Verify Everything — The Four-Step Check and the Trail
The course's spine, drilled to reflex: exists, says that, still good law — and the fourth step everyone skips — with a verification trail that would survive a judge's question, because in India it now may have to.
The hook
“Three confident authorities are on screen — produced minutes ago by the pipeline you built last week. Before anyone checks anything, commit: how many will survive? By the end of this hour you will have a number, a method, and a professional reflex you will use for the rest of your career — and you will know why the most dangerous authority is not the one that does not exist.”
What you'll be able to do
- Run the four-step authority check as a drilled habit — identify; verify it exists; confirm it says that, at the pinpointed paragraph; check it is still good law — on everything a model hands you, including the critic's counter-authorities.
- Classify failures into the three classes in ascending order of danger — does not exist; exists but says something else; exists, says it, no longer good law — and name the real Indian matter that exemplifies each.
- Trace any claim to its origin — the SIFT reflex: stop at the emotional spike, read laterally, find the primary record — whether the claim arrives from a chatbot, a headnote or a WhatsApp forward.
- Keep a verification trail another person could repeat — tool, query, what was checked, verdict, what was rejected — and say why that trail is about to acquire legal significance in India.
On the syllabus
- The reveal: the room predicts survival, then the four steps run live on the pipeline's own authorities
- The four steps: identify — exists — says that — still good law; the fourth is the one everyone skips
- The three failure classes, each with its real Indian case — in ascending order of danger
- Trace to origin: the SIFT reflex for forwards, headnotes and models alike
- Step four without a subscription: citators, and the honest limits of the free workaround
- The trail: what it contains, whom it protects, and why it is becoming a legal artefact in India
In short
Session 6 is the course's spine made explicit and drilled to reflex. The four-step check arrives not as advice but as choreography — identify, exists, says that, still good law — run live against the room's own prediction on authorities the class's own pipeline produced. The three failure classes are taught through real Indian matters: the invented case (the trial court's four phantom SCC citations in Gummadi), the real case with invented content (the Delhi High Court petition quoting paragraphs 73–74 of a judgment that has 27), and the case that was good law until it was not — the class everyone misses, because a case that exists and says the right thing feels finished. The free-tier reality is handled honestly: professional citators are step four's proper tool, and where students lack them, Indian Kanoon's citing-references plus a targeted overruling search is the poor lawyer's citator, with its limits stated. The hour closes on the trail — what it contains, how it protects you, and why India's draft court rules and the BCI's post-Pooja mandate are about to turn 'I checked' from a virtue into a record you produce.
Why it matters for using AI well
One workflow on autopilot and one instinct that never switches off: identify, exists, says that, still good law — logged as you go. It is the single most employable habit this course installs: it is what a journal's editorial board calls cite-checking, what a senior calls a junior worth keeping, and in the 2026 regulatory weather, what makes you the person nobody has to worry about.
What you can do on Monday
Take the last piece of work you submitted that cited authority — a moot memorial, a project, an internship note — and run all four steps on its three most load-bearing citations, with a trail. Whatever you find, you will never skip step four again.
What they leave with
The skill
Four steps and a trail: identify, exists, says that, still good law — every authority, every time, logged so another person could repeat the check.
The insight
The failure classes are ordered by invisibility: fabrications announce themselves, misdescriptions hide in real documents, and dead law looks identical to living law — which is why the step everyone skips is the one that decides cases.
The moment they remember
The tally against the prediction. The room committed — 'two of three will survive' — and then watched the checking happen live: one authority dissolving at step two, one surviving until the paragraph was actually read aloud and said something narrower, one sailing through to fail at the citator. The room's optimism, quantified and beaten, is the lesson no lecture could land: the dangerous authorities are the ones that survive the first check. Students leave running step four on cases their own seniors cite — which is exactly the reflex the profession is currently discovering it lacks.
In this session
- 01
The check, as choreography rather than checklist. Step one, identify: make the model list every authority with full citation and the proposition it supposedly supports — the Session 2 contract already does this. Step two, exists: the citation goes into a grounded database — Indian Kanoon first, SCC Online or Manupatra where the institution provides them — and either resolves to a document or does not. Step three, says that: open the document, go to the pinpointed paragraph, and read — does it support the proposition, at that width? A case that 'broadly supports' a proposition stated absolutely has failed step three. Step four, still good law: the citator question — appealed, overruled, distinguished into irrelevance, or statutorily superseded?
