diff --git a/examples/parallel-decisions.md b/examples/parallel-decisions.md
new file mode 100644
index 000000000..340da6753
--- /dev/null
+++ b/examples/parallel-decisions.md
@@ -0,0 +1,64 @@
+# Parallel Decisions through treg
+
+Connect your team's Parallel API key once. Calls with this key are billed by Parallel to your
+Parallel account and do not debit your treg balance.
+
+```bash
+treg secret add parallel --value "$PARALLEL_API_KEY"
+treg catalog get parallel.ai-judge.decide
+```
+
+One predicate question:
+
+```bash
+treg call parallel.ai-judge.decide --data '{
+  "state": {"company": "Parallel", "website": "https://parallel.ai"},
+  "questions": {
+    "has_api": {"type": "noul", "instructions": "Does Parallel offer an API?"}
+  },
+  "include_excerpts": false
+}'
+```
+
+Several questions can share the same state and web evidence in one request:
+
+```bash
+treg call parallel.ai-judge.decide --data '{
+  "state": {"company": "Parallel", "website": "https://parallel.ai"},
+  "questions": {
+    "has_api": {"type": "noul", "instructions": "Does Parallel offer an API?"},
+    "product_type": {
+      "type": "choice",
+      "instructions": "Which description best fits Parallel?",
+      "criteria": {"web_api": "Offers a web API", "desktop_app": "Offers a desktop app"}
+    },
+    "documentation_level": {
+      "type": "score",
+      "instructions": "How detailed are Parallel's public API docs?",
+      "criteria": ["No docs", "Partial docs", "Detailed docs"]
+    }
+  }
+}'
+```
+
+Read each result under `answers.<question_name>`. The response includes the vendor's `decision_id`,
+`mode`, token `usage`, and shared `citations`. A question can return `type: "refusal"` instead of a
+decision value. See [Parallel's question types](https://docs.parallel.ai/decisions/question-types)
+for criteria and limits.
+
+## Verification ledger
+
+Checked against Parallel's [API reference](https://docs.parallel.ai/api-reference/decisions/make-web-grounded-typed-decisions),
+[question types](https://docs.parallel.ai/decisions/question-types),
+[best practices](https://docs.parallel.ai/decisions/best-practices),
+[pricing](https://docs.parallel.ai/getting-started/pricing), and
+[rate limits](https://docs.parallel.ai/getting-started/rate-limits) on 2026-10-30.
+The documented rate is $0.001 per `fast` request with up to 100 questions.
+
+`tests/test_parallel_decisions.py` uses a mocked upstream for single and multiple questions,
+response fields and citations, upstream error statuses, own-key precedence, and the platform-key
+gate. `src/treg/catalog/examples/parallel.ai-judge.decide.json` is synthetic and sanitized; it is
+not a captured live response. No live Parallel Decisions call or vendor billing observation was
+made. Platform-funded access remains disabled until a maintainer verifies the platform key can
+call Decisions, confirms actual charges and error billing against the shared account, and selects
+settlement behavior from that evidence.
diff --git a/src/treg/catalog/examples/parallel.ai-judge.decide.json b/src/treg/catalog/examples/parallel.ai-judge.decide.json
new file mode 100644
index 000000000..078f0bf82
--- /dev/null
+++ b/src/treg/catalog/examples/parallel.ai-judge.decide.json
@@ -0,0 +1,25 @@
+{
+  "mode": "fast",
+  "answers": {
+    "has_api": {"type": "noul", "noul": 0.99},
+    "product_type": {
+      "type": "choice",
+      "choice": "web_api",
+      "probabilities": {"web_api": 0.95, "desktop_app": 0.05},
+      "confidence": 0.91
+    },
+    "documentation_level": {
+      "type": "score",
+      "score": 1.8,
+      "legend": {"0": "No public documentation", "1": "Partial documentation", "2": "Detailed documentation"},
+      "probabilities": {"0": 0.0, "1": 0.2, "2": 0.8},
+      "confidence": 0.86
+    },
+    "unanswerable": {"type": "refusal"}
+  },
+  "usage": {"input_tokens": 1200, "output_tokens": 0},
+  "citations": [
+    {"url": "https://example.com/docs", "title": "Example documentation", "excerpts": ["Example excerpt."]}
+  ],
+  "decision_id": "decisions_00000000000000000000000000000000"
+}
diff --git a/src/treg/catalog/parallel.yaml b/src/treg/catalog/parallel.yaml
index 23c9af457..7f5d2bab7 100644
--- a/src/treg/catalog/parallel.yaml
+++ b/src/treg/catalog/parallel.yaml
@@ -2,11 +2,44 @@ provider: parallel
 source:
   docs: https://docs.parallel.ai/getting-started/overview
   openapi: https://docs.parallel.ai/public-openapi.json
-  curated: '2026-10-09'
+  curated: '2026-10-30'
 pricing_url: https://docs.parallel.ai/getting-started/pricing
-limits: "Search, Extract and Entity Search each allow 600 requests per minute on a rolling 60-second window. HTTP 429 is a rate limit; HTTP 402 is insufficient credit, which in-flight requests count against."
