Training junior associates for legal research in the AI era

TL;DR: A junior who can prompt an AI tool but cannot independently trace a citation, read a judgment for its ratio, or tell a High Court decision from an SLP dismissal is not trained, whatever the output looks like. This piece sets out what a junior must do unaided before touching AI, a 90-day programme to build it, and how to review a research memo so the review itself becomes the training.

---

On this page

• Why the old training model stopped working
• What a junior must do unaided before touching AI
• The 90-day programme
• How to set a research task so the output is checkable
• Reviewing a junior’s research memo
• The verification habit
• When AI-assisted work is confidently wrong
• Measuring competence, not hours
• Frequently asked questions

---

Why the old training model stopped working

For decades, a junior associate’s judgement was built on drudgery. You were sent to the library or the digest volumes to find every case on a point, and in the process of reading forty judgments to find the three that mattered, you learned to recognise a weak ratio, an obiter remark dressed up as a holding, and a judgment that had been quietly distinguished into irrelevance. The grunt work was the curriculum. Nobody assigned it as training; it was simply how the research got done, and judgement was the byproduct.

AI tools do that grunt work now. A junior can generate a competent-looking answer to “find precedents on anticipatory bail in economic offences” in under a minute, complete with citations and summarised holdings. The byproduct is gone, but the need for the judgement it used to build has not gone anywhere. A senior still has to trust that when a junior tells them a point is settled, it is settled, and that trust has to be earned the same way it always was: by watching the junior get it right, and wrong, enough times to know the difference.

The Thomson Reuters 2026 Future of Professionals report, surveying more than 1,800 professionals across 62 countries, found that 91 percent say their organisation falls short of what AI could deliver. That gap is not a tooling problem in most firms. It is a training problem: people are using tools they were never taught to check. The Wolters Kluwer 2026 Future Ready Lawyer Survey, covering 810 lawyers across the United States, China, and nine European countries, found inadequate training was named as a barrier to AI adoption by 39 percent of respondents, tied with ethics and data privacy concerns at the same figure. Those numbers describe firms in the US, China, and Europe, not India, but the underlying mechanism, a firm handing juniors a capable tool without rebuilding the training that used to happen by accident, is not specific to any jurisdiction.

The stakes are concrete, not theoretical. On 2 July 2026 the Supreme Court decided Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. & Anr., 2026 INSC 668, a bench of Justice Pamidighantam Sri Narasimha and Justice Alok Aradhe setting aside NCLT and NCLAT orders because the tribunal had relied on precedents that were fake, wrongly cited, or contained paragraphs that do not exist in the real judgments. The Court held that citing an AI-generated precedent without verification is misconduct on the part of an advocate, and that a decision built on even a fragment of fake material is no decision in the eyes of the law. Somewhere in the chain that produced those citations was a person, junior or senior, who did not check. That is what training junior associates for the AI era has to prevent.

This is not an argument against juniors using AI tools. It is an argument that the unaided skill has to exist first, deliberately built rather than assumed, because the tool will not tell a junior when it is wrong, and a junior who cannot tell either is not a safeguard against that judgment, they are a conduit for it. Firms that have already written this down as a formal policy rather than an assumption can start from a law firm AI use policy for India or the broader legal AI rollout plan for a small Indian law firm; a training programme works better once it sits inside a stated policy rather than filling the gap left by having none.

---

What a junior must do unaided before touching AI

Before a junior is handed an AI research tool with any autonomy, they should be able to do the following without it. This is not a nostalgia list. Each item is a check the junior needs to run on AI output later, and a person cannot run a check they cannot perform themselves.

A junior solid across this table can be given AI tools with real latitude, because they will catch the tool’s mistakes before those reach a senior. A junior who is not should use AI for narrow, low-risk drafting scaffolding at most, with every substantive claim checked against a primary source. The programme below is how you get from the second state to the first.

