TL;DR: Below are 15 prompts an Indian advocate can paste into a general-purpose AI tool for issue framing, research, drafting, chronology building, and cross-examination prep. Every prompt instructs the model to cite only verifiable authority and is paired with a specific check you run before the output leaves your desk, because a general model will still get case names, section numbers, and quoted paragraphs wrong.

On this page

Why every prompt here ends with a check

On 2 July 2026, the Supreme Court of India decided Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. & Anr., Civil Appeal No. 11950 of 2025, 2026 INSC 668. The bench of Justice Pamidighantam Sri Narasimha and Justice Alok Aradhe set aside an NCLT order and an NCLAT judgment because six of the precedents relied on were fake, wrongly cited, or carried paragraphs that do not exist in the real reports. The Court held that citing AI-generated case law without verification is misconduct, that reliance by a judge is a serious lapse, and that a decision built on fake material is no decision in the eyes of the law even if only a fragment entered the reasoning. Full detail, including the citation table showing how each fake authority failed, is in AI-generated case law: the Supreme Court’s 2026 rule.

That is the environment every prompt below is written for. Asking a general model to find case law supporting an argument, without a verification instruction, invites exactly what put six fabricated citations into an NCLT order. So every prompt here instructs the model not to present unverified case law as fact, and pairs it with a human step. Skip the human step and the prompt is not safer than the one that got sanctioned.

This is not a tool comparison; that ground is covered in ChatGPT for lawyers in India, Claude vs ChatGPT vs Gemini for Indian legal work, and native legal AI vs generic GPT tools. This piece assumes you already have a model open and gives you the prompts themselves, verification folded in.

How to use this library

Each prompt below follows the same shape: what it is for, the exact text to paste, what it will get wrong even when it looks right, and how you check the output before it reaches a client file or a court. Replace the bracketed placeholders with your facts, and keep the verification instruction inside each prompt even if it feels redundant, since the model does not remember being told this in an earlier prompt in the same session.

A general-purpose model run without retrieval against a real database of Indian judgments cannot confirm a case exists. It can only produce text that resembles a citation. These prompts make that failure visible when it happens. Only checking the primary source removes it.

Issue framing from a brief

Turning a messy client narrative into a clean list of legal issues is where AI genuinely saves time; the task is organisational, not substantive, so the risk is lower. But the model can still invent facts absent from your brief or blur two causes of action into one.

You are assisting an Indian advocate in framing legal issues from a client brief. I will paste a fact narrative below. Read it and produce:

1. A numbered list of every distinct legal issue that arises on these facts, stated as a question (for example "Does the six-year limitation period under Article 113 of the Limitation Act, 1963 bar this claim?").
2. For each issue, name the branch of law it falls under (contract, tort, property, criminal, constitutional, or other) without citing any specific case or section number.
3. Flag any issue where the facts as given are insufficient to frame the question precisely, and state what additional fact you would need.

Do not cite any case law, section number, or statute in this response. Do not add facts that are not in the text I paste. If a fact is ambiguous, note the ambiguity rather than resolving it in the client's favour.

Brief:
[paste the client narrative here]

What it will get wrong: the model can merge two causes of action into one issue, or split one into two over a stray sentence. It may also infer a limitation period or forum despite being told not to cite statute numbers.

How to check: read the issue list against the brief line by line and confirm each issue traces to a specific fact, not an inference. Cross off anything that does not map to your brief, and add any issue it missed, particularly limitation and jurisdiction.

Finding authority on a point of law

This is the highest-risk category in the entire library, because it directly overlaps with what went wrong in Pooja Ramesh Singh. Use this prompt only as a starting point for a search you will run yourself on a primary source, never as the final answer.

You are helping an Indian advocate identify the general legal principle relevant to a point of law. Do not name specific cases with confidence; I will verify all citations independently through a primary source.

Question under Indian law: [state your legal question precisely, for example "Can a corporate guarantee survive a demerger of the guarantor company absent an express novation clause?"]

1. State the principle in plain terms, without inventing a case name, citation, or bench composition.
2. Name the governing statute and section, and note that I must confirm the section number against the current bare act.
3. If a landmark case widely known to decide this point exists, name it by party names only, without a citation number, and flag that I must confirm citation, year, and holding on Indian Kanoon, SCC Online, or the Supreme Court's judgment portal before use.