+limits: "Search, Extract, Entity Search and Decisions each allow 600 requests per minute on a rolling 60-second window. HTTP 429 is a rate limit; HTTP 402 is insufficient credit, which in-flight requests count against."
 
 endpoints:
+  - id: parallel.ai-judge.decide
+    capability: ai-judge.decide
+    platform: ai-judge
+    domain: decisions
+    scope: any_account
+    method: POST
+    path: /v1/decisions
+    platform_blocked: "Platform-key billing and settlement for Decisions have not been live-verified; connect your team's own Parallel key."
+    name: Make web-grounded decisions
+    summary: Answer up to 100 named predicate, choice or score questions using shared state and current web evidence.
+    input:
+      bodyType: json
+      body:
+        state: {type: any, required: true, example: {company: "Parallel", website: "https://parallel.ai"}, note: "Any JSON value shared by all questions. State plus the largest question is limited to 32,000 estimated input tokens."}
+        questions: {type: object, required: true, example: {has_api: {type: noul, instructions: "Does Parallel offer an API?"}}, note: "1–100 names matching [A-Za-z0-9_-]+. Each value has type noul, choice or score and instructions. Noul may include criteria {true, false}; choice requires 2–255 named option descriptions; score requires 2–10 ordered level descriptions. Each rendered question is limited to 5,000 characters."}
+        mode: {type: string, required: false, default: fast, enum: [fast], note: "Only fast mode is currently supported."}
+        include_excerpts: {type: boolean, required: false, default: true, note: "False keeps citation URLs and titles but returns empty excerpts arrays."}
+      note: "Answers are keyed by question name. Noul returns a probability; choice returns the chosen key and option probabilities; score returns a weighted mean, legend and probabilities. Choice and score may include confidence. Any answer may be a refusal. The raw response also includes mode, decision_id, token usage and citations shared across all questions; citations are not mapped to individual answers."
+    test_request: {body: {state: {company: "Parallel", website: "https://parallel.ai"}, questions: {has_api: {type: noul, instructions: "Does Parallel offer an API?"}}, include_excerpts: false}}
+    cost:
+      type: per_call
+      value: 0.001
+      currency: USD
+      unit: call
+      source: docs
+      source_url: https://docs.parallel.ai/getting-started/pricing
+      checked: '2026-10-30'
+      confidence: documented
+      note: "$0.001 per fast request, including up to 100 questions. Documented rate only; platform-key billing and settlement remain unverified, so this tool is team-key only. The response's usage is token counts, not a charge."
+    added: '2026-10-30'
+    example_response: examples/parallel.ai-judge.decide.json
+    docs_url: https://docs.parallel.ai/decisions/decisions-quickstart
+
   - id: parallel.web.search
     capability: web.search
     platform: web
diff --git a/tests/test_parallel_decisions.py b/tests/test_parallel_decisions.py
new file mode 100644
index 000000000..144f24853
--- /dev/null
+++ b/tests/test_parallel_decisions.py
@@ -0,0 +1,165 @@
+"""Parallel Decisions catalog contract and faithful own-key relay (mocked upstream)."""