---

The 90-day programme

Ninety days is long enough to see a junior work through several full research cycles and short enough that a firm can commit to it as a defined block rather than ad hoc supervision. The phases below describe the shape of a programme; specific case files and volume of work will vary with practice area and firm size.

flowchart TD
    A["Day 1-2: baseline skills check"] --> B["Week 1-2: unaided primary-source research"]
    B --> C["Week 3-4: memo writing without AI"]
    C --> D["Day 30: gate review<br/>pass or repeat phase 1"]
    D --> E["Week 5-6: AI-assisted research,<br/>every claim verified by junior"]
    E --> F["Week 7-8: AI-assisted drafting,<br/>senior spot-checks verification log"]
    F --> G["Day 60: gate review<br/>pass or repeat phase 2"]
    G --> H["Week 9-10: independent case ownership,<br/>AI as first pass only"]
    H --> I["Week 11-12: junior reviews another junior's memo"]
    I --> J["Day 90: competence assessment<br/>and sign-off"]

Days 1 to 2: baseline skills check. Before assigning any real matter, run the junior through a short unaided exercise built from the table above: find a section, read a judgment for its ratio, compute a limitation date, and identify whether a given case is still good law. This is diagnostic, not a pass or fail gate. It tells the supervising senior which of the ten skills need the most attention over the next four weeks, and it gives the junior a concrete starting point rather than a vague sense of being new.

Weeks 1 to 2: unaided primary-source research. The junior works real, low-stakes research questions, drawn from closed matters or genuine but non-urgent live questions, entirely without AI tools. Every citation gets pulled from the actual reported judgment or from a primary source like India Code, the Supreme Court of India website, or the eCourts case status system, not from a summary. This is deliberately slower than it needs to be. The point is not efficiency in weeks one and two; it is building the muscle that later lets the junior spot when an AI tool’s summary of a judgment has drifted from what the judgment actually says. A grounding in the difference between primary and secondary legal sources belongs here, alongside practical fluency with whichever case law search engine the firm uses, whether that is SCC Online, Manupatra, or Indian Kanoon, and an honest look at where a free tool like Indian Kanoon is strong and where it falls short is useful reading at this stage; see Indian Kanoon’s limitations for legal research. A junior should also be able to cite a judgment correctly by the end of this phase, since a memo full of correct research undermined by malformed citations is still not usable.

Weeks 3 to 4: memo writing without AI. The junior writes full research memos from that unaided research, in the firm’s standard structure, with no AI assistance on drafting either. This tests something distinct from research accuracy: can the junior organise a finding into an argument a senior can act on. A junior who finds the right cases but cannot write a memo that states the issue, states the answer, and shows the reasoning has not finished this phase, whatever the citation accuracy looks like.

Day 30: gate review. The senior reviews the accumulated memos against the rubric in the next section and makes an explicit call: the junior has the unaided baseline and moves to phase two, or specific gaps remain and phase one repeats with a narrower focus on those gaps. Skipping this gate is the single most common way this programme fails. Without an explicit stop, firms drift into letting a junior “get the hang of it” over an undefined period, which is exactly the informal, unmeasured model that AI broke.

Weeks 5 to 6: AI-assisted research, every claim verified by the junior. The junior now uses AI tools for research, but with a hard rule: every citation, every quoted proposition, and every statutory reference in the output gets checked against the primary source before it goes into a memo, and the junior keeps a verification log showing what was checked and against what. This is the phase where the duty to verify AI legal output stops being an abstract professional obligation and becomes a habit performed on every task. A tool that shows whether a case has been affirmed, distinguished, or overruled without the junior leaving the research screen, such as Niyam’s good-law checking, turns this verification step from a chore that gets skipped under deadline pressure into something that fits inside the normal research flow.

Weeks 7 to 8: AI-assisted drafting, senior spot-checks the verification log. Drafting assistance is added, still gated by the same verification discipline built in the previous two weeks. The senior does not re-verify everything; they spot-check a sample of the junior’s verification log against the underlying sources, which tests whether the verification habit is real or performative.

Day 60: gate review. Same structure as day 30, now assessed against the verification log rather than raw research accuracy. A junior whose spot-checked entries hold up moves to phase three. A junior whose log shows verification that was recorded but not actually performed, a case marked as checked with no evidence the primary source was opened, goes back to closer supervision, because that failure mode is more dangerous than simple inexperience: it looks compliant while carrying the exact risk the Supreme Court flagged in Pooja Ramesh Singh.

Weeks 9 to 10: independent case ownership, AI as first pass only. The junior owns research on a real matter end to end, using AI tools as a first pass to locate candidate authorities and structure a first draft, with the junior responsible for every substantive claim reaching a senior. This is closer to how the junior will actually work once the programme ends.