Do not state as fact any citation, paragraph number, or bench composition you cannot verify. If unsure a case is real and correctly described, say so instead of guessing.

What it will get wrong: the model may name a real case but attribute a holding to a different, similarly named one, or invent a citation for a reporter reference it does not actually know, exactly the failure mode where two of the six fake authorities in Pooja Ramesh Singh were real cases carrying invented paragraphs.

How to check: search every case name independently on Indian Kanoon or your paid database, aware of the limitations of manual keyword search on that platform. Confirm the case exists, confirm the citation matches, then open the judgment and confirm the attributed paragraph is actually there. If you cannot find it, discard the citation rather than softening it into hedged language.

A system that only returns authority from a fixed corpus of real Indian judgments cannot cite a case outside that corpus. Niyam works this way for statute and case search, with every result tied back to the judgment it came from.

Distinguishing an adverse judgment

When the other side cites authority against you, the useful AI task is not finding a counter-case but structuring the distinguishing argument once you already have the adverse judgment open.

You are assisting an Indian advocate in distinguishing an adverse judgment that has been cited against my client. I will paste the full text or a substantial extract of the adverse judgment below. Do not search for or cite any other case in your response; work only from the text I give you.

Based only on the text below:
1. Identify the material facts the court relied on to reach its holding.
2. State the ratio decidendi as narrowly as the text supports, not as broadly as the headnote might suggest.
3. List every factual or legal distinction between the facts in this judgment and the following facts of my case: [state your client's facts].
4. Rank the distinctions from strongest to weakest, with one sentence on why each ranking is placed where it is.

Do not introduce facts about my case that I have not stated. Do not cite any other authority. Flag if the adverse judgment's facts are close enough to mine that a genuine distinction is difficult to make honestly.

What it will get wrong: the model tends to state the ratio more broadly than the court held, and it can manufacture a distinction that sounds persuasive but is unsupported by the text, particularly around procedural posture.

How to check: read the stated ratio against the judgment’s paragraph numbers and confirm it is not broader than what the court said. For each distinction, locate the specific sentence that supports it and drop any you cannot tie to a passage.

A legal notice is largely a repeatable structure with facts substituted in, which makes it a good AI drafting task, provided the demand and the deadline are yours, not generated.

Draft a legal notice under Indian civil practice. Do not add any relief, deadline, or legal ground I have not specified below. Do not cite any case law. You may cite the specific statutory provision I name, but add no other section on your own.

Facts: [state the facts in chronological order]
Cause of action: [state it, for example "dishonour of cheque under Section 138 of the Negotiable Instruments Act, 1881"]
Statutory provision to cite, if any: [name it exactly]
Relief sought: [state the exact amount or action demanded]
Compliance deadline: [state the number of days]
Sender and recipient details: [names, addresses]

Structure the notice with a heading, a numbered recital of facts, the demand, the statutory basis given, the deadline stated as an absolute date, and a closing paragraph reserving the right to initiate proceedings on non-compliance. Flag if the deadline is inconsistent with the statutory minimum for this cause of action.

What it will get wrong: the model can miscalculate the deadline date from a stated number of days, particularly around month-end, and it can silently add a second provision (commonly an old Indian Penal Code reference, which no longer applies since the Bharatiya Nyaya Sanhita, 2023 replaced it from 1 July 2024) you did not ask for.

How to check: recompute the compliance date yourself, confirm no additional Act or section appears beyond what you specified, and confirm every recital traces to your facts. For a Section 138 notice, cross-check the 30-day statutory demand window against the dishonour intimation date, not the drafting date.

A reply notice needs the same discipline as the notice itself, with the added risk that a general model will try to argue law on your behalf rather than simply structuring your instructed defence.

Draft a reply to the legal notice below on behalf of my client. I will paste the notice text and my client's version of events. Structure the reply to: (1) acknowledge receipt without admitting any allegation, (2) deny each factual allegation specifically rather than with a general denial, (3) state my client's version of the facts, (4) state the specific defence or ground I give you below, and (5) close by reserving my client's rights.

Do not introduce a legal ground, statutory defence, or case law that I have not specified below. Do not admit any fact by implication through careless phrasing; every paragraph that addresses an allegation from the notice must use explicit denial language.

Notice received: [paste the notice text]
Client's version of facts: [state it]
Specific defence or ground to raise: [state it, for example "the cheque was issued as security and not in discharge of a legally enforceable debt"]

What it will get wrong: general denial language is a common shortcut the model reaches for, which is legally weaker than paragraph-by-paragraph denial. It may also volunteer a defence adjacent to the one you specified.

