_id", "edge_confidence"):
+        out["edge_type_term_id"] = v
+    if v := keep("figure_term_ids", "figure_confidence"):
+        out["figure_term_ids"] = v
+    if v := keep("grade_term_id", "grade_confidence"):
+        out["grade_term_id"] = v
+    for k in ("character", "inclusions", "voids", "checks"):
+        if parsed.get(k):
+            out[k] = parsed[k]
+    return out
+
+
+@router.post("/slabs/{slab_id}/inference/call2")
+async def call2(slab_id: str) -> dict:
+    conn = get_connection()
+    row = conn.execute("SELECT * FROM slabs WHERE id=? AND deleted_at IS NULL", (slab_id,)).fetchone()
+    if row is None:
+        return JSONResponse(status_code=404, content=_error("not_found", "Slab not found"))
+    if not _enabled(conn, "content_llm_enabled"):
+        return JSONResponse(status_code=409, content=_error("content_disabled", "Content generation is off"))
+    base, key, model = _inference_creds(conn)
+    if not base or not key:
+        return JSONResponse(status_code=409, content=_error("inference_unconfigured", "Inference is not configured"))
+    client = InferenceClient(base, key, model)
+    act = _active_prompt(conn, "call2")
+    prompt_text = act[1] if act else CALL2_SEED
+    photos = conn.execute(
+        "SELECT processed_blob FROM slab_photos WHERE slab_id=? AND processed_blob IS NOT NULL ORDER BY seq LIMIT 8",
+        (slab_id,),
+    ).fetchall()
+    image_parts = [
+        {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{base64.b64encode(p[0]).decode()}"}}
+        for p in photos
+    ]
+    brand = conn.execute("SELECT value FROM settings WHERE key='brand_voice'").fetchone()
+    geo = conn.execute("SELECT value FROM settings WHERE key='geo_context'").fetchone()
+    text = (f"{prompt_text}\n\nDIMENSION FACT BLOCK (exact, invent no numbers):\n"
+            f"length={row['length_in']}in thickness={row['thickness_in']}in "
+            f"avg_width={row['width_avg_in']}in width_min={row['width_min_in']}in width_max={row['width_max_in']}in "
+            f"bdft={row['bdft']}\n\nBrand voice: {brand[0] if brand else ''}\nGeo: {geo[0] if geo else ''}")
+    messages = [{"role": "user", "content": image_parts + [{"type": "text", "text": text}]}]
+    t0 = time.time()
+    try:
+        raw = client.chat(messages)
+    except InferenceConnectionError as e:
+        return JSONResponse(status_code=503, content=_error("inference_unreachable", str(e), {"recoverable": True}))
+    except InferenceError as e:
+        return JSONResponse(status_code=502, content=_error("inference_failed", str(e), {"recoverable": True}))
+    latency = int((time.time() - t0) * 1000)
+    _log_call(conn, slab_id, 2, act[0] if act else None, model, messages, raw, latency)
+    # Store as description, mark content_source=llm.
+    conn.execute("UPDATE slabs SET description=?, content_source='llm', updated_at=? WHERE id=?",
+                 (raw.strip(), _now(), slab_id))
+    conn.commit()
+    return {"ok": True, "description": raw.strip()}
+
+
+def server_description(slab: dict) -> str:
+    """Inference-off long description (METADATA-70)."""
+    return (
+        f"A {slab.get('title', 'slab')} from the mill, measured at "
+        f"{slab.get('length_in')}in long, {slab.get('width_avg_in')}in average width, and "
+        f"{slab.get('thickness_in')}in thick ({slab.get('bdft')} bdft). "
+        "Ready to ship."
