AnrakConductor — grounded legal reasoning as a model

Two model families

The model string picks the pipeline family and how hard it thinks. Any other model id returns an OpenAI-shaped 404 (model_not_found).

Legal Conductor — anrak-conductor*

The full Indian-law pipeline: planner → live research (case law, statutes, court records, web) → citation manifest → grounded scoring → the 3-cop verifier panel.

Pure Ensemble — anrak-ensemble*

The same multi-model panel with no legal machinery: no planner, no research lanes, no citation manifest. Your messages (system prompt included) go straight to the four experts, a Trusted Researcher consolidates the drafts, and the route-aware three-cop panel verifies the buffered final answer before release. Use it for general reasoning, code, strategy — anything where you want ensemble quality without Indian-law grounding.

Durable Deep Ensemble — anrak-*-long

Long tasks use a separate asynchronous endpoint so ordinary OpenAI-compatible completions keep their expected response shape. Submit the work once, close the client, and poll the task URL. Long drafts are planned and generated section by section without restarting completed sections; every section targets all four model families, requires a three-model quorum, and is withheld unless the cops pass it.

# Durable Deep Ensemble — submit, close the client, poll later.
import time, requests

headers = {
    "Authorization": "Bearer " + ANRAK_API_KEY,
    "Content-Type": "application/json",
}

task = requests.post(
    "https://anrak.legal/api/v1/orchestra/tasks",
    headers=headers,
    json={
        "model": "anrak-conductor-long",
        "task_type": "long_draft",  # or "answer"
        "messages": [
            {"role": "user", "content": "Draft a comprehensive technology services agreement..."}
        ],
    },
).json()

while task["status"] in ("pending", "processing"):
    time.sleep(3)
    task = requests.get(
        "https://anrak.legal/api/v1/orchestra/tasks/" + task["id"],
        headers=headers,
    ).json()

print(task["output"][0]["content"])

Per-request override: anrak.legal_research

Both families accept an optional anrak object in the request body. {"legal_research": false} makes an anrak-conductor* call behave pure; {"legal_research": true} makes an anrak-ensemble* call run the full legal pipeline. The depth carries across (router/ensemble ↔ full panel, deep ↔ deep, fast ↔ fast).

# The pure family with the OpenAI SDK — any general question,
# your own system prompt, no legal machinery:
resp = client.chat.completions.create(
    model="anrak-ensemble",
    messages=[
        {"role": "system", "content": "You are a terse staff engineer."},
        {"role": "user", "content": "Design a rate limiter for a multi-tenant API."},
    ],
)

# Flip either family per-request with the anrak.legal_research override:
resp = client.chat.completions.create(
    model="anrak-conductor",                     # legal family…
    messages=[{"role": "user", "content": "..."}],
    extra_body={"anrak": {"legal_research": False}},  # …run PURE this time
)
# and the reverse: model="anrak-ensemble" + {"legal_research": True}
# runs the full legal pipeline under the pure family's name.

Pricing

The Model API bills per request (per section for the durable ids), not per token. Your cost is driven by fan-out you never see — four drafters, a consolidation pass, and a verifier panel, plus live research on the legal family — so a per-token price would be both unstable and misleading. A long answer never costs you more than a short one.

Per-model prices are listed in the tables above; each model page carries a monthly estimator.

Try it live

Runs against your signed-in account (enable AnrakConductor in Labs first). Ask an Indian-law question and watch the panel deliberate and verify.