- 02
The three failure classes, in ascending order of danger, each now with an Indian face. Class one, does not exist: the four phantom SCC citations the trial court relied on in the matter that became Gummadi — loud, checkable, and caught by step two. Class two, exists but says something else: the Delhi High Court petition that quoted 'paragraphs 73–74' of Raj Narain — a real, famous judgment with twenty-seven paragraphs — and fabricated extracts of four more real cases; invisible to step two, caught only by opening the text at step three. Class three, no longer good law: the authority that survives every step but the fourth — the Shafhi Mohammad problem, a real Supreme Court decision, correctly quoted, and overruled by Arjun Panditrao two years later. Class one embarrasses you; class two misleads a court; class three loses the case — and only class one is what most people mean when they say they checked.
- 03
Why the fourth step is the one people skip: a case that exists and says the right thing produces the feeling of completion, and nothing in the document announces its own overruling. The discipline is structural, not moral — the step happens because it is on the table you fill in, not because you remembered to be virtuous. This is also where the Ladder's legal reading completes: statement → fact → authority → binding authority → settled law, and each rung is one of the steps.
- 04
Step four's tools, honestly priced: SCC Online's citator and Manupatra's citation analysis are the professional instruments, and students who have campus access should form the habit there. Without them: Indian Kanoon's 'cited by' listing on a judgment, read newest-first for treatment; a targeted search of the case name with 'overruled' and with the doctrine's leading later cases; and — for anything the matter turns on — the honest admission that a free check of step four is a screen, not a certification. The workspace corollary from Session 4 completes the thought: your bundle does not know it has been overruled either; bundle currency is step four applied to your own files.
- 05
Trace to origin, as one reflex rather than a second method: the forwarded 'Supreme Court has just held' message, the headnote's confident proposition, and the model's fluent summary are the same epistemic object — a statement wearing authority's clothes. The moves: stop, especially at the emotional spike (outrage is a targeting signal); read laterally — what do other sources say about this source; and land on the primary record — the judgment on the court's site or Indian Kanoon, the order itself, the Gazette text. The classroom version is fast; the professional version is the same thing with a citation at the end.
- 06
Grounded tools do not exempt themselves: the Stanford benchmark's finding — professional legal-AI tools hallucinating on one query in six to one in three — is on the syllabus precisely so that 'but the tool is citator-backed' never ends a verification. And the course's free grounded workhorse now carries its own AI layer: Indian Kanoon's premium tier sells generative answers and even outcome prediction on top of its database — grounding raises the floor; it never removes the duty; the duty is yours, whatever the tool cost.
- 07
The trail, specified like a document you might one day produce: for each authority — the tool and query used, the date, what was found, the step-three paragraph read, the step-four treatment found, the verdict (verified / corrected / rejected), and what replaced anything rejected. Ten minutes of logging that converts 'I used AI responsibly' from a character claim into a record — the difference, in the reported cases, between a professional and a cautionary tale. The protective logic is the same one lawyers already know from attendance notes: the trail is what you have when memory and good faith are not enough.
- 08
Why the trail is acquiring legal edge in India, stated precisely and dated: Pooja Ramesh Singh makes citing unverified AI authority misconduct — which makes the record of verification the natural defence; the Supreme Court's draft Regulations for AI in courts (June 2026, draft as of the date checked) would require parties to disclose AI assistance in prepared material, with responsibility expressly non-delegable; and the BCI, directed in Pooja to frame guiding principles and disciplinary consequences, was constituting its committee as this course went to press. Each cohort re-checks what has been notified — and the course's own habit of dating every regulatory claim is, itself, step four applied to the law about the tools.
The mirror
Run it on the class. Then on the machine.
An experiment on the room, the same effect explained in the model, a live demonstration on a real tool, and a named takeaway skill.
Why this shape
The authority-check reveal is one of the course's designed moments and it must stay an experiment: the room commits to how many authorities will survive before the checking starts, and the gap between prediction and tally is what installs the habit. Craft sessions build; this session confronts — and then drills.
The authority-check reveal
Experiment on the classOn the class
Three authorities produced by last session's pipeline go on screen, full citations, confident propositions. The room commits on their phones: how many survive all four steps — and which one fails, if any?
In the model
The three failure classes are baked into how a predictor works: it composes citations (class one), widens propositions toward the statistically common phrasing (class two), and knows nothing after its training about what has been overruled (class three).
Live on the model
The presenter runs all four steps on each authority, live — Indian Kanoon, the opened paragraph read aloud, the citing-references screen — and the tally lands against the room's prediction. The room is almost always too optimistic, and the survivor that fails at step four is the moment the session exists for.