+    )
+
+
+def _log_call(conn, slab_id: str, call: int, rev_id: int | None, model: str,
+              request: list, response: str, latency_ms: int) -> None:
+    conn.execute(
+        "INSERT INTO inference_log(slab_id, call, prompt_revision_id, model, request_payload, response_payload, latency_ms, created_at) "
+        "VALUES (?,?,?,?,?,?,?,?)",
+        (slab_id, call, rev_id, model, json.dumps(request)[:200000], response, latency_ms, _now()),
+    )
+    conn.commit()
+
+
+@router.post("/admin/test-inference")
+async def test_inference() -> dict:
+    conn = get_connection()
+    base, key, model = _inference_creds(conn)
+    if not base or not key:
+        return {"ok": False, "mode": "fail", "vision_capable": False, "detail": "Inference is not configured"}
+    client = InferenceClient(base, key, model, timeout=20.0)
+    t0 = time.time()
+    try:
+        client.chat([{"role": "user", "content": "Reply with the single word: ok"}])
+    except InferenceConnectionError as e:
+        return {"ok": False, "mode": "fail", "vision_capable": False, "detail": str(e)}
+    except InferenceError as e:
+        return {"ok": False, "mode": "fail", "vision_capable": False, "detail": str(e)}
+    detail = f"reachable in {int((time.time()-t0)*1000)}ms"
+    return {"ok": True, "mode": "stateless", "vision_capable": True, "detail": detail}
diff --git a/backend/tests/test_inference.py b/backend/tests/test_inference.py
new file mode 100644
index 0000000..9e2b2ce
--- /dev/null
+++ b/backend/tests/test_inference.py
@@ -0,0 +1,128 @@
+"""S7 inference tests (METADATA-105/106, 45, 55, 70, SETTINGS-45)."""
+
+import base64
+import json
+import os
+import shutil
+import tempfile
+
+import pytest
+from fastapi.testclient import TestClient
+
+import backend.db as db_mod
+
+
+@pytest.fixture
+def env(monkeypatch):
+    tmp = tempfile.mkdtemp(prefix="inf_")
+    os.environ["SLAB_DB_PATH"] = os.path.join(tmp, "i.db")
+    os.environ["SLAB_IMAGES_DIR"] = tmp
+    os.environ["WOO_BASE_URL"] = "https://x.com"
+    os.environ["SLAB_AES_KEY"] = base64.b64encode(b"0123456789abcdef0123456789abcdef").decode()
+    db_mod._reset_for_tests()
+    from backend.crypto import encrypt
+
+    conn = db_mod.init_db()
+    conn.execute("INSERT OR REPLACE INTO settings(key,value,value_enc) VALUES ('inference_base_url',NULL,?)", (encrypt("https://infer.example.com"),))
+    conn.execute("INSERT OR REPLACE INTO settings(key,value,value_enc) VALUES ('inference_api_key',NULL,?)", (encrypt("sk-test"),))
+    conn.execute("INSERT OR REPLACE INTO settings(key,value,value_enc) VALUES ('inference_model','gpt-test',NULL)")
+    conn.commit()
+    from backend.app import app
+
+    yield TestClient(app), conn
+    db_mod._reset_for_tests()
+    shutil.rmtree(tmp, ignore_errors=True)
+
+
+def test_prompt_seed_and_revision(env):
+    client, conn = env
+    r = client.get("/api/v1/prompts/call1")
+    assert r.status_code == 200
+    body = r.json()
+    assert body["active_content"] is not None
+    assert len(body["revisions"]) == 1  # seed only
+
+    r = client.post("/api/v1/prompts/call1/revisions", json={"content": "revised prompt v2"})
+    assert r.status_code == 201
+    new_rev = r.json()["id"]
+    # new revision becomes active
+    r = client.get("/api/v1/prompts/call1")
+    assert r.json()["active_revision_id"] == new_rev
+    assert r.json()["active_content"] == "revised prompt v2"
+    assert len(r.json()["revisions"]) == 2
+
+    # activate back to seed
+    seed_rev = 1
+    r = client.post("/api/v1/prompts/call1/activate", json={"revision_id": seed_rev})
+    assert r.status_code == 200
+    assert client.get("/api/v1/prompts/call1").json()["active_content"] is not None
+
+
+def test_call1_confidence_gate(env, monkeypatch):
+    client, conn = env
+    # enable inference
+    conn.execute("INSERT OR REPLACE INTO settings(key,value,value_enc) VALUES ('inference_enabled','true',NULL)")
+    conn.execute("INSERT INTO woo_categories(id,name,slug,parent_id,woo_id) VALUES (1,'Black Walnut','black-walnut',NULL,100)")
+    conn.execute("INSERT INTO slabs(id,sku,length_in,thickness_in,status,server_rev,client_rev,created_at,updated_at) VALUES ('s1','SLAB-01',100,2,'calibrated',1,1,'t','t')")
