# AI output verification workflow: a 5-step bench protocol **TL;DR:** Verifying AI-drafted legal work is a five-step check you run on every citation before it leaves your desk: confirm the case exists in a primary source, confirm the quoted passage says what the draft claims, confirm the case is still good law, confirm any statute section is current under BNS/BNSS/BSA numbering, and confirm the proposition survives if you drop the weakest citation. Each step has a fixed home in a real database, a realistic time cost, and a place to record that you did it. --- ## On this page - [Why a checklist, not judgment](#why-a-checklist-not-judgment) - [Before you start: what this protocol assumes](#before-you-start-what-this-protocol-assumes) - [Step 1: does the cited case exist](#step-1-does-the-cited-case-exist) - [Step 2: does it say what the draft claims](#step-2-does-it-say-what-the-draft-claims) - [Step 3: is it still good law](#step-3-is-it-still-good-law) - [Step 4: are the statute sections current under BNS/BNSS/BSA](#step-4-are-the-statute-sections-current-under-bnsbnssbsa) - [Step 5: does the proposition survive without the weakest citation](#step-5-does-the-proposition-survive-without-the-weakest-citation) - [What to do when a citation cannot be found](#what-to-do-when-a-citation-cannot-be-found) - [How long the full protocol actually takes](#how-long-the-full-protocol-actually-takes) - [Where to record that you did it](#where-to-record-that-you-did-it) - [Verified vs unverified: what each looks like in a draft](#verified-vs-unverified-what-each-looks-like-in-a-draft) - [A verification log template](#a-verification-log-template) - [Common failure patterns and how each step catches them](#common-failure-patterns-and-how-each-step-catches-them) - [Training a junior to run this without supervision](#training-a-junior-to-run-this-without-supervision) - [Frequently asked questions](#frequently-asked-questions) --- ## Why a checklist, not judgment An AI-drafted note, opinion, or set of pleadings does not need a senior lawyer's discretion to be checked. It needs a fixed sequence that a junior can run the same way every time, on a Sunday night, without asking anyone what "verify" means. This post sets out that sequence: five steps, each with a specific source to open, a specific question to answer, and a specific amount of time it should take. The reason this has to be procedural rather than a matter of feel is stated plainly by the Supreme Court of India in *Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. & Anr.*, 2026 INSC 668, Civil Appeal No. 11950 of 2025, decided on 2 July 2026. The bench of Justice Pamidighantam Sri Narasimha and Justice Alok Aradhe held that courts must adopt a zero-tolerance mode for producing, citing, or using AI-generated precedents without verification, that citing such material without verification is misconduct on the part of an advocate, and that a decision resting on fake or hallucinated material is no decision in the eyes of the law even if only an iota of it entered the reasoning. Three of the six citations relied on by the National Company Law Tribunal in that case did not exist at all. Two more were real, correctly cited Supreme Court judgments with quoted paragraphs that do not appear anywhere in them. That second category is the one a casual check misses, which is why this protocol runs five separate steps rather than one. The full holding and the six-citation table are covered in [the Supreme Court's 2026 rule on AI-generated case law](/blog/sc-ai-generated-case-law-2026). The professional duty behind the checking obligation is set out in [a lawyer's duty to verify AI output](/blog/lawyer-duty-verify-ai-output). What follows is the bench routine: the steps in order, what each one involves, how long it takes, and where the record of having done it lives. Bharatiya Nyaya Sanhita, 2023 (BNS), Bharatiya Nagarik Suraksha Sanhita, 2023 (BNSS), and Bharatiya Sakshya Adhiniyam, 2023 (BSA) came into force on 1 July 2024, replacing the Indian Penal Code, 1860, the Code of Criminal Procedure, 1973, and the Indian Evidence Act, 1872. Any AI-drafted note that cites a criminal-law section needs its own check against that transition, which is Step 4 below. --- ## Before you start: what this protocol assumes This is a verification workflow for output that an AI tool has already produced: a memo, a set of proposed citations, a first-draft argument, a summary of a judgment. It assumes you did not write the draft from primary sources yourself and are now checking someone else's, or something else's, work before you put your name on it. It also assumes a specific division of failure modes, because that division is what the five steps are built around. A citation can fail in three ways: the case does not exist, the case exists but the quoted text does not, or both exist but the case is no longer good law. A statute reference can fail in one further way unique to Indian practice right now: the section number is correct for the old code and wrong for the current one, because BNS, BNSS, and BSA renumbered the criminal statutes on 1 July 2024. Steps 1 through 4 map onto these four failure modes directly. Step 5 checks the argument as a whole once the first four steps have done their work. Run the five steps in order. Step 2 only makes sense once Step 1 has confirmed the document exists to open. Step 3 only matters for a case that has cleared Steps 1 and 2. Doing them out of order wastes time on questions that a cheaper, earlier check would have made moot. --- ## Step 1: does the cited case exist **The question.