The skill
Identify, exists, says that, still good law — every authority, every time, and especially the ones that feel finished after step three.
The case that exists but says something else
DemonstrationThe room predicts
A real judgment's name and a confident quoted 'extract' go up. The room votes: real quotation or not? The citation, everyone can verify, is genuine — which is exactly why the room splits.
What is going on
Class two is the predictor's signature failure: real case, plausible paragraph, invented content — fluent enough to survive any check that stops at existence.
Shown live
The judgment opens live; the pinpointed paragraph is read aloud; it says something else — or does not exist at that number, as in the Delhi High Court matter where 'paragraphs 73–74' were quoted from a twenty-seven-paragraph judgment.
The skill
Step two proves the case; only step three proves the proposition. Open the paragraph and read it — the quotation is not evidence of itself.
Trace the forward
Experiment on the classOn the class
A forwarded 'The Supreme Court has just held…' message — engineered to outrage — goes up, and the room votes fast: share, sit on it, or check? Then: what is the first check?
In the model
The forward, the headnote and the chatbot answer are the same object: a statement dressed as authority, optimised for transmission rather than truth — and models trained on the internet inherit precisely that optimisation.
Live on the model
The class runs the trace together: lateral search on the claim, the news cycle found or not found, and the landing on the primary record — the order itself, or the demonstrated absence of one.
The skill
Stop at the spike, read laterally, land on the primary record. The stronger your reaction, the earlier the checking starts.
The legal thread
Candour to the tribunal is the duty this session serves: you certify what you file, and since Pooja Ramesh Singh the certification of an unverified AI authority is not a risk but a defined misconduct. The trail is the duty's working form — and with the draft court Regulations proposing mandatory AI-disclosure and the BCI framing disciplinary norms, the record of checking is on its way from good practice to expected artefact. Verified as of August 2026; the session re-dates this slide each cohort.
Hands-on · 30 minutes · on your own laptop
The Authority-Check Drill
The drill that makes it a reflex, run on your own material: the verification work list your Session 5 pipeline produced — your draft's authorities and your critic's counter-authorities — through all four steps, with a trail another person could repeat. Your variant is built in: nobody else has your work list.
1 · Watch — the instructor demonstrates
On
Indian Kanoon (no account needed) in one tab; the judgment PDFs in another; the trail spreadsheet on screen
The exact prompt
No prompt — this lab's tool is the database. The trail template, filled live for one authority: Authority | Proposition claimed | Step 2: query used + found/not found (date) | Step 3: paragraph opened + supports/narrower/other | Step 4: citing references read + treatment found | Verdict: verified / corrected / rejected | If rejected: replaced with what.
Point at
The speed — a competent four-step check on one authority takes under four minutes, and the room should see that timed. Then the verdict language: 'corrected', not just 'wrong', when a real case supports a narrower proposition — and what the corrected entry now usefully says.
Roughly what comes back
Of the demo's three authorities (seeded from the Session 5 runs): one verified with its pinpoint, one corrected to a narrower proposition, one rejected — at whichever step it actually fails on the day. The trail row for each, filled in real time, is the artefact.
If it misbehaves
If Indian Kanoon strains under sixty simultaneous users, the direct document URLs for every seeded authority are on the slide — searching is the step that parallelises badly, so the links bypass it; the reading and the citator screen still happen live.
2 · Your turn — a variant, not a copy
Your turn, on your own work list: run all four steps on at least four authorities — your draft's two most load-bearing, and your critic's two counter-authorities — logging the trail as you go. Classify every failure into its class. If everything survives, run the fourth step one level deeper: find the most recent judgment citing your best authority and confirm the treatment is still approval.
Free tier
No AI quota consumed at all — this lab runs on Indian Kanoon (free, no account) and the documents. The only shared-resource risk is sixty simultaneous searches; the direct-URL sheet removes it. Students with campus SCC Online or Manupatra run step four there and note the difference in what the citator shows them.
3 · The reveal
The room's aggregate goes up by failure class — existence failures, says-something-else corrections, step-four casualties — beside the Session 6 opening prediction, now thoroughly refuted. Then one number stays on screen: how many students' critics had fabricated a counter-authority. The machine that attacked your argument needed checking too, which is the whole course in one statistic.
Deliverable
A four-row (or more) verification trail — authority, steps, verdicts, corrections — complete enough that a stranger could repeat every check. Feeds the Hallucination Audit directly: this is the same artefact at half scale.
Run of show · 30 minutes
- 0–10 min — Watch: three authorities through all four steps, timed, with the trail filled live.