+
+import json
+from pathlib import Path
+
+import httpx
+import pytest
+
+from treg import api as A
+from treg.config import get_settings
+from treg.domain.catalog import store as catalog_store
+
+
+ENDPOINT = "parallel.ai-judge.decide"
+FIXTURE = (Path(__file__).resolve().parents[1] / "src/treg/catalog/examples"
+           / f"{ENDPOINT}.json")
+
+
+class _JSONStream(httpx.AsyncByteStream):
+    def __init__(self, payload):
+        self.payload = json.dumps(payload).encode()
+
+    async def __aiter__(self):
+        yield self.payload
+
+
+def _response(status, payload, *, headers=None):
+    return httpx.Response(status, headers={"content-type": "application/json", **(headers or {})},
+                          stream=_JSONStream(payload))
+
+
+async def _balance(clients):
+    org_id = (await clients.get("/orgs")).json()[0]["org_id"]
+    return (await clients.get(f"/orgs/{org_id}/balance")).json()["balance_micro"]
+
+
+def test_decisions_catalog_contract():
+    catalog = catalog_store.load()
+    endpoint = catalog.by_id[ENDPOINT]
+    assert endpoint["capability"] == "ai-judge.decide"
+    assert (endpoint["method"], endpoint["path"]) == ("POST", "/v1/decisions")
+    assert endpoint["input"]["body"]["state"]["type"] == "any"
+    assert endpoint["input"]["body"]["questions"]["type"] == "object"
+    assert endpoint["input"]["body"]["include_excerpts"]["default"] is True
+    assert endpoint["cost"]["value"] == 0.001
+    assert endpoint["cost"]["confidence"] == "documented"
+    assert not catalog.platform_eligible(endpoint)
+    response = json.loads(FIXTURE.read_text())
+    assert {"noul", "choice", "score", "refusal"} == {
+        answer["type"] for answer in response["answers"].values()
+    }
+
+
+async def test_decisions_is_discoverable(clients):
+    response = await clients.get(f"/catalog/endpoints/{ENDPOINT}")
+    assert response.status_code == 200
+    assert response.json()["endpoint"]["id"] == ENDPOINT
+    assert response.json()["endpoint"]["platform_eligible"] is False
+    assert response.json()["example_response"]["decision_id"].startswith("decisions_")
+
+
+@pytest.mark.parametrize("state", [None, "Parallel", {"company": "Parallel"}])
+async def test_decisions_single_question_preserves_json_state_and_response(clients, monkeypatch, state):
+    secret = await clients.post("/secrets", json={"name": "parallel", "value": "OWN-PARALLEL"})
+    assert secret.status_code == 200, secret.text
+    request_body = {
+        "state": state,
+        "questions": {"has_api": {"type": "noul", "instructions": "Does Parallel offer an API?"}},
+        "include_excerpts": False,
+    }
+    reply = json.loads(FIXTURE.read_text())
+    seen = []
+
+    def serve(request):
+        seen.append((request.url.path, request.headers.get("x-api-key"), json.loads(request.content)))
+        return _response(200, reply)
+
+    before = await _balance(clients)
+    async with httpx.AsyncClient(transport=httpx.MockTransport(serve)) as upstream:
+        monkeypatch.setattr(A.app.state, "http", upstream)
+        response = await clients.post(f"/call/{ENDPOINT}", json=request_body)
+    assert response.status_code == 200, response.text
+    assert response.json() == reply
+    assert seen == [("/v1/decisions", "OWN-PARALLEL", request_body)]
+    assert "x-treg-cost-micro" not in response.headers
+    assert await _balance(clients) == before
+
+
+async def test_decisions_multiple_question_types_and_citations_are_relayed(clients, monkeypatch):
+    await clients.post("/secrets", json={"name": "parallel", "value": "OWN-PARALLEL"})
+    body = {
+        "state": {"company": "Parallel", "website": "https://parallel.ai"},
+        "mode": "fast",
+        "questions": {
+            "has_api": {"type": "noul", "instructions": "Does Parallel offer an API?",
+                        "criteria": {"true": "There is an API.", "false": "There is no API."}},
+            "product_type": {"type": "choice", "instructions": "What kind of product is it?",
+                             "criteria": {"web_api": "A web API", "desktop_app": "A desktop app"}},
+            "documentation_level": {"type": "score", "instructions": "How detailed are the docs?",
+                                    "criteria": ["None", "Partial", "Detailed"]},
+        },
+    }
+    reply = json.loads(FIXTURE.read_text())
+
+    def serve(request):
+        assert request.headers["x-api-key"] == "OWN-PARALLEL"
+        assert json.loads(request.content) == body
+        return _response(200, reply)
+
+    async with httpx.AsyncClient(transport=httpx.MockTransport(serve)) as upstream:
+        monkeypatch.setattr(A.app.state, "http", upstream)
+        response = await clients.post(f"/call/{ENDPOINT}", json=body)
+    assert response.status_code == 200, response.text
+    assert response.json() == reply
+    assert response.json()["answers"]["documentation_level"]["confidence"] == 0.86
+    assert response.json()["citations"][0]["excerpts"] == ["Example excerpt."]
