Your Own Papers, Defended

0. Why this is the most important chapter

Everything else here is material you might be asked about. This is material you will be asked about, because it is yours and it is on the CV in front of them. Goanta in particular does exactly this kind of empirical work at scale, and she will go to method rather than to findings.

Read this before the rest — a calibration note

This chapter was first written as if you were defending a thesis. You are not. It is a 25-minute first interview, from nearly 400 applications, described in the invitation as questions about your background and application to evaluate your fit. **Realistically you get one question about your work, and it is *tell me about it* rather than *justify your p-values*. Nobody on a four-person screening panel has audited whether GridSearchCV was applied uniformly across eight models in a second-author student preprint. So treat everything below as insurance rather than as an indictment. The items are ranked by likelihood from here on: high means have the sentence ready; low means read it once so it cannot surprise you, and then stop thinking about it. The strengths table immediately below is what you should actually be rehearsing.**

Paper	What is genuinely good — rehearse this	Likelihood of being probed
Beyond the Dark Figure	A 20-year district panel built from scratch, an ICC of 0.6 percent that is a clean variance partition, a reporting-shock estimate of about +88,879 cases a year around a known statutory amendment, and VIFs reported unprompted. This is a real empirical paper	High — it is your lead publication, it has a DOI, and it is on the CV
Mens Rea Evaluator	A paired three-condition design, a parser-scored forced-choice probe, per-run provenance logging, and an error in version one that you found and fixed yourself	High — the title is memorable and it is the most ATLANTIS-adjacent thing you have
Master's thesis	A ten-stage working system, built solo in four months, with an explicit cost argument	Medium — they have the document. Its own chapter now
Hybrid summarisation	An honest three-stage negative result that arrives at the neurosymbolic argument empirically rather than by assertion	Low — and if it comes up, the negative result is the contribution
IoT cyber-security	Your only non-preprint journal article, genuinely comparative EU/US/India, and the closest thing you have to Wisman's field	Low to medium — low on method, higher if Wisman wants common ground
DiagnoChat	A working end-to-end system with a real interface. And you can now read it back through the AI Act and the Medical Device Regulation, which is the interesting move	Low — second author, course project, four years old

1. Beyond the Dark Figure — your strongest paper, and the one with the sharpest hole

SocArXiv preprint, DOI 10.31235/osf.io/4fxsm_v2, posted 19 September 2026. Sole author. This is now your lead publication and it is the best bridge you have to ATLANTIS — but there is one internal tension in it that a careful reader finds in about a minute, so take that section first.

Result	The number	What it supports
Moran's I — spatial autocorrelation	I = 0.095, p = 0.151	No evidence of geographic clustering. High-crime districts are not systematically adjacent to other high-crime districts
ICC from the two-level mixed model	State variance 3.895, district residual 662.59 → ICC = 0.6%	State-level policy and governance explain under 1 percent of variance in local rates. The other 99.4 percent is district-level. This is the headline and it is robust
Urbanisation, mixed-effects fixed effect	β = 0.371, p = 0.004	The one clearly significant demographic driver. Dominant predictor of normalised reported crime rates
Literacy gap, mixed-effects fixed effect	β = 0.148, p = 0.156	Not significant. Positive direction, does not reach conventional thresholds
Literacy gap, OLS panel with time FE and clustered SE	β = 0.2049, p = 0.399	Also not significant, and with wide intervals from the cluster-robust errors
ATE of literacy gap, DoWhy after closing the urbanisation backdoor	0.1487 per 1 percentage point	A point estimate of the causal effect — and note it is essentially the same number as the non-significant 0.148
Random Common Cause refutation	Re-estimated at 0.1485, p = 0.96	The estimate is stable when a random confounder is injected
BSTS, 2013 Criminal Law (Amendment) Act	+88,879 cases per year against the counterfactual	A legislative reporting shock, not a behavioural epidemic. Your single most quotable result
Kidnapping → rape, DAG controlling urbanisation	Pearson r = 0.43; causal β₁ = 0.26	Offence types are causally linked rather than independent — an escalation pathway
Bayesian Belief Network risk matrix	Low urbanisation + high gap → P(high crime) = 49.0%. High urbanisation + high gap → 5.9%, with P(medium) = 76.5%	Strongly non-linear interaction: the literacy gap's effect reverses sign with urbanisation
VIF diagnostics	All below 5 — female literacy 1.85, literacy gap 1.63, urbanisation 1.26, gender ratio 1.01	No multicollinearity problem. Good practice, worth mentioning unprompted