Weeks 11 to 12: the junior reviews another junior’s memo. Having a junior apply the rubric from the next section to a colleague’s work, under senior oversight, is one of the fastest ways to convert passive competence into active judgement. Reviewing forces the junior to articulate why a citation is weak rather than just sensing that it might be, which is a different and harder skill than producing correct research themselves.

Day 90: competence assessment and sign-off. A final review against the same baseline exercise from day one, now compared to the day one result, plus a review of the full body of memos produced across the twelve weeks. The output is a specific, written statement of what the junior can now be trusted to do unsupervised and what still needs a senior’s eye, not a generic “completed training” note.

Firms with lighter caseloads can compress or extend individual weeks, but the gate structure, an explicit checkpoint at 30 and 60 days with a real decision attached, should not be cut regardless of pace.

---

How to set a research task so the output is checkable

A badly framed research assignment produces a memo that cannot be reviewed properly no matter how good the junior is, because the reviewer has nothing to check the answer against. The fix is in how the task is set, not in how hard the junior works.

Give a specific, answerable question rather than an open topic. “Research the law on anticipatory bail” produces a survey. “Can our client, charged under Section 316 BNS for an alleged criminal breach of trust involving Rs 40 lakh, get anticipatory bail given no prior convictions and full cooperation” produces an answer that can be right or wrong, and forces the junior to identify which facts actually matter, a skill the first version never tests.

State what counts as a complete answer before the junior starts. If the task needs case law from a specific High Court because that is where the matter is pending, say so; a junior left to guess whether Bombay or Delhi High Court precedent is more relevant will guess, and the guess will not always be visible in the final memo.

Set a source floor. Require that every proposition of law in the memo trace to a specific section of a named Act or a specific paragraph of a named judgment, not a general citation to “the case” without a pinpoint reference. A memo that says “the Supreme Court has held that mere delay is not fatal” without a paragraph number cannot be checked in under twenty minutes; one that says “at paragraph 14 of [case name]” can be checked in two. A research platform that surfaces the exact paragraph a proposition comes from, rather than a headnote summary, such as Niyam, makes this floor realistic to enforce rather than a standard juniors quietly work around under deadline pressure.

Require the junior to flag genuine uncertainty rather than resolve it silently. If two High Courts have taken different views and the Supreme Court has not settled the split, the memo should say that plainly, name both lines of authority, and state which one is binding on the court where the matter is pending. A junior who quietly picks one line without disclosing the conflict is producing a memo that looks more settled than the law actually is, which is a worse outcome than an honest “this is unsettled.”

Ask for the negative case too. A memo that only cites authority favouring the client’s position, with no acknowledgment of contrary authority the other side is likely to raise, has not actually researched the question; it has advocated without testing it. This matters more once AI tools are involved, because a prompt phrased around the desired conclusion will readily return authority that supports it while missing authority that does not.

---

Reviewing a junior’s research memo

The review is where most of the actual training happens, more than any assigned reading or formal session, because it is specific to real work the junior just did and the feedback lands while the reasoning is still fresh in their mind. Use a fixed rubric so the review is consistent across juniors and across reviewers, and so the junior can see exactly what will be checked before they submit.

LEGAL RESEARCH MEMO REVIEW RUBRIC

1. ISSUE STATEMENT
   [ ] States the precise legal question, not a general topic
   [ ] Ties the question to the actual client facts
   [ ] Notes the applicable jurisdiction and forum

2. LAW STATED CORRECTLY
   [ ] Statutory citations give exact section and Act as currently in force
   [ ] BNS/BNSS/BSA used where applicable, IPC/CrPC/Evidence Act cited
       only where the pre-1 July 2024 numbering is relevant to the facts
   [ ] Case citations include party names, court, year, and a pinpoint
       paragraph reference for any quoted or relied-on proposition
   [ ] No citation appears that the reviewer cannot locate independently
       within ten minutes using the citation given

3. SOURCE VERIFICATION
   [ ] Every cited case has been checked against the primary judgment,
       not a summary, and the memo or verification log shows this
   [ ] Currency of authority checked: not overruled, not under a stay,
       not distinguished into irrelevance on materially similar facts
   [ ] Binding versus persuasive status is correctly identified for
       every authority relied on