** Is there a real, reported judgment matching the case name, court, and year the draft gives you. **Where to check it.** For Supreme Court judgments, use the [e-SCR portal at scr.sci.gov.in](https://scr.sci.gov.in/), which publishes judgments in the neutral citation format the Court itself assigns, or the [Supreme Court of India's own judgment search](https://www.sci.gov.in/). For High Court and subordinate court judgments, use [judgments.ecourts.gov.in](https://judgments.ecourts.gov.in/), the National Judicial Data Grid's judgment search. [Indian Kanoon](https://indiankanoon.org/) is a fast third option that indexes a broad range of Indian courts and tribunals, useful as a first pass, though a citation load-bearing enough for a bench needs confirmation against the court's own portal too. **How to run it.** Search by party names first, then again by the citation string alone. The second search matters more than it looks. In *Pooja Ramesh Singh*, one of the six fabricated citations gave a correct reporter reference, 2020 SCC OnLine SC 341, attached to the wrong case name. A name search returned nothing. A citation search returned a real judgment between entirely different parties, M. Subramaniam v. S. Janaki, and revealed the citation had been repurposed. Searching only one way would have missed it. **What counts as a pass.** The case name, court, and year in the draft match a document you can open on one of the sources above. A pass here does not mean the citation is usable. It means the document exists. Steps 2 through 4 tell you whether it is usable. **What counts as a fail.** No match on either a name search or a citation search, across all relevant sources. Three of the six citations in *Pooja Ramesh Singh* failed at exactly this step: State Bank of India v. M/s Shree Ram Urban Infrastructure Ltd., ICICI Bank Ltd. v. Urban Infrastructure Real Estate Ltd., and V.S. Dempo & Co. Ltd. v. Reliance Communications Ltd. did not exist under those names at all. This is the fastest failure mode to catch, and should be the first thing you check for exactly that reason. **Time.** Two to four minutes per citation on a functioning search. A citation that returns no result on the first search and needs a second, citation-string search takes closer to five. --- ## Step 2: does it say what the draft claims **The question.** Open the actual judgment. Find the specific paragraph the draft is relying on. Does that paragraph exist in those words, and does it support the proposition the draft attributes to it. **Why this is a separate step from Step 1.** Two of the six citations in *Pooja Ramesh Singh* passed Step 1 outright. Everest Kento Cylinders Ltd. v. Union of India, (2015) 2 SCC 1, and Canara Bank v. N.G. Subbaraya Setty & Anr., (2018) 16 SCC 228, are genuine, correctly cited Supreme Court decisions. The paragraphs the National Company Law Tribunal attributed to them do not exist in either judgment. A verification process that stops once the case is confirmed real passes both of these and misses the fabrication, because the invented text was inserted into a real authority rather than attached to a fake one. **How to run it.** Open the full text on the same source you used in Step 1, not a summary or a headnote from a search engine's preview. Locate the paragraph number the draft cites. Read it, then read the paragraph before and after it, because a quotation taken out of surrounding context can be technically present in the judgment and still misrepresent what the court held. Check the paragraph count of the judgment against the paragraph number claimed. A cited paragraph 73 in a judgment that has 27 paragraphs is an immediate fail, the exact pattern in a documented Delhi High Court petition drafted with a chatbot that quoted paragraphs 73 and 74 of Raj Narain v. Indira Nehru Gandhi, a judgment running to 27 paragraphs. **What counts as a pass.** The paragraph exists, in substantially those words, and a plain reading of it supports the proposition the draft is using it for. **What counts as a fail.** The paragraph does not exist in the judgment at all, or it exists but says something narrower, different, or contrary to what the draft claims. Both are fails. Do not distinguish between "wrong paragraph number" and "wrong content" for logging purposes; either one means the citation cannot be used as drafted. **Time.** Five to ten minutes per citation for a straightforward statutory or contract point. Fifteen to twenty minutes for a constitutional or multi-issue judgment where the relevant reasoning is spread across several paragraphs and you need to read enough surrounding text to be confident about context. --- ## Step 3: is it still good law **The question.** Has this judgment been overruled, reversed on appeal, referred to a larger bench that has since decided differently, or so consistently distinguished that the ratio you are relying on no longer has force. **Why it is a separate step from Steps 1 and 2.