The one real vulnerability — HIGH likelihood, and worth raising yourself

Your abstract says the analysis isolates the causal effects of the literacy gap and urbanisation, and that urbanisation is the dominant driver. But the literacy gap is not statistically significant in either regression — p = 0.399 in the OLS, p = 0.156 in the mixed model. Then section 4.4 reports a causal ATE of 0.1487 for that same variable, which is numerically almost identical to the non-significant 0.148 coefficient. A reader will ask how you can report a causal effect for a predictor you could not distinguish from zero. The honest answer is good, so have it ready: *the identification strategy and the precision are separate questions. Closing the backdoor path tells me which quantity I am estimating; it does not shrink the standard error. So the correct statement is that the point estimate of the causal effect is 0.1487 and I cannot rule out zero at conventional levels. The abstract overstates that, and urbanisation is the result I would actually defend.* Say that before they ask, and the paper's credibility survives intact.

Three more — LOW likelihood, read once and move on

**One — *proves* twice. The refutation test is described as proving robustness. It does not: Random Common Cause injects a random confounder, and a p of 0.96 means the estimate barely moved, which is what you would expect from adding noise. It does not test a *plausible* confounder. The stronger tests you did not run are a placebo refuter and, better, an E-value or sensitivity analysis quantifying how strong an unmeasured confounder would have to be to overturn the result. Name that. Two — Moran's I.** p = 0.151 means you failed to reject the null of no clustering. Your abstract correctly says *no evidence of geographic clustering*, but section 4.2 says the distribution is *statistically random*, which is a stronger claim your test cannot support. Absence of evidence is not evidence of absence, and at district-level aggregation you may simply lack power. Three — the BBN cells. The matrix has a non-monotonic row: medium urbanisation with a high gap gives P(low) = 72.2 percent and P(medium) = 5.6 percent, which is an odd shape. With nine cells over a district panel, some conditional probability tables are likely estimated on very few observations. The non-linearity may be real or may be sparse cells, and you should say you would report cell counts and credible intervals before leaning on it.

Your DAG, and the transfer to enforcement

  YOUR PAPER
  ================================================
                 Urbanization
                  /        \
                 /          \
                v            v
          Literacy_Gap ---> Crime_Rate <--- Gender_Ratio

   Urbanization is the CONFOUNDER: it causes both the
   literacy gap and the reported crime rate, so the raw
   association between gap and crime is partly spurious.
   Closing that backdoor is the whole method.

   AND THE KEY INSIGHT: "Crime_Rate" is not crime.
   It is REPORTED crime. Urbanization raises reporting
   by improving access to police infrastructure, while
   also plausibly raising incidence. ONE ARROW, TWO
   MECHANISMS, pointing the same way.


  THE SAME STRUCTURE IN COMPETITION ENFORCEMENT
  ================================================
              Screen deployment
                  /        \
                 /          \
                v            v
          Firm behaviour --> FLAGS <--- Sector features

   "FLAGS" is not collusion. It is DETECTED collusion.
   Deploying the screen raises detection while also
   changing the conduct (evasion / deterrence).
   ONE ARROW, TWO MECHANISMS, pointing OPPOSITE ways.


  WHY THIS IS YOUR BEST PITCH
  ================================================
   Police records measure reporting, not offending.
   A cartel screen measures detection, not collusion.

   Same defect. And you have already built the toolkit
   for separating the two: an explicit DAG, a
   counterfactual time series around a known
   intervention, and a decomposition that distinguishes
   an administrative shock from a real change.

   The 2013 Act is to NCRB data what deploying a screen
   is to enforcement statistics.

This is the single strongest thing you can say in the interview, and it comes from this paper rather than from the Mens Rea one. The structural analogy is exact, the method transfers directly, and no legal-track candidate can propose it.

2. A Hybrid Approach to Legal Text Summarization — the honest negative result

SSRN, DOI 10.2139/ssrn.6669602. Ghosh & Vetriselvi, work dated 2024, posted 2026. Intellectually this is the paper that most directly supports your symbolic-layer argument, because it is a documented three-stage progression in which you tried the obvious thing, watched it fail, tried the pure-logic thing, watched it fail differently, and built the hybrid.