4. ANALYSIS
   [ ] Precedent facts are mapped onto client facts, with material
       differences stated explicitly, not glossed over
   [ ] Contrary authority is acknowledged, named, and addressed
   [ ] Genuine splits in authority are disclosed, not silently resolved
   [ ] Conclusion follows from the analysis stated, not asserted ahead
       of the reasoning that supports it

5. USABILITY
   [ ] A senior could act on this memo without independently redoing
       the research
   [ ] Uncertainty is flagged where it exists, not hidden behind
       confident language
   [ ] Memo is the right length for the question: no padding to look
       thorough, no compression that drops a material point

6. IF AI-ASSISTED
   [ ] Verification log shows what was checked and against what source
   [ ] No AI-generated citation appears in the memo without an
       independent primary-source check recorded against it
   [ ] Junior can explain, without the tool open, why each key
       authority supports the conclusion

REVIEWER NOTE: mark each unchecked item with the specific fix needed,
not a general comment. "Check paragraph 14, not the headnote" is
actionable. "Needs more rigour" is not.

Deliver the review as a conversation, not a document dropped in an inbox. Walk through the two or three items that matter most rather than every checkbox, and ask the junior to explain their reasoning on the weakest point before offering the correction; a junior who can articulate why they thought a case was on point, even when wrong, is closer to independent judgement than one who simply accepts the correction and moves on.

---

The verification habit

Verification is not a step at the end of research. Treated that way, it becomes the first thing cut when a deadline tightens, which is precisely when it matters most. It has to be built as something the junior does while researching, not after.

The habit worth building is narrow: before any citation, quoted proposition, or statutory reference leaves the junior’s hands, in a memo, an email to a senior, or a draft, the junior has personally opened the primary source and confirmed the citation says what it is being cited for. Not that a summary says so. Not that the AI tool’s output looks internally consistent. The primary source, opened and read.

This matters more, not less, once AI tools are involved, for two structural reasons. First, a hallucinated citation from an AI tool is often stylistically indistinguishable from a real one; it will have plausible party names, a plausible year, and a plausible-sounding holding, which is exactly what made the fake precedents in Pooja Ramesh Singh pass through a tribunal’s reasoning undetected. Second, AI tools are consistently confident regardless of whether they are right, so a junior cannot use the tool’s own tone as a signal of reliability the way they might, cautiously, trust a hedging colleague.

Build the check for currency into the same habit as the check for existence. A real citation that has since been overruled, or distinguished into irrelevance on the specific point relied on, is arguably more dangerous than a fake one: a fake citation gets caught the moment anyone tries to pull it and fails, while a real but stale citation survives a basic existence check and only fails when someone specifically asks whether it is still good law. A dedicated walkthrough of how to run that check is in how to vet legal AI citation accuracy, and the underlying mechanics are covered in good-law checking.

Make the habit visible, not just internal. A verification log, even a simple table of citation, source checked, date checked, currency confirmed, gives a senior something to spot-check without redoing the research, and gives the junior a record showing the habit is real rather than assumed.

---

When AI-assisted work is confidently wrong

At some point in the first year, a junior will hand up work built on an AI-assisted research error stated with total confidence, a fabricated citation, a real case cited for a proposition it does not actually support, or a superseded provision presented as current law. How that moment is handled determines whether the junior treats verification as central or as theatre for the next twelve months.

Separate the error from the consequence deliberately. The immediate task is fixing what is in front of the client or the court, not assigning blame; a junior who expects a blowup for surfacing an error will start hiding uncertainty instead of flagging it, which is worse than the original mistake. Praise the disclosure explicitly if the junior caught it themselves before it went further, even though it should not have happened. Those two things are not in tension: the mistake gets corrected and taken seriously, and the honesty about it gets reinforced.

Trace the failure to the specific point in the workflow where the verification step should have caught it and did not: was the citation never checked against a primary source at all, checked against a secondary summary that repeated the error, or checked against the right document but misread. Each has a different fix. The first means the verification habit was skipped under time pressure, and the fix is process, probably a mandatory verification log for a defined period. The second means the junior does not yet reliably distinguish a primary source from a summary, a knowledge gap the training programme should revisit. The third is a reading skill, closer to the ratio-versus-obiter distinction than to a verification failure, and needs targeted practice on judgment reading rather than a stricter checklist.