** A citation can pass both existence and paragraph-accuracy checks and still be dead law. A real judgment, correctly quoted, that a larger bench overruled two years later is not safe, and a court will treat citing it without disclosure as a gap in candour rather than an honest oversight. This step is independent of whether the AI hallucinated anything; a human-drafted brief with a stale citation fails it the same way. **How to run it.** The mechanics of a good-law check, including the full taxonomy of overruled, distinguished, doubted, per incuriam, and statutorily superseded, are set out in [what "good law" means and how to check it](/blog/good-law-checking). In short: check bench strength (a two-judge bench decision yields to a conflicting three-judge bench decision of the same court), check whether the case has been cited approvingly or adversely in recent decisions, and check whether the statutory provision it interpreted has since been amended. A citator built into a legal research tool can surface this signal quickly; a general-purpose AI chatbot has no reliable mechanism for this, because it has no structured record of subsequent judicial treatment to draw on. **What counts as a pass.** No adverse treatment found, the bench strength is not undercut by a later, larger-bench decision on the same point, and the statutory provision interpreted is unchanged or the ratio survives any amendment. **What counts as a fail.** The case has been overruled, reversed, or is the subject of an unresolved reference to a larger bench where the more recent, more authoritative view cuts against your proposition. A case that has been distinguished heavily on facts closely resembling yours is a soft fail: usable with disclosure, not usable as unqualified authority. **Time.** Five minutes for a routine check on a well-known, unambiguous case. Fifteen to thirty minutes for a case in an area of law that has moved recently, or where a citator flags any adverse treatment and you need to read the later judgment yourself to understand its scope. --- ## Step 4: are the statute sections current under BNS/BNSS/BSA **The question.** If the draft cites a criminal-law section, is that section number correct under the code presently in force, or is it a pre-1 July 2024 IPC, CrPC, or Indian Evidence Act number that an AI tool carried over from its training data without updating. **Why this is its own step.** BNS, BNSS, and BSA came into force on 1 July 2024, and the renumbering is not a simple one-to-one shift. Some numbers were reused for a completely different offence. Section 144 of the old CrPC governed the power to prohibit unlawful assemblies; Section 144 of BNSS governs maintenance for wives, children, and parents, the direct successor to CrPC Section 125. Section 482 of the old CrPC was the High Court's inherent powers provision; Section 482 of BNSS is anticipatory bail, the successor to CrPC Section 438, while inherent powers moved to BNSS Section 528. An AI model trained substantially on pre-2024 material can produce a section number correct under the old code and wrong under the new one. That is not a hallucinated case; it is a real, existing number that means something else now, which is why a quick glance will not catch it. **How to run it.** Check the current text of any cited section against the [official electronic versions of Central Acts on India Code](https://www.indiacode.nic.in/), maintained by the Ministry of Law and Justice, which carries the current, amended text of BNS, BNSS, and BSA alongside the repealed IPC, CrPC, and Indian Evidence Act for reference. Confirm the section number, read the operative text, and confirm it matches what the draft says the section covers. Where the draft cites an old IPC, CrPC, or Evidence Act section for a matter arising after 1 July 2024, flag it for renumbering before doing anything else with it. For sections that changed number, and sometimes subject matter entirely, see [BNSS Section 528 on inherent powers](/blog/bnss-section-528-inherent-powers) (successor to CrPC 482), [anticipatory bail under BNSS Section 482](/blog/anticipatory-bail-bnss-section-482) (successor to CrPC 438), [default bail under BNSS Section 187](/blog/default-bail-bnss-section-187) (successor to CrPC 167), [maintenance under BNSS Section 144](/blog/maintenance-bnss-section-144) (successor to CrPC 125), [BNS Section 152 on sedition](/blog/bns-section-152-sedition) (successor to IPC 124A), and [BSA Section 63 on electronic evidence](/blog/bsa-section-63-electronic-evidence) (successor to Evidence Act Section 65B). **What counts as a pass.** The section number, the Act, and the operative text all match the current statute, and if an old-code reference is retained for historical accuracy, it is clearly labelled as such alongside the current number. **What counts as a fail.** A section number is stated without specifying which Act it belongs to, an old-code number is used for a matter governed by the new codes without cross-reference, or the section cited under its current number does not actually say what the draft claims. **Time.