How to check: read the reply against the notice paragraph by paragraph and confirm every allegation gets an explicit, individual denial. Remove any defence or fact introduced beyond your instruction.

Summarising a long judgment

Summarisation is one of the safer AI tasks because the model works entirely from text you supply rather than from its training data, provided you paste the actual judgment and do not ask it to supplement from memory.

Summarise the following judgment. Work only from the text I paste; do not add facts, holdings, or context from outside it, even if you recognise the case.

1. A two-sentence summary of the outcome (who won, what relief was granted or denied).
2. The material facts, chronologically, in no more than 200 words.
3. The issues the court framed, in the court's own words where explicit.
4. The holding on each issue, with the paragraph number where it appears.
5. Any dissent or separate opinion, summarised separately, with paragraph numbers.

Cite a paragraph number for every substantive claim. If none is visible in the source, say "paragraph not numbered in source" rather than guessing.

Judgment text:
[paste the full judgment or the relevant extract]

What it will get wrong: on long judgments, the model tends to compress or skip minority opinions, and can misattribute a paragraph number when OCR has dropped or duplicated markers in a scanned PDF.

How to check: open the source at each cited paragraph number and confirm the holding is actually stated there. Confirm no material issue was left out, particularly costs orders and interim directions.

Extracting the ratio decidendi

Separating the ratio from obiter is a genuinely hard legal skill, and a general model will often present the widest possible reading of a judgment as its ratio because broad statements are more common in training text than narrow, fact-bound ones.

Analyse the following judgment and separate the ratio decidendi from obiter dicta. Work only from the text pasted; do not draw on outside knowledge of this case.

1. State the ratio as narrowly as the material facts require, in one or two sentences, quoting the passage that supports this framing.
2. List every statement of law broader than what the facts required, labelled obiter, with paragraph number.
3. Explain in one sentence why each obiter statement was unnecessary to the decision, tied to the facts.
4. State any point where the ratio is genuinely ambiguous, and give both readings.

Do not state the ratio more broadly than the facts justify. Quote, do not paraphrase, the passage that states it.

Judgment text:
[paste the full judgment or the relevant extract]

What it will get wrong: models default to headnote-style generalisation, the opposite of a proper ratio analysis, and can present a passage from counsel’s submissions, not the court’s own view, as if it were the holding.

How to check: confirm the quoted passage comes from the court’s reasoning, not a paragraph beginning “learned counsel submitted”. Confirm the ratio would change if the material facts changed, the working test for ratio at all.

Building a chronology from documents

Chronology building from a bundle of documents is a strong AI use case because it is mechanical extraction rather than legal reasoning, but the model will still misread dates in Indian date formats and duplicate entries across documents that describe the same event.

Build a chronology of events from the documents I paste below. Work only from dates and facts explicitly stated in the text; do not infer a date that is not written in the source.

For each entry:
1. State the date in DD Month YYYY format.
2. State the event in one sentence.
3. Name the source document and, if visible, the page or paragraph it comes from.
4. If two documents describe what appears to be the same event with different dates, flag the conflict explicitly rather than picking one.

Sort the chronology in ascending date order. If a document uses a date format that is ambiguous between DD/MM and MM/DD, flag it rather than guessing, and state both possible readings.

Documents:
[paste the extracted text of each document, labelled by source]

What it will get wrong: DD/MM versus MM/DD ambiguity is the most common error, particularly with foreign correspondents or US-format systems, and the model can create one entry per document even when two documents describe the same event.

How to check: re-verify every date that could plausibly be read either way against the original document, and merge any entries describing the same event.

Cross-examination question sets

Question generation is useful for structuring an examination but the model has no access to the witness’s actual prior statements unless you give them to it, so unsupported “gotcha” questions are the main risk here.

Draft a set of cross-examination questions for the following witness, based only on the prior statement and documents I paste below. Do not invent any fact, admission, or inconsistency that is not present in the text I give you.

Witness's role: [state it]
Prior statement or affidavit: [paste it]
Documents that may contradict the statement: [paste them, labelled]

Produce:
1. A set of foundational questions that lock the witness into their prior statement before any contradiction is raised.
2. A set of questions that put each specific document to the witness, one document at a time, without revealing the contradiction until the witness has committed to an answer.
3. For each question, note in brackets which paragraph of the prior statement or which document it is drawn from.