Use the specific error as the next training exercise rather than a generic lecture on being careful. Have the junior write a short note explaining exactly how the error would have been caught if the verification step had been performed correctly, and file that note alongside the corrected memo. This converts one mistake into a concrete example the junior will remember far longer than an abstract warning, and gives the reviewing senior a written record of what was actually understood.

Escalate only where the pattern, not the single incident, warrants it. A first confidently wrong AI-assisted memo six weeks into a research-heavy role is a training gap to close. The same error from the same junior a third time, after the gap has been named and addressed twice already, is a different problem: the honest conversation then is whether the junior is ready for the autonomy currently given, not whether the tool needs replacing.

---

Measuring competence, not hours

“Six months in the role” and “competent to research unsupervised” are different claims, and treating the first as a proxy for the second is exactly the assumption AI broke, because time spent no longer correlates with skill built the way it did when the grunt work forced repetition. Measure the actual thing.

Track a small set of concrete indicators rather than a vague sense of how a junior is coming along. The rate at which a senior’s review catches a substantive error, wrong citation, missed contrary authority, misread ratio, per memo submitted, tracked over time, is the single most informative number: a falling error rate across real matters is direct evidence of judgement improving, not just familiarity with the tools. The proportion of a junior’s verification log entries that hold up when a senior spot-checks them is a second, since a junior can produce a plausible-looking log without the underlying checks having actually been done. Time to a checkable, complete memo on a comparable task, tracked across the twelve weeks rather than compared across different juniors, shows whether research is becoming more efficient without becoming less careful, which is the actual goal, not raw speed on its own.

Resist turning any of these into a single score a junior optimises against directly. A junior told their number is “errors per memo” will start submitting fewer, thinner memos to keep the count down, which defeats the purpose. Use the indicators as inputs to a qualitative judgement a senior makes and can defend in specific terms, not as an automated gate.

Reassess the gate decisions on a real schedule after the programme ends, at six months and at a year, rather than treating day 90 sign-off as final. Competence built under close supervision on a defined set of exercises does not automatically transfer to unsupervised work on unfamiliar matters months later, and a firm that never checks again is trusting a snapshot rather than a trend.

A capable AI research tool changes what a junior spends their time on, not what a senior needs to be able to trust about their judgement. A citator that shows a case’s subsequent history without leaving the research screen, of the kind Niyam builds, makes the verification step faster to perform. It does not make the step optional, and a training programme that treats speed as a substitute for the underlying skill will produce juniors who are fast and wrong, which is a worse combination than slow and wrong, because it is far harder for a senior to catch in time. Firms weighing whether the underlying cost of a research platform is justified by the training gains should read what legal research software actually costs an Indian firm, and firms already committed to switching tools can use the checklist for switching legal research platforms to time that change around, not against, a training cycle.

---

Frequently asked questions

How long should a firm expect before a junior can research unsupervised?

There is no fixed timeline that applies across practice areas, but a structured 90-day programme with explicit gate reviews at 30 and 60 days gives a firm a defined checkpoint rather than an open-ended sense of “getting there.” Full unsupervised trust on unfamiliar matters typically takes longer, closer to six months to a year, and should be reassessed on a schedule rather than assumed once granted.

Should juniors be allowed to use AI tools during their first weeks at a firm?

Restrict AI use to low-stakes, closely reviewed tasks in the first two to four weeks so the junior builds unaided primary-source research skills first. A junior who never develops the unaided skill cannot later catch an AI tool’s errors, because they have no independent basis to know the output is wrong.

What is the single most important thing to check in a junior’s research memo?

Whether every citation traces to a specific, pinpointed reference in a primary source that the reviewer can independently locate. A memo with no traceable citations cannot be checked properly regardless of how well written or confidently stated it is, and this single check catches most of the serious failure modes described in this piece.

How does the Pooja Ramesh Singh judgment change how firms should train juniors?

It sets a legal, not just a practical, floor for verification. The Supreme Court held that citing an AI-generated precedent without verification is misconduct on the part of an advocate. A firm’s training programme now has to demonstrate that juniors are taught, and actually practise, primary-source verification as a defined step, not as an assumed background competence. See AI-generated case law and the Supreme Court’s 2026 rule for the full holding.

Is it acceptable for a junior to use AI to draft a first version of a research memo?