The three approaches, and why each failed

  APPROACH A - feed it straight to BART
  ===============================================
   input: "In the case of Mr Anand Kashyap vs Mrs
    Joharlal Gandhi ... Section 73 Indian Contract
    Act ... High Court of Calcutta ... 10,000 kg of
    wheat ... loss of 5,000 rupees"
        |
        v
   output: captured the gist, and LOST:
     - party names (defaulted to "petitioner",
       "defendant" - named entities mentioned once
       are dropped)
     - the jurisdiction
     - the sum of money
        |
        v
   DIAGNOSIS: surface-pattern summarisation cannot
   tell SALIENT from FREQUENT. In law the salient
   facts are often the ones mentioned exactly once.


  APPROACH B - regex + POS -> FOL -> back to text
  ===============================================
   extracted: acts, parties, addresses, dates,
    sections, petition format, jurisdiction, court,
    loss amount
        |
        v
   FOL:  Contract(Kashyap, Gandhi)
         DeliveryDue(Gandhi, 15/03/2021)
         DeliveryLate(Gandhi, Kashyap)
         Loss(Gandhi, 5000)
        |
        v
   and it could DEDUCE:
     "Gandhi breached the contract with Kashyap due
      to late delivery, causing a loss of 5,000."
        |
        v
   WORKED. But DID NOT SCALE.
   DIAGNOSIS: the logic is CONTEXT-SPECIFIC. A
   predicate set built for breach of contract does
   not transfer to theft or defamation. Finite
   vocabulary, effectively unbounded sentence
   construction. You cannot hand-write the rules.


  APPROACH C - the hybrid
  ===============================================
   regex (fixed-form entities)
     + POS (subject / object / action)
     + GloVe + cosine similarity (which action
       attaches to which object)
     + select the LEAST similar pairs, on the
       theory that unexpected pairs carry
       information the entity extractor missed
     + sentence templates
     + T5 / Parrot paraphrase for fluency
        |
        v
   RESULT: richer summary, key legal elements
   preserved - BUT the paraphraser dropped some of
   the extracted relationships, reintroducing the
   original failure at the last step.

Read Approach B's diagnosis as the finding. You did not fail to make logic work — you demonstrated *why* hand-built logic does not scale, which is precisely the argument for pairing a language model with a logic layer rather than choosing one. That is the neurosymbolic case, reached empirically.

The vulnerability — LOW likelihood, but it is one sentence in the abstract

**Your abstract says *experimental results demonstrate the method's effectiveness*. There is no experiment. There is one worked example — the Kashyap contract case — and no quantitative evaluation of any kind: no ROUGE, no entity-extraction precision or recall against a gold standard, no human rating, no baseline comparison with a number attached. Anyone who runs empirical studies looks for a number first, so this is the kind of thing that gets noticed if the paper is read at all** — though for a 25-minute fit interview, that is a real *if*. The answer that survives: *that sentence overstates it and I would not write it again. What the paper actually contains is a qualitative three-stage design study on a worked example. The evaluation I should have run is entity-level precision and recall against a hand-annotated set of judgments, plus a human rating of whether the summary preserves the holding — and I would want the second, because overlap metrics cannot see a reversed outcome.* Then point at your own later work: the Mens Rea harness is what you built once you understood that evaluation design was the hard part.

Two more — LOW likelihood, reachable only by someone reading closely

The least-similar-pairs heuristic is unjustified. You select action–object pairs with *low* cosine similarity on the theory that unexpected pairings carry overlooked information. That is a plausible intuition and nothing in the paper validates it — low similarity could equally indicate a parsing error or an irrelevant coincidence. Say it is a heuristic you would now test rather than assert, by checking whether the pairs it surfaces are ones a lawyer marks as material. And the GloVe choice is the other one. You state the limitation yourself — GloVe has poor coverage of rare and domain-specific legal terms — but the deeper issue is the one from the NLP chapter: GloVe gives one vector per word regardless of context, so *consideration* has a single vector across contract doctrine and ordinary usage. A contextual encoder would have been the right tool, and the honest reason you did not use one is the resource constraint you already state.