+    conn.execute("INSERT INTO slab_photos(id,slab_id,seq,processed_blob,created_at) VALUES ('p1','s1',1,X'89504E47','t')")
+    conn.commit()
+    from backend.inference import InferenceClient
+
+    def fake_chat(self, messages):
+        return json.dumps({
+            "species_id": 1, "species_confidence": 0.9,
+            "wood_category_ids": [1], "wood_category_confidence": 0.5,  # below threshold
+            "grade_term_id": 7, "grade_confidence": 0.3,               # below
+            "figure_term_ids": [2], "figure_confidence": 0.8,          # at/above
+        })
+    monkeypatch.setattr(InferenceClient, "chat", fake_chat)
+    r = client.post("/api/v1/slabs/s1/inference/call1")
+    assert r.status_code == 200, r.text
+    s = r.json()["suggestions"]
+    assert s.get("species_id") == 1
+    assert s.get("figure_term_ids") == [2]
+    assert "wood_category_ids" not in s  # gated out
+    assert "grade_term_id" not in s       # gated out
+    # inference_log written with call=1
+    assert conn.execute("SELECT call FROM inference_log").fetchone()["call"] == 1
+
+
+def test_call2_stores_description(env, monkeypatch):
+    client, conn = env
+    conn.execute("INSERT OR REPLACE INTO settings(key,value,value_enc) VALUES ('content_llm_enabled','true',NULL)")
+    conn.execute("INSERT INTO slabs(id,sku,length_in,thickness_in,width_avg_in,bdft,title,status,server_rev,client_rev,created_at,updated_at) VALUES ('s1','SLAB-01',100,2,20,18,'Walnut Slab','ready',1,1,'t','t')")
+    conn.execute("INSERT INTO slab_photos(id,slab_id,seq,processed_blob,created_at) VALUES ('p1','s1',1,X'89504E47','t')")
+    conn.commit()
+    from backend.inference import InferenceClient
+
+    monkeypatch.setattr(InferenceClient, "chat", lambda self, messages: "<p>A fine walnut slab.</p>")
+    r = client.post("/api/v1/slabs/s1/inference/call2")
+    assert r.status_code == 200, r.text
+    assert r.json()["description"] == "<p>A fine walnut slab.</p>"
+    row = conn.execute("SELECT description, content_source FROM slabs WHERE id='s1'").fetchone()
+    assert row["description"] == "<p>A fine walnut slab.</p>"
+    assert row["content_source"] == "llm"
+
+
+def test_inference_off_call1_409(env):
+    client, conn = env
+    conn.execute("INSERT INTO slabs(id,sku,status,server_rev,client_rev,created_at,updated_at) VALUES ('s1','SLAB-01','calibrated',1,1,'t','t')")
+    conn.commit()
+    r = client.post("/api/v1/slabs/s1/inference/call1")
+    assert r.status_code == 409  # inference_enabled default off
+
+
+def test_server_description_inference_off(env):
+    from backend.inference import server_description
+
+    d = server_description({"title": "Figured Walnut Slab", "length_in": 114.0, "width_avg_in": 24.0, "thickness_in": 1.75, "bdft": 18.4})
+    assert "114.0in" in d and "18.4" in d
+
+
+def test_test_inference_unconfigured(env):
+    client, conn = env
+    conn.execute("UPDATE settings SET value_enc=NULL WHERE key='inference_api_key'")
+    conn.commit()
+    r = client.post("/api/v1/admin/test-inference")
+    assert r.status_code == 200
+    assert r.json()["ok"] is False and r.json()["mode"] == "fail"
diff --git a/frontend/src/App.svelte b/frontend/src/App.svelte
index b19ba3f..692f8f8 100644
--- a/frontend/src/App.svelte
+++ b/frontend/src/App.svelte
@@ -14,6 +14,7 @@ import SlabList from './create/slabs/SlabList.svelte';
 import PhotoAttach from './create/photos/PhotoAttach.svelte';
 import Review from './create/review/Review.svelte';
 import Publish from './create/publish/Publish.svelte';
+import GenerateText from './create/inference/GenerateText.svelte';
 
   const gate = typeof window !== 'undefined' ? supported() : 'window unavailable';
   const session: CaptureSession = createCaptureSession();
@@ -26,6 +27,7 @@ import Publish from './create/publish/Publish.svelte';
   let showPhotos = false;
   let showReview = false;
   let showPublish = false;
+  let showGenerate = false;
 
   const unsubCap = session.capture.subscribe((c) => {
     cap = c;
@@ -104,6 +106,7 @@ import Publish from './create/publish/Publish.svelte';