** Two to three minutes per section reference for a section you already know the mapping for. Five to ten minutes for an unfamiliar section, since you need to pull the current text and confirm it against the drafted proposition, not just confirm the number exists. --- ## Step 5: does the proposition survive without the weakest citation **The question.** After Steps 1 through 4 have been run on every citation supporting a given proposition, remove the weakest one, the one you had the most doubt about or the one that only barely passed. Does the argument still stand on the remaining authority. **Why this step exists.** Steps 1 through 4 are citation-level checks. This step is proposition-level. An argument built on four citations, one of which is marginal, is fragile in a way individual citation checks will not reveal, because each citation can technically pass its own check while the argument as a whole leans on the weakest link. This is also the step that answers a question opposing counsel will ask directly: if I take away your worst authority, what is left. If the honest answer is "not much," the proposition needs a stronger foundation, not a fifth citation bolted on to compensate. **How to run it.** List every citation supporting the proposition, ranked by strength after Steps 1 through 4. Strike the weakest one, mentally or literally, and read the proposition again with it removed. If it still holds, the proposition is sound and the marginal citation can stay as supporting authority or be dropped for cleanliness. If it collapses, find a stronger authority to replace the weak link, or reduce the proposition to what the remaining, solid citations actually support. **What counts as a pass.** The proposition survives on the remaining authorities, even if in a narrower form than originally drafted. **What counts as a fail.** The proposition depends entirely on the citation you just removed. This is not, by itself, a reason to abandon the point; it is a reason to either find a better authority or narrow the claim to what the evidence actually supports. **Time.** Five to fifteen minutes per proposition, once the individual citations underneath it have already cleared Steps 1 through 4. This step is cheap precisely because it reuses work already done; it does not require opening any new documents, only re-reading what you have already verified. --- ## What to do when a citation cannot be found A citation that fails Step 1 needs a decision, not a shrug. Three outcomes are available, and the correct one depends on what the citation was doing in the draft. **Delete the citation and the sentence that depends on it.** This is the default and, in most cases, the correct response. If a proposition rests on a citation that does not exist, the proposition as drafted has no support. Removing it is not a concession; it is what Step 5 would have forced you to do anyway once the weak link was gone. **Search harder before deleting, but only once.** A name search returning nothing does not always mean the case does not exist; it sometimes means the case name in the draft is slightly wrong. Run one further citation-string search, as described in Step 1, and check for common transcription errors: a swapped party order, a citation year off by one, a reporter abbreviation error. If a second, targeted search still returns nothing, treat it as non-existent and move to deletion. **Escalate rather than substitute silently.** If the proposition genuinely needs authority and none of your own research turns up a real case supporting it, that signals the proposition itself may be weaker than the draft assumed, not a licence to find any nearby case and cite it in place of the fabricated one. Flag it to whoever is supervising the matter before filling the gap. A hurried substitute citation, checked less carefully than the rest of the document, is exactly how a second error gets into a document that was supposed to have been cleaned of the first one. Never re-ask the AI tool that produced the citation whether the citation is correct. A model that fabricated a citation will, when asked to confirm it, frequently confirm it again; this was the second failure in the *Mata v. Avianca* sanctions matter in the United States in 2023, where counsel asked the tool that had invented the cases whether the cases were real, and it said yes. Verification has to run against an independent primary source, never against the tool that generated the claim. --- ## How long the full protocol actually takes For a single citation, running Steps 1 through 4 in full takes roughly fifteen to twenty-five minutes: two to four minutes for existence, five to ten for the paragraph check, five to fifteen for the good-law check, and two to ten for the statute-numbering check if a criminal-law section is involved. Step 5 runs once per proposition rather than per citation, after the citations under it have each been checked, and adds another five to fifteen minutes. For a draft with twenty citations across eight propositions, budget two to three hours for a complete first pass. It runs shorter with practice, once you can tell a routine citation (a well-known Supreme Court case cited for a proposition you have seen it support before) from one that needs the full treatment (an unfamiliar High Court decision, a section number you have not personally checked since the code changed, a citation that only barely