Yes, once the junior has demonstrated the unaided baseline skills and has an active verification habit, because the risk at that stage is a drafting error, not an undetected factual or legal error passing through unchecked. Before that baseline is established, AI-drafted first versions risk the junior editing language around a citation they cannot independently verify.

How should a firm handle a junior who consistently produces error-free memos very quickly?

Verify that speed is not coming from skipped verification before treating it as a strength. Spot-check the verification log against primary sources on a sample of the junior’s fast, clean memos before concluding the speed reflects genuine skill rather than confident but unchecked output.

What is the difference between reviewing a memo for accuracy and reviewing it for usability?

Accuracy review checks whether every stated proposition of law is correct and every citation supports what it is cited for. Usability review checks whether a senior could act on the memo without redoing the research, which includes whether uncertainty is flagged honestly and whether the memo is structured so the conclusion and its reasoning are easy to find. A memo can pass one and fail the other.

Should junior associates be told which AI tool the firm uses and why?

Yes. A junior who understands what the tool is actually doing, retrieving and summarising source material rather than reasoning about the law independently, is better placed to know where its output needs the most scrutiny. Treating the tool as a black box teaches juniors to trust its output rather than to interrogate it.

How does a firm train for the new BNS, BNSS, and BSA numbering specifically?

Build it into the baseline skills check on day one and revisit it explicitly in memo review for the first several weeks, since juniors trained partly on older material or AI outputs trained on mixed-era text can easily conflate old and new numbering. The rule is straightforward: BNS, BNSS, and BSA replaced the IPC, CrPC, and Indian Evidence Act from 1 July 2024, published on the official gazette and summarised for the profession by PRS Legislative Research, and current filings must use the current numbering.

What should a junior do if two High Courts disagree and the Supreme Court has not resolved it?

State both lines of authority explicitly in the memo, name the deciding High Courts and the leading cases on each side, and identify which line binds the court where the matter is pending. Silently picking one position without disclosing the split is a more serious memo failure than acknowledging genuine uncertainty in the law.

How much should a review session cost a senior’s time per memo?

Aim for a focused fifteen to twenty minute conversation on the two or three most important findings rather than a line-by-line read-through of the entire rubric out loud. The rubric guides what the reviewer checks beforehand; the conversation should concentrate on what will actually change how the junior works next time.

Can a junior be trusted with client-facing research before they finish the 90-day programme?

Only with a senior reviewing every substantive claim before it reaches the client, regardless of how the junior is performing internally. Client-facing exposure and internal training gates are separate decisions, and a junior can be progressing well through the programme while still needing that external safety net for anything a client or opposing counsel will see directly.

What if the firm is too small to run a formal 90-day programme?

Compress the phases rather than skip the gates. Even a solo practitioner training one junior can run a shortened version, a week of unaided research, a week of unaided memo writing, a review checkpoint, then AI-assisted work with verification logging, as long as the explicit checkpoint and the rubric-based review survive the compression.

How does a firm know if its verification habit is real rather than performed for review?

Spot-check, not full re-verification. A senior periodically pulls the primary source themselves for a small random sample of entries in a junior’s verification log and confirms the check was actually done as described. A log with no discrepancies over several spot-checks, across several weeks, is stronger evidence than a junior’s own assurance that they checked everything.

Does training change for a junior moving from litigation to a corporate or transactional practice?

The verification discipline stays the same, but the applicable law and the shape of the research task differ, and a junior trained entirely on litigation research will need a deliberate onboarding period on transactional sources and diligence workflows before being treated as competent there. See legal research for litigators versus corporate lawyers for how the underlying research skills diverge by practice area.

Should in-house legal teams follow the same training approach as law firms?

The core discipline transfers directly: unaided baseline first, checkable task framing, rubric-based memo review, and a visible verification habit. In-house teams often have a narrower subject-matter scope and fewer juniors, which usually means the 90-day programme can be tailored to the specific regulatory or contractual domain the team works in rather than run as a general litigation curriculum. See legal research for in-house counsel in India for how in-house research needs differ from firm practice.

How do you stop a junior from just re-prompting an AI tool until it produces a citation that checks out?

Require the verification log to record the first citation the tool produced and whether it was accepted or rejected on primary-source review, not just the final citation used. A junior who re-prompts repeatedly until something passes is still performing the verification step correctly on the citation that ends up in the memo; the log makes that process visible to the reviewer rather than hidden inside a chat history nobody looks at again.