Do not phrase any question that assumes a fact not stated in the material I gave you. Do not draft a question implying a document says something it does not say.

What it will get wrong: the model can draft a question that reveals the contradiction before the foundational sequence is complete, and it can misjudge which fact is actually contested versus already admitted.

How to check: read the sequence in order and confirm no question tips off the witness early. Verify each contradiction is genuine by placing the statement and the document side by side yourself.

Translating a vernacular order

Trial courts and many High Court benches issue orders in Hindi and other regional languages. Translation is a defensible AI use case, but a mistranslated term of art can change the substance of what you tell a client.

Translate the following court order from [source language, for example Hindi] into English. This is a legal document, so preserve legal terms of art precisely rather than translating them loosely.

Rules:
1. Translate paragraph by paragraph, keeping the same paragraph numbers as the source.
2. Where a term has a specific legal meaning under Indian procedure (for example, terms relating to bail, remand, stay, or limitation), state the English legal term used in Indian courts, not a literal dictionary translation.
3. If a word or phrase is ambiguous or you are not fully confident of the translation, mark it with [UNCERTAIN: original text] rather than guessing silently.
4. Do not summarise. Translate the full text.

Order text:
[paste the vernacular order]

What it will get wrong: procedural terms of art are the highest-risk category, since a mistranslation can turn “interim bail” into “bail” or blur an order dismissing an application with one merely adjourning it, opposite outcomes for a waiting client.

How to check: have a fluent reader, ideally with legal training, check every paragraph marked uncertain, and separately verify the operative paragraph word for word.

Explaining a case to a client

Clients need the substance of an order without the procedural language. The model’s plain-language instinct helps here, provided the underlying facts come from you and not the model’s guess at the case.

Explain the following court order to a client with no legal training. Do not add interpretation, prediction, or next-step advice unless I ask separately; explain only what the order says and means right now.

1. One sentence: what the court decided.
2. What this means practically (what changes, what stays the same, any deadline now applying).
3. Any date by which the client or advocate must act, stated as an absolute date if the order specifies a number of days from today, which is [state today's date].
4. Plain-language definitions of any Latin or technical term in the order.

Use short sentences. Do not speculate about appeal chances or strategy.

Order text:
[paste the order]

What it will get wrong: date arithmetic again, when an order says “within four weeks from today” and the model does not reliably use the order’s actual date, and it can drift into strategic advice despite being told not to.

How to check: verify the deadline against a calendar using the order’s actual date, not the date you ran the prompt, and delete any sentence that strays into advice about appeal prospects.

Drafting written submissions or a synopsis

A synopsis or written submission needs your arguments and your authorities; the model’s job is structuring and prose, not sourcing the law.

Draft written submissions for the following matter, using only the arguments, authorities, and facts I give you below. Do not add a new argument, ground, or case citation that I have not supplied.

Case: [name, court, case number]
Arguments to include, in the order I want them presented: [list your arguments]
Authorities to cite for each argument, exactly as I state them: [list case name, citation, and the specific proposition each supports]
Facts relevant to each argument: [state them]

Structure the submissions with a brief statement of the case, then each argument as a numbered heading, with the facts and the authority I gave you under each heading, closing with a prayer for relief matching: [state the relief sought].

Do not paraphrase the propositions I attributed to each authority into something broader than what I stated. Do not add a case law citation anywhere that I did not supply in the list above.

What it will get wrong: even when told not to add citations, the model can broaden the proposition attributed to a case, making the authority sound stronger than you stated.

How to check: read each heading against the exact proposition you fed the model and confirm nothing was strengthened beyond what you wrote. Confirm zero citations appear beyond your list.

Interpreting a statutory provision

Interpretation prompts are useful for identifying the questions a section raises, not for answering them with invented case law layered on top.

I will paste the text of a specific statutory provision below. Analyse its plain language only. Do not cite any case law interpreting this provision; I will research that separately.

Provision text: [paste the exact text of the section, as in force today]

Produce:
1. A plain-language restatement of what the provision requires or prohibits.
2. Every condition or element that must be satisfied for the provision to apply, listed separately.
3. Any term in the provision that is undefined in the text and would need a definition from elsewhere in the Act or from judicial interpretation.
4. Any internal ambiguity in the wording itself, stated without resolving it.