- 10–24 min — Your turn: four authorities from your own work list through the four steps, trail logged, failures classified.
- 24–28 min — Deepen: one step-four treatment check on your best surviving authority.
- 28–30 min — Reveal: class-wide failure tallies by class; fabricating-critic count; the prediction from the top of the hour, revisited.
For the instructor · before the session
- Seed the demo set from the actual Session 5 runs (the class's own authorities land harder than a prepared list); verify each yourself the night before so the reveal has no surprises you don't control.
- Build the direct-URL sheet for every seeded authority and the practice fallbacks; test Indian Kanoon from the venue network.
- Prepare the trail template as a copyable spreadsheet and post it before the session.
- Confirm the Greenopolis extracts and the Raj Narain paragraph count against sources once more before teaching class two from it.
- Check the draft Regulations' status the week of the session — if anything has been notified, the closing slide changes and says so.
- Launch the Session 6 poll deck: the opening survival prediction, the real-quote vote, the forward vote, and the closing tallies.
Key sources & cases
Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd., 2026 INSC 668 (SC, 2 July 2026)
The stakes, now doctrinal: unverified AI citations are advocate misconduct; a decision resting on hallucinated material is no decision; the BCI directed to frame disciplinary consequences. The six defective authorities in the NCLT's order — three non-existent, three real-but-misdescribed — are the failure classes appearing in the wild, in one order. Verified to the judgment text 2026-08-27.
Greenopolis Welfare Association v. Narender Singh (Delhi HC, 25 Sept 2025)
Class two's Indian face: a petition quoting 'paragraphs 73–74' of Raj Narain v. Indira Nehru Gandhi — a judgment with twenty-seven paragraphs — plus altered extracts of four real authorities; withdrawn once exposed. Secondary sources only (Bar & Bench; Business Standard); the W.P. number remains untraced — VERIFY TO SOURCE and treat details as reported, not found.
Shafhi Mohammad v. State of Himachal Pradesh (2018) 2 SCC 801, overruled by Arjun Panditrao Khotkar (2020) 7 SCC 1
Class three's teaching pair, drawn deliberately from the Session 4 bundle: a real Supreme Court decision, quotable and quoted, that was good law for two years and then was not. Nothing in its text announces this. Arjun Panditrao verified 2026-08-26; Shafhi enters the codebook with this session — VERIFY TO SOURCE.
Magesh et al., 'Hallucination-Free?' (JELS 2025); Indian Kanoon PRISM (premium AI tier, checked 2026-08-27)
Why no tool ends the duty: professional grounded tools hallucinated on ~17–33% of queries in the Stanford evaluation (vendors contested the methodology; teach the dispute too); and the course's free grounded database now sells a generative layer — including outcome prediction — on its premium tier. Grounding raises the floor; the duty stays yours.
Mike Caulfield, SIFT / Web Literacy for Student Fact-Checkers (CC BY 4.0)
The trace-to-origin reflex's source, compressed in this course to its load-bearing moves: stop at the spike, read laterally, land on the primary record. Attribution required by licence; the four moves' wording verified against the source before any slide quotes it.
Supreme Court of India, draft Regulations for Use of Artificial Intelligence in Courts, 2026 (published for consultation, June 2026)
The trail's coming legal edge: draft reg. 43 would require disclosure of AI assistance in material prepared for a court, with responsibility non-delegable; the profession's association objected as unworkable; status as of 2026-08-27 — draft, not notified. RE-CHECK EACH COHORT: this slide is rewritten the day any version is notified.
Citators: SCC Online (Note Up), Manupatra citation analysis; Indian Kanoon 'cited by' as the free screen
Step four's instruments, priced honestly: professional citators where campus access exists; the free workaround — citing-references read newest-first plus targeted overruling searches — taught with the explicit caveat that it is a screen, not a certification. Feature names re-verified at teaching time.
Readings
- Pooja Ramesh Singh, 2026 INSC 668 — paragraphs 7, 9, 15 and 17: the holding, the BCI direction, and the six authorities
- Arjun Panditrao Khotkar (2020) 7 SCC 1 — the paragraphs overruling Shafhi Mohammad; watch law die on the page
- Magesh et al., 'Hallucination-Free?' (JELS 2025) — abstract and the vendor-response coverage
- Caulfield, SIFT — the four moves, one page
- The Supreme Court's draft AI-in-Courts Regulations (June 2026) — reg. 43, and the reported objections to it
Sixteen hours, one professional discipline.
Using AI well is not a knack — it is craft, competence and verification, practised until they are habits you could defend in court.