survived Step 2). This is not materially longer than the manual research a diligent junior would have done before AI drafting tools existed. What has changed is the shape of the work: research time has fallen, verification time has not. Treating the speed gained at drafting as time saved overall is the mistake. It has moved to verification, and skipping it there is the exact failure the Supreme Court addressed in [*Pooja Ramesh Singh*](/blog/sc-ai-generated-case-law-2026). ```mermaid flowchart TD A[AI-drafted output with a citation] --> B{Step 1: case exists in a primary source} B -->|No| X1[Search citation string once more] X1 -->|Still not found| D1[Delete citation and dependent sentence] B -->|Yes| C{Step 2: paragraph exists and says what draft claims} C -->|No| D1 C -->|Yes| E{Step 3: still good law} E -->|Overruled or reversed| D2[Cite the later authority instead, or drop the point] E -->|Yes| F{Step 4: statute section current under BNS/BNSS/BSA} F -->|Old-code number used unlabelled| D3[Correct to current section, relabel old number if kept] F -->|Yes| G{Step 5: proposition survives without weakest citation} G -->|No| H[Strengthen with better authority or narrow the claim] G -->|Yes| I[Safe to file, record the check in the verification log] ``` --- ## Where to record that you did it If a bench, a supervising partner, or opposing counsel later asks how a citation reached your submission, "I checked it" is a weaker answer than a dated record showing what you checked, where, and when. Keep the record with the matter file, not in a separate system you will not think to open when the question comes up. A workable minimum: a single log entry per citation, appended to the matter file, recording the citation as given, the source checked, the result of each of the four citation-level steps, and the date. The template in the next section is built for exactly this. It needs to exist, be dated, and be attached to the specific document it verified. For matters where verification runs through a research tool such as [Niyam](https://niyam.ai) that already links every proposition back to the judgment it came from, the record can be lighter, because the tool's own citation trail does part of the logging work. That does not remove the requirement to read the underlying judgment yourself; it reduces the time Steps 1 and 2 take, a different thing from removing the obligation to run them. --- ## Verified vs unverified: what each looks like in a draft | Signal in the draft | Verified | Unverified | |---|---|---| | Case citation format | ✓ Full cause title, court, year, and neutral or reporter citation, checked against a primary source | ✗ Case name and year only, no reporter or neutral citation given | | Quoted paragraph | ✓ Paragraph number given and confirmed to exist in the judgment, in those words | ✗ A block quote with no paragraph number, or a paragraph number higher than the judgment's total | | Good-law status | ✓ Checked against a citator or recent judgments on the same point, dated | ✗ No mention of subsequent treatment, older case cited as though settled | | Statute section | ✓ Act named explicitly (BNS, BNSS, or BSA) with current section number, cross-checked against India Code | ✗ Bare section number with no Act specified, especially for criminal-law points | | Supporting authority for a proposition | ✓ Multiple citations, proposition survives with the weakest one removed | ✗ A single citation carrying the entire proposition, unexamined for whether it is the strongest available | | Record of the check | ✓ Dated verification log entry attached to the matter file | ✗ No record; "I checked this" with nothing to show for it | --- ## A verification log template Copy this into the matter file for each citation you verify. One block per citation. ``` Citation as drafted: [full case name, year, reporter/neutral citation as given by the AI tool] Proposition it supports: [one sentence] STEP 1 - Existence Source checked: [scr.sci.gov.in / sci.gov.in / judgments.ecourts.gov.in / indiankanoon.org] Result: [Confirmed real / Not found / Found under different cause title] Notes: STEP 2 - Paragraph accuracy Paragraph cited: [number] Total paragraphs in judgment: [number] Passage checked matches quoted text: [Yes / No] Passage supports the stated proposition: [Yes / No / Partially - narrower than claimed] STEP 3 - Good law status Citator or manual check used: [tool/source name] Bench strength: [e.g. two-judge / three-judge / Constitution Bench] Adverse treatment found: [None / Distinguished / Doubted / Overruled / Referred to larger bench] Notes: STEP 4 - Statute currency (if applicable) Act and section as drafted: [e.g. "Section 144" - specify which code] Current Act and section (per India Code): [e.g. BNSS Section 144, successor to CrPC Section 125] Confirmed current text matches proposition: [Yes / No] STEP 5 - Proposition survives without this citation: [Yes / No - see other citations under same proposition] Verified by: [name] Date: [DD Month YYYY] Outcome: [Safe to file / Deleted / Replaced with [citation] / Flagged for supervisor] ``` --- ## Common failure patterns and how each step catches them **A confident citation with no reporter reference.