+
+
+@pytest.mark.parametrize("status", [401, 422, 429, 503])
+async def test_decisions_upstream_errors_follow_the_relay(clients, monkeypatch, status):
+    await clients.post("/secrets", json={"name": "parallel", "value": "OWN-PARALLEL"})
+    error = {"type": "error", "error": {"ref_id": "example", "message": "Example error"}}
+
+    def serve(request):
+        return _response(status, error, headers={"retry-after": "3"})
+
+    before = await _balance(clients)
+    async with httpx.AsyncClient(transport=httpx.MockTransport(serve)) as upstream:
+        monkeypatch.setattr(A.app.state, "http", upstream)
+        response = await clients.post(f"/call/{ENDPOINT}", json={
+            "state": {}, "questions": {"q": {"type": "noul", "instructions": "Is it true?"}},
+        })
+    assert response.status_code == status, response.text
+    assert response.json() == error
+    assert await _balance(clients) == before
+
+
+async def test_decisions_never_uses_platform_key_until_billing_is_verified(clients, monkeypatch):
+    monkeypatch.setenv("TREG_PLATFORM_KEY_PARALLEL", "PLATFORM-PARALLEL")
+    monkeypatch.setenv("TREG_PLATFORM_PROVIDERS", "parallel")
+    get_settings.cache_clear()
+    seen = []
+    body = {"state": {}, "questions": {"q": {"type": "noul", "instructions": "Is it true?"}}}
+
+    def serve(request):
+        seen.append(request.headers.get("x-api-key"))
+        return _response(200, json.loads(FIXTURE.read_text()))
+
+    try:
+        before = await _balance(clients)
+        async with httpx.AsyncClient(transport=httpx.MockTransport(serve)) as upstream:
+            monkeypatch.setattr(A.app.state, "http", upstream)
+            missing = await clients.post(f"/call/{ENDPOINT}", json=body)
+            assert missing.status_code == 404
+            assert seen == []
+            assert await _balance(clients) == before
+
+            await clients.post("/secrets", json={"name": "parallel", "value": "OWN-PARALLEL"})
+            own = await clients.post(f"/call/{ENDPOINT}", json=body)
+            assert own.status_code == 200, own.text
+            assert seen == ["OWN-PARALLEL"]
+            assert "x-treg-cost-micro" not in own.headers
+            assert await _balance(clients) == before
+    finally:
+        get_settings.cache_clear()
diff --git a/tests/test_rankings.py b/tests/test_rankings.py
index 99894a622..22ccddd7c 100644
--- a/tests/test_rankings.py
+++ b/tests/test_rankings.py
@@ -5,6 +5,7 @@
 import re
 from datetime import date, timedelta
 
+import pytest
 from httpx import AsyncClient
 
 from treg.application import rankings as rankings_app
@@ -333,8 +334,8 @@ def test_a_model_that_moved_vendor_compares_with_all_its_earlier_calls():
             cat.endpoints.remove(ep); cat.by_id.pop(ep["id"])
 
 
-def test_a_decision_model_is_its_own_type_and_counts_in_all():
-    (jev,) = [e for e in store.load().endpoints if e["platform"] == "ai-judge" and store.browsable(e)]
-    board = rankings.build(_rows(jev["id"], {0: 50}), store.load(), LAST)["ai_models"]["7d"]
+@pytest.mark.parametrize("endpoint_id", ["openrouter.ai-judge.decide", "parallel.ai-judge.decide"])
+def test_a_decision_model_is_its_own_type_and_counts_in_all(endpoint_id):
+    board = rankings.build(_rows(endpoint_id, {0: 50}), store.load(), LAST)["ai_models"]["7d"]
     (model,) = board["models"]
     assert model["type"] == "Decision" and model["calls"] == 50