       <button on:click={() => (showPhotos = !showPhotos)}>Photos (dev)</button>
       <button on:click={() => (showReview = !showReview)}>Review (dev)</button>
       <button on:click={() => (showPublish = !showPublish)}>Publish (dev)</button>
+      <button on:click={() => (showGenerate = !showGenerate)}>Generate (dev)</button>
     </div>
     {#if showSettings}
       <Settings />
@@ -120,6 +123,9 @@ import Publish from './create/publish/Publish.svelte';
     {#if showPublish}
       <Publish />
     {/if}
+    {#if showGenerate}
+      <GenerateText />
+    {/if}
   </main>
 {/if}
 
diff --git a/frontend/src/create/inference/GenerateText.svelte b/frontend/src/create/inference/GenerateText.svelte
new file mode 100644
index 0000000..66baa50
--- /dev/null
+++ b/frontend/src/create/inference/GenerateText.svelte
@@ -0,0 +1,103 @@
+<script lang="ts">
+  let slabId = '';
+  let contentEnabled = false;
+  let reachable = false;
+  let description: string | null = null;
+  let call1Suggestions: Record<string, unknown> | null = null;
+  let error = '';
+  let busy = false;
+
+  async function checkReadiness() {
+    contentEnabled = false;
+    reachable = false;
+    try {
+      const [inf, test] = await Promise.all([
+        fetch('/api/v1/settings').then((r) => r.json()),
+        fetch('/api/v1/admin/test-inference', { method: 'POST' }).then((r) => r.json()),
+      ]);
+      contentEnabled = inf.content_llm_enabled === 'true' || inf.content_llm_enabled === true;
+      reachable = test.ok === true;
+    } catch {
+      reachable = false;
+    }
+  }
+
+  async function generate() {
+    if (!slabId) {
+      error = 'Enter a slab id.';
+      return;
+    }
+    busy = true;
+    error = '';
+    description = null;
+    try {
+      const res = await fetch(`/api/v1/slabs/${encodeURIComponent(slabId)}/inference/call2`, { method: 'POST' });
+      const body = await res.json();
+      if (res.ok) description = body.description;
+      else error = body.error?.message || `call2 failed (${res.status})`;
+    } catch {
+      error = 'network error';
+    }
+    busy = false;
+  }
+
+  async function classify() {
+    if (!slabId) {
+      error = 'Enter a slab id.';
+      return;
+    }
+    busy = true;
+    error = '';
+    call1Suggestions = null;
+    try {
+      const res = await fetch(`/api/v1/slabs/${encodeURIComponent(slabId)}/inference/call1`, { method: 'POST' });
+      const body = await res.json();
+      if (res.ok) call1Suggestions = body.suggestions;
+      else error = body.error?.message || `call1 failed (${res.status})`;
+    } catch {
+      error = 'network error';
+    }
+    busy = false;
+  }
+
+  checkReadiness();
+</script>
+
+<div class="inference">
+  <h2>Generate text</h2>
+  <p class="state">
+    Content generation: {contentEnabled ? 'on' : 'off'} · endpoint: {reachable ? 'reachable' : 'unreachable'}
+  </p>
+  {#if !contentEnabled || !reachable}
+    <p class="note">The Generate text control is disabled: it appears only when content generation is
+      enabled and the inference endpoint is reachable.</p>
+  {:else}
+    <div>
+      <label>Slab ID</label>
+      <input bind:value={slabId} placeholder="uuid" />
+    </div>
+    <div class="row">
+      <button on:click={generate} disabled={busy}>{busy ? 'Working…' : 'Generate description (Call 2)'}</button>
+      <button on:click={classify} disabled={busy}>Classify metadata (Call 1)</button>
+    </div>
+    {#if error}<p class="err">{error}</p>{/if}
+    {#if description}<p class="desc">{description}</p>{/if}
+    {#if call1Suggestions}
+      <pre class="sug">{JSON.stringify(call1Suggestions, null, 2)}</pre>
+    {/if}
+  {/if}
+  <button class="mini" on:click={checkReadiness}>Recheck readiness</button>
+</div>
+
+<style>
+  .inference { max-width: 560px; color: #eee; }
+  label { display: block; margin: 8px 0 4px; }
+  input { width: 100%; padding: 6px; }
+  .row { margin-top: 12px; display: flex; gap: 8px; }
+  .state { color: #aaa; }
+  .note { color: #e8a; }
+  .err { color: #e88; }
+  .desc { margin-top: 12px; }
+  .sug { background: #1c1c1c; padding: 8px; }
+  .mini { margin-top: 12px; font-size: 12px; }
+</style>