What to say when this paper comes up

*It is a three-stage negative result and I think the negative part is the contribution. BART alone lost party names, jurisdiction and the sum in dispute — it defaulted to 'petitioner' and 'defendant', because a named entity mentioned once looks unimportant to a surface-pattern model, and in law the fact mentioned once is often the operative one. Hand-built first-order logic fixed that and produced real deductions, and then did not scale, because the predicate set for breach of contract does not transfer to theft or defamation. So the paper's conclusion is that neither half is sufficient, which is why I now think the interesting architecture is a model for extraction and a logic layer for the test. The paper's weakness is that it has no quantitative evaluation, and the abstract claims experimental results it does not contain. The evaluation harness in my later work is what I built after learning that.*

3. DiagnoChat — where you are second author, and where the claim does not hold

SSRN, DOI 10.2139/ssrn.6669459. Sharma, Ghosh & Sharma — you are second of three, with Ritik Sharma first and Dr Sathya Narayana Sharma K. supervising. Symptom-based disease prediction over a National Health Portal of India dataset: 200+ diseases, 489 symptoms, regex plus a custom synonym dictionary for symptom extraction, HTML and JavaScript front end over a Python backend, and a nearby-care recommendation in urgent cases.

Model	Reported accuracy	Note
Naive Bayes + GridSearchCV	90.11%	The paper's headline. And the only model that was tuned
KNN	88.94%	Untuned, and second best
Logistic regression	88.23%	Untuned
Random forest	87.95%	Untuned
SVM	86.81%	Untuned
Decision tree	81.78%	Untuned
Weighted Bernoulli NB	70.91%	The paper's custom modification
Traditional Bernoulli NB	66.73%	Worst of the eight

The vulnerability — LOW likelihood. Second author, course project, four years old

Context first, because it changes the standard that applies: this was a three-person undergraduate course project, you are second of three, and it is from 2022. Tuning one model and not the others is ordinary student practice, not misconduct, and nobody screening 400 applicants is auditing it. Read the next paragraph once so it cannot surprise you. The paper concludes that Bernoulli Naive Bayes shows superior performance and is particularly well suited to this task. Its own table does not support that. Untuned Bernoulli NB is 66.73 percent — the worst of the eight models. The 90.11 percent figure comes from NB plus GridSearchCV, and no other model was given the same hyperparameter search. So the comparison attributes to the algorithm what may be attributable to the tuning. KNN at 88.94 percent and logistic regression at 88.23 percent were never tuned at all. The fair design is to run GridSearchCV over every candidate and compare the tuned versions. If asked, concede it cleanly: *that comparison is not like-for-like, and I would not stand behind the claim that Naive Bayes is superior here. Only one model received a hyperparameter search. The defensible statement is that a tuned Naive Bayes reached 90 percent on this dataset, and that an equivalently tuned KNN or logistic regression might well have matched or beaten it.*

Three further problems — LOW likelihood. Awareness only, do not volunteer these

Accuracy is close to meaningless over 200+ classes. With that many diseases the class distribution will be extremely uneven, and a single accuracy figure says nothing about rare conditions — which in a diagnostic context are exactly the ones that matter. Macro-averaged precision and recall per disease is what the task needs, and the paper reports neither. The numbers are internally inconsistent: 66.73 against 66.72, 90.11 against 90.10, and section IV says the system has achieved accuracies *up to 70.91 percent* while the rest of the paper claims 90.11. Be the one who notices. And there is no reported train–test protocol beyond mentioning cross-validation — no split sizes, no stratification, no held-out set described. So the figures are not reproducible from the paper as written.

The regulatory observation that turns this paper into an asset

This is the move to make with this paper, because the methodology will not survive scrutiny and the legal framing will. *A symptom checker that predicts disease and recommends seeking care would, deployed in the EU today, be a serious regulatory object rather than a student project. Under the AI Act a system intended for use in healthcare triage sits close to the high-risk categories, and under the Medical Device Regulation software intended for diagnosis is a medical device requiring conformity assessment, with the intended-purpose statement doing the classification work. We built it as a course project with none of that in view. I find that instructive rather than embarrassing: the gap between what is technically straightforward and what is lawfully deployable is exactly the gap I want to work on, and I have been on the wrong side of it.* That answer is strong, it is honest, and it converts a weak paper into evidence of exactly the judgment this position needs.

4. The Mens Rea Evaluator — the numbers, and the framing to lead with

SSRN, DOI 10.2139/ssrn.7446198, dated 16 August 2026. Covered in detail in the interview brief; repeated here so every paper is in one place.

