How Nyave checks a Special Leave Petition: four pieces, all deterministic.
Every page of the filing is read. Every rule carries its consequence. Facts are checked across the whole paper book. Every finding is checked again before it is reported.
Page-Level Reading
Every page of a filing is read on its own terms: typed, scanned or vernacular, body text or stamp and seal. A poor scan never yields false certainty.
A Consequence-Weighted Rule Base
Each of the 106* rules carries the outcome the Registry attaches to it, measured on 14,000+* defect instances, so the report ranks what stops a filing ahead of what draws a remark.
Cross-Document Consistency
The same fact is checked wherever it appears in the paper book: the impugned order's particulars against every document citing them, party names and dates across the documents that carry them.
| Information Being Checked | Documents That May Need to Agree |
|---|---|
| Party names and representation | Impugned judgment, petition, vakalatnama, affidavit |
| Particulars of the impugned order | Every document citing them |
| Dates and document counts | Petition, affidavit and related applications |
A Self-Check Gate
Every finding is checked again before it is reported. Where Nyave cannot be sure, it says so.
It fails loudly: where the engine is not sure, the report says so and the Advocate-on-Record confirms.
Affidavit not filed, blank, or not properly attested/identified
Checklist item 4 · Order XXI Rule 3(1)(e)
Engine uncertain. AOR must manually confirm.
The rule engine detects. AI explains.
Output in minutes, by severity and confidence, each finding cited to its rule. AI explains; it never decides, never sees the document.
Nyave uses a purely deterministic rule engine to discover defects. Generative models are restricted exclusively to explaining those defects in plain language; artificial intelligence never decides whether a violation exists.
rules in the knowledge layer, built on Supreme Court Rules 2013, the Supreme Court Handbook 2025, and 13* Registry circulars.
in ~5 mins*.
The same filing produces the same result every time, and each finding can be checked against the rule it cites.
Built with Indian OCR.
A Special Leave Petition filing is typed English alongside scanned annexures: handwritten Hindi FIRs, faint photocopies of lower-court orders, pages carrying stamps, seals and signatures.
Nyave ran its own benchmark of OCR engines on 492* pages from 13* real Special Leave Petition filings: 22* document types, in English, Hindi and Gujarati.
The pages were sampled to mirror a real filing, then weighted with the hard cases: dense tables, degraded scans, Hindi and Gujarati pages.
Each engine was tested at its own documented operating point; an earlier round that had not done so was set aside and run again.
Each page was graded on two things separately: how faithfully the text was reproduced, and whether the engine invented anything.
Before the full run of 3,444* page readings (492* pages, seven engines), 300* page readings were graded manually by the founder, and the grading was calibrated against them. Those pages stayed inside the run to detect drift.
No single engine reads all six kinds of page a filing contains. The hybrid is a finding, not a preference.
So Nyave reads filings with a hybrid: its own self-hosted engine, and Sarvam. The engine has read every page of 11* complete real filings: 4,695* pages, scanned or typed.
Part of Sarvam for Startups. Reading uses Indian OCR, including Sarvam's service under Sarvam's terms.
Runs on AWS. Part of AWS Activate.
106* rules, each traced to its source.
Every flagged error is traceable back to its underlying statutory provisions and official registry checklist items.
Before any rule was written the Court's own sources were read in full: the Supreme Court Rules, 2013, the Handbook on Practice and Procedure, the e-filing manual, the Registry's default list and its circulars.
Every rule traces to a primary source, and the Registry's actual defect language takes precedence over assumed language.
- 1The Supreme Court Rules, 2013Amended by gazette six times since 2019.
- 2The Handbook and the Registry's Circulars13* Registry circulars, 2020 to 2026. In January 2025 the 21-item checklist was replaced by a 17-item checklist for the Registry and a 15-item checklist for advocates.
- 3Real FilingsA library of 12* real Special Leave Petition filings, with their actual Registry defect sheets, is the ground truth.
- 4The DatasetIt confirms how often. The official sources confirm what is correct.
- 5Lawyer InputTaken and then validated against the dataset before use; it advises and determines nothing.
Tested in stages, ending on real Special Leave Petitions and their actual Registry defect notices.
The measure of the product is whether its results match the Registry's actual defect record.
A confident wrong answer is worse than a slow right one.
It tells an advocate what the Registry is likely to flag. It never tells a lawyer how to do their job.
- The rule each finding is cited to, so it can be checked.
- The rules version it was checked against.
- The statement that Registry officers exercise discretion.
- Data protection built in from the first day: consent at upload, 24-hour deletion, no training use of uploaded documents, a published privacy notice with a named grievance contact.
- Draft petitions.
- Advise on merits.
- Replace the AOR's judgment.
- No templates, no case-law search, no other courts, no other filing types.
* Figures are from Nyave's own dataset and records, or from public data; actual figures may differ.