** AI tools frequently produce a case name, a year, and a court, with no neutral or reporter citation attached. This is not itself proof of fabrication, but it removes the fastest verification path. Ask for the citation string first; if the tool cannot supply one, that is itself informative. **A real case, a fake paragraph.** This is the failure that made *Everest Kento Cylinders* and *Canara Bank v. N.G. Subbaraya Setty* dangerous in [*Pooja Ramesh Singh*](/blog/sc-ai-generated-case-law-2026). Step 2 exists specifically because Step 1 alone will not catch it. **A citation that survives Steps 1 and 2 but is dead law.** A judgment can be entirely real and correctly quoted, and still have been overruled. This failure mode is equally common in ordinary, non-AI-assisted drafting, which is why Step 3 is basic practice that AI-assisted drafting has simply made more urgent to run every time. **An old-code section number carried over into a post-2024 matter.** A model trained substantially on pre-1 July 2024 material can produce a CrPC or IPC section number as though the codes were still in force. The renumbering trap is worse than a plain error because the number often exists and means something under the current codes too, just not what the draft claims. **A proposition standing on one thin citation, dressed up with three decorative ones.** Four citations look more authoritative than one. If three are marginal and only one is load-bearing, Step 5 surfaces that before a court does. --- ## Training a junior to run this without supervision The protocol above is written to be followed without judgment calls, which is the point: a junior should be able to run it correctly the first time, on an unfamiliar matter, without asking a senior colleague what "verify" means at each step. Three things make that possible in practice. Give the junior the exact sources named in Step 1, not "go check it," so there is no ambiguity about where a negative result on one source should trigger a second search on another. Require the verification log filled in as the checks run, not reconstructed afterward from memory, because a log written after the fact tends to record what the junior meant to check rather than what they actually found. Treat a Step 1 or Step 2 failure as routine rather than alarming; a junior who deletes an unverifiable citation without escalating is doing the job correctly, and treating every fail as a crisis teaches juniors to hide failures rather than report them. The underlying professional duty does not shift with seniority. The lawyer who signs the document owns every citation in it, described in full in [a lawyer's duty to verify AI output](/blog/lawyer-duty-verify-ai-output). What changes with a written protocol is that the duty becomes something a junior can discharge correctly on their own. A research tool that links every generated proposition back to the specific judgment it drew from removes a meaningful share of the search time in Steps 1 and 2, because the document is one click away rather than a fresh database search. [Niyam](https://niyam.ai) is built on a corpus of Indian judgments with source attribution on every retrieved answer, which shortens the existence and paragraph checks without removing the obligation to run them. The good-law check in Step 3 is faster with a citator that flags adverse treatment automatically, provided the flagged treatment is still read and assessed rather than accepted on faith. For the fuller mechanics behind each step, see [good law checking in Indian legal research](/blog/good-law-checking), [how to vet a legal AI tool for citation accuracy](/blog/how-to-vet-legal-ai-citation-accuracy), and [the Supreme Court's 2026 ruling on AI-generated case law](/blog/sc-ai-generated-case-law-2026). --- ## Frequently asked questions ### What is the single most important step in this protocol? Step 2, the paragraph check. Step 1 catches citations that do not exist at all, which is the easier failure to spot. Step 2 catches citations that are real but carry fabricated or mischaracterised text, which is harder to notice and was the failure mode behind two of the six citations in [*Pooja Ramesh Singh*](/blog/sc-ai-generated-case-law-2026). A verification process that stops at confirming a case exists will miss this category entirely. ### Do I need to run all five steps on every citation, or can I skip steps for well-known cases? Run Steps 1 through 4 on every citation you intend to file, including well-known cases, because Step 2 and Step 3 failures can occur even in familiar authorities if the AI tool has mischaracterised the holding or you are relying on it after a subsequent overruling you have not tracked. You can move faster through familiar cases, but skipping a step entirely defeats the purpose of having a fixed protocol. ### How is this different from ordinary legal research verification? It is not fundamentally different; it is the same due diligence competent practice has always required, made explicit and sequenced because AI-drafted output produces citations faster than a human researcher would, which increases the volume that needs checking and the temptation to check less carefully per item. Step 4, the BNS/BNSS/BSA numbering check, is the one addition specific to the current Indian legal moment, because the 1 July 2024 transition