Measured, not marketed
Every number below comes from a recorded run you could demand to see: the questions, the judge protocol, the scores, and the rounds it took to get there. Where the numbers vary, we say so. Where we have no number, we say that too.
The certified result
Local certification on the deployed code, July 2026. Median of three external-judge passes per answer.
Cheque 9.2 · bail 9.0 · property 9.5 · consumer 9.0 · motor 9.0 — the certification run, per domain.
Four clean live runs on production: 9.06, 9.05, 8.90, 8.82. Per-run variance stated, not hidden.
Time to the first streamed content in certification runs. Live infrastructure varies — details below.
The method — five questions, one hostile judge
The evaluation set is five realistic client fact patterns across distinct practice areas — reproduced verbatim below. Each answer is produced by the live pipeline, then scored out of 10 by an external AI judge instructed as a strict Indian senior counsel, with graded rubrics and the current date. Each answer is judged three times; the median counts. The judge grades — it never helps draft.
s.138 NI Act · certified score, median of 3
“My client's Rs 8 lakh cheque from a supplier bounced twice with 'funds insufficient'. We sent the demand notice 25 days after the second return memo; the drawer's 15 days expired yesterday with no payment. Criminal and civil options, and what will the drawer argue?”
BNS / BNSS · certified score, median of 3
“FIR under s.318(4) BNS and s.316(2) BNS against my client, a woman and first-time accused, alleged amount Rs 12 lakh. Police have issued a s.35 BNSS notice. Anticipatory bail strategy, and what will the prosecution press against us?”
Specific Relief Act · certified score, median of 3
“A neighbour built a compound wall 3 feet into my client's vacant plot in Hyderabad six months ago. Client holds the registered sale deed and remains in possession of the rest of the plot. What is the correct remedy and forum?”
CPA 2019 · RERA 2016 · certified score, median of 3
“A builder has delayed flat possession by 3 years; my client paid Rs 85 lakh consideration and wants a refund with interest. Consumer commission or RERA, which commission has jurisdiction, and what is the limitation position?”
MV Act s.166 · certified score, median of 3
“My client's husband died in a truck collision 8 months ago — salaried Rs 60,000/month, aged 38, two minor children. The insurer says the claim is time-barred. Advise on maintainability and realistic compensation.”
Ask → answer on the live pipeline →
judge three times → keep the median.
No cherry-picked runs.
What we measure against
Did the answer identify every live legal issue in the fact pattern — including the ones the client did not ask about?
Are the statutes and judgments real, current, and correctly applied — with superseded law flagged rather than recycled?
Does it state the opponent's best arguments honestly and rebut them on their failing premise — no strawmen?
Right court, right tier, right procedural route — jurisdiction, limitation, and mandatory pre-steps.
Could an advocate act on this tomorrow morning — concrete filings, deadlines, and what still needs verification?
Does the answer claim only what the law supports — contested positions labelled as contested, never as settled?
Sober counsel's prose: structured, precise, no bluster and no filler.
Seven criteria, graded per answer.
The overall score is out of 10.
The judge is not kind
Three deductions, quoted verbatim from live run logs. This is the level of scrutiny behind every score on this page — half a point lost for a tax-slab miscalculation, a full point for overstating a developing line of authority as settled law.
“Missed citing Newtech Promoters (2021), the landmark Supreme Court judgment settling the jurisdictional split between RERA Authority (refund) and Adjudicating Officer (compensation).”
Consumer answer, live run — scored 8.8
“Incorrect tax math: 7.2L annual salary minus 50k standard deduction yields 6.7L, which incurs zero tax under the current Section 87A rebate.”
Motor-accident answer, live run — scored 8.0
“Overclaims contested law: DAR as an absolute bypass to the 6-month limitation is a developing High Court workaround, not settled Supreme Court law.”
Motor-accident answer, live run — same 8.0
How it got there — 6.25 to 9.14
The first fully judged build scored 6.25. Certification took roughly twenty-five recorded rounds of the same loop: run all five questions, let the judge tear the answers apart, trace every deduction to its root cause, fix it deterministically, run again. Every defect found on the way — from citation gaps to a bottom line that contradicted its own arithmetic — is logged in an internal error register with its cause and its fix. The bars below are the recorded round averages from that campaign’s final stretch.
Judged round averages, rounds 11–22 (median-of-3 each) · R22 = certification, all five areas ≥ 9 · earlier rounds climbed from 6.25
Speed, honestly stated
The measured band across certification runs: from question sent to the first streamed content on screen.
Identical code on live infrastructure has measured anywhere in this range. Upstream model latency varies with load; we publish the spread instead of quoting the best day.
A complete answer — issues, authorities, counter-analysis, strategy — finishes streaming in this range on recorded runs.
What this number does not mean
The same answer, judged twice, can move by ±0.5–1 points. That is why we certify on the median of three judge passes and on trends across runs — never on a single lucky score. Every figure on this page follows that protocol.
An average of 9.14 does not promise a 9 on your next question. Individual answers vary — which is exactly why every production answer still passes through the citation verification gate, regardless of what the benchmark says.
The certified score covers the five-question evaluation set described above. It is evidence of the pipeline's discipline, not a coverage claim for every domain of Indian law. We extend the set before we extend the claim.
The external judge is a frontier model instructed as a strict senior counsel with graded rubrics and the current date. It is not a human bench. We publish the method precisely so it can be criticised.
You will not find competitor scores here, because we have not run a sourced, like-for-like comparison we could defend. Head-to-head comparisons are in progress and will be published with full methodology.
The benchmark tells you how the pipeline performs on a defined set. What protects each individual answer is the verification gate — read how we verify for the mechanism.
Don’t take the average. Read an answer.
Three answers from a recorded live run are published verbatim — questions, timings, scores and all.