Anrak Ensemble: Deep
anrak-ensemble-deepanrak/ensemble-deepThe same AB-MCTS search as AnrakConductor: Deep, reusing the same engine over the general-reasoning panel. The difference is the scorer, and it matters. With no retrieval corpus to check against, branches are ranked by an LLM judge on consistency, completeness and instruction-following. That is a strictly weaker signal than a symbolic citation check, and a confident fabrication can fool it in ways the grounded scorer cannot. We publish this rather than let you discover it.
The Panel
Four independent jurists draft the same question in parallel. They are deliberately different model families, so their errors are largely independent — the condition under which fan-out plus aggregation actually beats one strong model.
| Jurist | Lens | What it contributes |
|---|---|---|
| Athena | The structured analyst | Decomposes the question and covers every sub-issue methodically. Usually the most complete draft. |
| Vikram | The frontier reasoner | Pushes maximum depth on the hardest part of the problem. The rollout engine for deep search. |
| Themis | The meticulous verifier | Careful step-by-step chains that flag their own uncertainty rather than leaping to a conclusion. |
| Chanakya | The pragmatic strategist | Trained closest to Indian legal language and drafting convention; on the pure family, the practical-implications lens. |
Pricing
One flat charge per call, whatever the answer's length. A request to this id runs ~26 underlying model calls — that fan-out, not your token count, is what it costs us, so that is what we price on.
Estimate a month
Requests per month to anrak-ensemble-deep.
- Per request
- ₹150.23
- Per month
- ₹45,06,900
- Per 1,000 calls
- ₹1,50,230
Comparing against a per-token model?
On a reference request (2,000 tokens in, 2,000 out) this price works out to an effective $170.72 / $682.86 per 1M tokens.
That number looks high next to a raw frontier model, and it should: a raw model bills you for one call, while this id bills for 26 and a verification pass. We publish the comparison rather than hide it — but we bill per request, so a long answer does not cost you more.
Verification
Logic Cop only
With no retrieval corpus there is nothing to check citations against, so a single consistency reviewer reads the answer for internal contradictions and fabricated-looking specifics. It annotates; it does not correct. Treat factual specifics as unverified.
The verdict rides on every response as anrak.verify, the drafts and their grounded scores as anrak.orchestration, and the ordered trace as anrak.reasoning. When streaming, the same narration arrives live as delta.reasoning_content.
Quickstart
Two changes from an ordinary OpenAI call: the base_url and your ANRAK- key.
from openai import OpenAI
client = OpenAI(
base_url="https://anrak.legal/api/v1",
api_key="ANRAK-xxxxx",
)
resp = client.chat.completions.create(
model="anrak-ensemble-deep",
messages=[{"role": "user", "content": "Design a rate limiter for a multi-tenant API."}],
)
print(resp.choices[0].message.content)
# What it cost, and whether a downgrade happened:
extra = resp.model_extra or {}
print(extra["anrak"]["billing"]) # {'model': 'anrak-ensemble-deep', 'price': '₹150.23', ...}
print(extra["anrak"]["downgraded"]) # None unless the pipeline was downgradedLimits and trade-offs
- The judge has no corpus. A confident fabrication can score well. Prefer anrak-conductor-deep for anything about Indian law.
- Professional plan or above; otherwise it runs as anrak-ensemble. Read anrak.downgraded to detect it.
- Pays for a judge call per rollout on top of the rollouts themselves.
Best for
- • Hard non-legal problems where you want search, not a single draft
- • Problems whose quality is judgeable from the text alone — proofs, designs, plans
FAQ
What exactly am I charged for?
Which permissions does a key need?
What happens on a plan below Professional?
anrak.downgraded names what you asked for, what ran, and why.