created a category of error that did not exist before it. ### What if the AI tool tells me it already verified the citation? Treat that claim the same as any unverified citation. A model that generated a fabricated citation has no independent mechanism for checking its own output against a primary source, and asking it to confirm its own work tends to produce a confirmation regardless of accuracy. Run Steps 1 and 2 yourself against a primary source, every time, regardless of what the tool claims about its own verification. ### How do I handle a citation where the case exists but I cannot access the full judgment text? Do not use it until you can. Try an alternative source from Step 1's list before concluding the text is inaccessible. If none of the listed sources have the full text, treat the citation as unverified for Step 2 purposes and either find an alternative authority or, where the citation matters enough, obtain a certified copy through the relevant court's registry. ### What counts as "the weakest citation" in Step 5? The citation you have the least confidence in after running Steps 1 through 4: the one from a lower court, the one where the paragraph only loosely supports the proposition, the one oldest relative to a fast-moving area of law, or the one where a good-law check surfaced adverse treatment short of outright overruling. If several are roughly equal in strength, remove the one easiest for opposing counsel to attack. ### Does this protocol apply to statutory provisions as well as case citations? Directly, in Step 4, for criminal-law provisions affected by the BNS/BNSS/BSA transition, such as [sedition under BNS Section 152](/blog/bns-section-152-sedition). For other statutes, the same discipline applies in principle: confirm the section is current, confirm it says what the draft claims, and confirm any amendment history that might affect the drafted proposition, checked against the current text on [India Code](https://www.indiacode.nic.in/). ### How do I verify a citation from a tribunal, like the NCLT or NCLAT, rather than a court? Tribunal orders are generally not indexed on the court-focused sources listed in Step 1. Check the tribunal's own website, and cross-check on [Indian Kanoon](https://indiankanoon.org/), which indexes tribunal decisions alongside court judgments. The good-law equivalent is confirming whether the order has been appealed to the NCLAT or, from there, to the Supreme Court, and what the outcome was, since reversed tribunal orders are common and are exactly the pattern at issue in *Pooja Ramesh Singh* itself. ### What should I do differently for a citation going into an urgent, same-day filing? Compress the time spent on each step rather than skipping any of them, and prioritise the propositions that carry the most weight. Steps 1 and 2 for the two or three citations the outcome actually turns on should never be skipped, even under deadline. If time genuinely does not permit full verification of every citation, flag the unverified ones to whoever is reviewing the filing before it goes out, rather than letting one pass silently. ### How does bench composition affect Step 3? Within the Supreme Court, a decision from a larger bench outweighs a conflicting decision from a smaller bench on the same point. When running Step 3, confirm the bench size of the case you are relying on, and check specifically whether a larger-bench decision exists on the same question. A two-judge bench decision that a later three-judge bench has taken a different view on needs disclosure of the conflict, not citation as though the position were settled. ### What is the difference between a citation being "distinguished" and being fabricated? A distinguished citation is a real judgment, correctly quoted, that a later court has held does not apply to different facts; it remains good law on its own facts and fails only Step 3, and only partially, since it can often still be cited with disclosure. A fabricated citation fails Step 1 (does not exist) or Step 2 (exists but the quoted content does not); it cannot be used in any form, and should be deleted rather than qualified. ### How often should I re-check a citation I verified previously, if I reuse it in a later matter? Re-run [Step 3, the good-law check](/blog/good-law-checking), before reusing any previously verified citation in a new matter, because the case may have been overruled or distinguished since. Steps 1, 2, and 4 generally do not need re-checking for a citation already personally verified against a primary source, unless significant time has passed or the statute involved has since been amended. ### Where does the duty to run this protocol come from if my firm has no written policy? The duty exists independently of any internal firm policy. It flows from the general professional duties of competence and candour to the court under the Advocates Act, 1961 and the Bar Council of India Rules made under it, and the Supreme Court restated it directly in [*Pooja Ramesh Singh*](/blog/sc-ai-generated-case-law-2026), holding that citing AI-generated precedents without verification is misconduct. A firm without a written verification policy does not have a lawyer without the duty; it has a lawyer whose duty has not yet been written down.