Do not tell me what courts have held about this provision. Work only from the words on the page.

What it will get wrong: the model can import a definition from a similarly worded statute, for example an Indian Penal Code reading into a Bharatiya Nyaya Sanhita, 2023 provision with different wording, a live risk since the BNS, BNSS, and BSA replaced the IPC, CrPC, and Evidence Act from 1 July 2024.

How to check: confirm the provision text is the current version against India Code, and confirm no definition was imported from a different Act than the one you pasted.

Preparing an RTI application

RTI applications are short and formulaic, which makes them a low-risk, high-value AI drafting task, provided the information sought is specific enough to survive a “vague query” rejection.

Draft a Right to Information application under the Right to Information Act, 2005 addressed to the Public Information Officer below, seeking the information specified. Do not add any request beyond what I specify.

Public authority and PIO: [name and address]
Information sought: [state precisely what records, dates, and categories you want]
Applicant details: [name, address; note citizenship, since Section 3 restricts the right to citizens]

Draft with: (1) the statutory heading citing Section 6(1), (2) each item of information as a separate numbered point specific enough that a PIO cannot call it vague, (3) the fee enclosed or fee waiver request if applicable, (4) closing details.

Flag if any information sought falls within the Section 8 exemptions, so I can decide whether to reword.

What it will get wrong: the model can be overcautious and flag a legitimate request as exempt under Section 8, or under-flag one genuinely touching a Section 8(1) exemption such as cabinet papers.

How to check: read the Section 8 flag against the actual text of Section 8, Right to Information Act, 2005 yourself, and confirm every information item is specific enough to survive a vagueness objection at first appeal.

Checking a limitation period

Limitation calculations are pure arithmetic dressed up as law, exactly where confident but wrong date math causes the most damage, so this prompt forces the model to show its working.

Calculate the limitation period for this claim under the Limitation Act, 1963. Show the calculation step by step, not only a final date.

Cause of action: [state it]
Date the cause of action arose: [state the exact date]
Applicable Article of the Schedule, if known: [state it, or leave blank and ask the model to identify the likely Article, flagged for my verification]
Any period to be excluded (for example time under Section 14, or a stay order): [state it]

Show: (1) the Article applied and the period it prescribes, (2) the start date, (3) any exclusion and why, (4) the resulting final filing date, calculated day by day or month by month.

Flag explicitly if uncertain which Article applies, rather than picking one silently.

What it will get wrong: the model can apply the wrong Article of the Schedule for a cause of action with more than one plausible characterisation, and its date arithmetic can be off by a day around months of different lengths.

How to check: verify the cited Article matches your cause of action by reading the Schedule text yourself, since limitation periods range from 30 days to 12 years, and recompute the final date by hand. A limitation date is the one error here a court will not let you fix later.

Where a general model holds up and where it does not

TaskGeneral modelWhy
Structuring an argument you already have✓No new facts or citations needed, purely organisational
Summarising a judgment text you paste in✓Works from supplied text, not memory
Translating a vernacular order✓ with human checkStrong at prose, weak on legal terms of art
Building a chronology from documents you supply✓Mechanical extraction, verifiable against source
Drafting boilerplate notices with your facts✓Structure is repeatable, facts are yours
Explaining an order in plain language✓ with human checkGood at register, weak at date arithmetic
Naming a case that supports your argument✗Fabrication risk demonstrated in Pooja Ramesh Singh, 2026 INSC 668
Giving a citation number from memory✗No reliable way to confirm a reporter reference without retrieval
Predicting how a specific bench will rule✗Not a task language models are built to answer reliably
Confirming a judgment is still good law✗Requires a live citator against subsequent judgments, not recall
Calculating a limitation deadline as a final answer✗Date arithmetic errors are common and consequential

The pattern is the same throughout: a general model is reliable working from text you supplied, and unreliable the moment it has to produce a fact from memory, whether a case name, a citation, or a computed date. A citator that checks whether a judgment is still good law closes that gap against a live database rather than training data. Niyam’s citator runs this as a standing feature, not a one-off prompt.

The same logic applies to finding authority in the first place. A similar-judgment search grounded in an actual case database cannot fabricate a case, since it has nothing to fabricate from.

The verification workflow

Every prompt above follows the same four-step path from draft to filing. The step most guides skip is the third one.

flowchart LR
    A[Prompt with verification instruction] --> B[Model output: draft text or leads]
    B --> C{Verification step}
    C -->|Case or citation| D[Check on Indian Kanoon, SCC Online, or citator]
    C -->|Date or deadline| E[Recompute by hand against a calendar]
    C -->|Translation or summary| F[Check against source text or fluent reviewer]
    D --> G[Corrected draft]
    E --> G
    F --> G
    G --> H[Filed document or client communication]

A prompt that skips straight from B to H is the pattern that put six fabricated citations into an NCLT order in Pooja Ramesh Singh. The verification step is not optional polish; it is the difference between a working draft and a document you can put your name to.

Frequently asked questions

Yes, for tasks where the model works from text you supply rather than memory of case law. For finding new authority or checking whether a judgment is still good law, treat any output as an unverified lead and confirm every case name and citation independently before it enters a filed document.

What exactly went wrong in the Supreme Court case about AI-generated citations?

In Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. & Anr., 2026 INSC 668, decided 2 July 2026, the NCLT relied on six precedents that were fake, wrongly cited, or attributed with paragraphs that do not exist. The Supreme Court set both orders aside. Full detail is at AI-generated case law: the Supreme Court’s 2026 rule.

No. The Supreme Court in Pooja Ramesh Singh did not prohibit AI tools; it held that citing AI-generated precedents without verification is misconduct. The distinction is between drafting with a tool and putting unverified output before a court as checked fact.

How do I verify whether a case name an AI model gave me is real?

Search the exact case name and citation on Indian Kanoon or a paid database such as SCC Online. If it does not appear, it is likely fabricated. If it does, open the judgment and confirm the attributed paragraph is genuinely there.

Why does a citation existing not mean the quoted paragraph is real?

A model that fabricates content does not necessarily fabricate the whole citation. Two of the six fake authorities in Pooja Ramesh Singh were genuine Supreme Court judgments, but the quoted paragraphs did not exist. Verifying a case exists is the first check, not the last.

Can AI reliably calculate a limitation period?

Not reliably enough to use without checking. A general model can misidentify which Article of the Schedule to the Limitation Act, 1963 applies, and its arithmetic is prone to off-by-one errors. Always recompute the final date yourself against the Schedule text.

Should I paste confidential client information into a general AI chatbot?

Treat this separately from accuracy. Consumer AI tools may use submitted text for training or store it outside India, raising confidentiality and Digital Personal Data Protection Act, 2023 concerns. Check the tool’s data handling terms before pasting unredacted client facts.

Asking the model to name supporting case law, then using the answer without opening the judgment. This is the pattern the Supreme Court addressed in Pooja Ramesh Singh: the fabricated citations were not caught before reaching the NCLT’s order, and unwinding that cost two orders set aside a year later.

Can I use AI to draft a full written submission without giving it any authorities?

You can ask it to draft the structure and prose, but do not let it select the authorities. The written submissions prompt here has you supply every case name and proposition; letting a model choose authorities recreates the risk behind the NCLT order that was set aside.

How do I check an AI translation of a vernacular court order without knowing the source language myself?

Have someone with genuine fluency, ideally with legal training, review it, particularly any paragraph marked uncertain and the operative paragraph containing the actual order. Do not rely on a second AI tool to check the first one’s translation.

Is a chronology built from AI-extracted dates safe to file in court without checking?

No. Treat it as a first draft that speeds up reading a document bundle, not a verified filing. Ambiguous date formats, particularly DD/MM versus MM/DD, are a common error source, and duplicate entries for one event across two documents can distort the sequence unmerged.

What is the difference between distinguishing a case and finding a counter-case?

Distinguishing works from the text of the adverse judgment you already have; finding a counter-case asks the model to name new authority from memory, the higher-risk task. Treat any counter-case named as an unverified lead needing the same check as any other citation.

Grounding a system in a real corpus of Indian judgments removes the failure mode of inventing a case never decided, since it can only retrieve documents that actually exist. You should still confirm a retrieved citation matches the proposition relied on. Niyam’s good-law checking applies this against a live database rather than a model’s memory.

Is it safe to rely on AI to identify which statutory provision governs a set of facts?

Use it to narrow the search, not conclude it. The BNS, BNSS, and BSA replaced the IPC, CrPC, and Evidence Act from 1 July 2024, and a model trained on pre-2024 material can default to old section numbers. Always confirm the current text against India Code.