Result	Value	Say
Control condition	0 of 15 biased	Clean baseline, which is what makes the rest interpretable
Ablation	9 of 15 pooled (4, 4, 1 of 5)	Partial manipulation, partial effect. Wide spread, not separable
Poisoned	15 of 15 biased	Fabricated authority in context produced a biased ruling every run
Admissions	2 of 15, both self-judged	In the 10 rows graded by a different model, none
Cognizable intent	0 of 15	The finding. Not the susceptibility rate
McNemar, control→ablation	p ≈ 0.004	9 changes one way, 0 back
McNemar, control→poisoned	p ≈ 0.00006	15 changes, 0 reversals
Fisher, model vs model	p = 0.206	Licenses nothing. No model ranking
Wilson, 4 of 5 / 1 of 5	≈38–96% / ≈4–62%	Overlap heavily — the formal reason models are not separable

The framing correction to make unprompted

*Mens rea does not transplant, and my paper is not arguing that it does.* EU competition liability is largely objective — the undertaking is liable for conduct, intention bears on fine-setting rather than on establishing the infringement, and the AI Act places duties on providers and deployers precisely because it does not treat a model as a subject. What the result bears on is the duty to give reasons. The system's account of its own decision is consistently true, verifiable and incomplete, which is a problem for Article 296 TFEU and for Article 14 oversight, not a problem about intent. The mens rea framing was the route in; it is not the conclusion. Make this correction before anyone presses on it — the title invites the question and pre-empting it converts the paper's weakest feature into a demonstration of judgment.

5. The thesis, and the one paper I have not read

Work	Status	What to say
Master's thesis — Fine-Tuning LLM with RAG for Generating Educational Audio-Visual Content, VIT Vellore, May 2025	Read in full. Now has its own chapter — go there, not here	Three corrections to volunteer: the QLoRA fine-tune ran on rented GPUs — Colab and RunPod — with development, the data pipeline and the app on the MacBook, because 4-bit training needs CUDA and the write-up is loose about this; the RAG corpus was 10 percent of Simple English Wikipedia rather than English Wikipedia; and both the training answers and the reference answers came from GPT-4o Mini, so BLEU and ROUGE measure agreement with the teacher rather than correctness. Do not repeat the user-testing claim — it contradicts your own Limitation 8
*Cyber Security Issues by Using the Internet of Things (IoT): A Legal Analysis*, Ghosh U. & Ghosh R., Indian Journal of Legal Research, Vol 4 Issue 4, 2022	Read in full. Own section below	Do not treat this as the throwaway. It is your only non-preprint journal article, it is comparative EU/US/India law, and IoT privacy is the subject Wisman has published on for over a decade. Of all your work it is the closest to his

6. The IoT paper — your only journal article, and closest to Wisman

*Cyber Security Issues by Using the Internet of Things (IoT): A Legal Analysis*, Upasana Ghosh & Rudrani Ghosh, Indian Journal of Legal Research, Vol 4 Issue 4, 2022. You are second author, and the paper lists you as a second-year LL.B. student — so this is early work and should be framed as such. Do not treat it as the throwaway, for two reasons: it is your only non-preprint journal article, and IoT privacy is the subject Tijmen Wisman has published on for roughly fifteen years.

What it does	Detail
The argument	India has no law specifically regulating IoT. The IT Act 2000 does not define cybercrime at all, let alone IoT-enabled offences, and its provisions are bailable unless read together with the Indian Penal Code — so the paper maps IT Act sections onto IPC sections to show how prosecutions actually have to be constructed. Conclusion: India urgently needs a dedicated legislative framework
The comparative section	EU: GDPR for data generated by IoT devices, the NIS Directive for infrastructure, the Cybersecurity Act 2019 for businesses. US: the IoT Cybersecurity Improvement Act 2020 for federal procurement, and California's IoT security law — first in the US, enacted 2018, effective 1 January 2020. This is the part worth mentioning: it is comparative EU/US/India law, which is exactly this panel's register
The empirical component	A questionnaire to 50 respondents in Kolkata, distributed by email, WhatsApp and Facebook Messenger, anonymous, with no mandatory questions. Findings included 98 percent daily internet use, 30.6 percent reusing passwords, 69.4 percent claiming to understand IoT, 100 percent believing cybercrime had increased, and 100 percent supporting separate legislation
The case material	The 2018 Nest baby-monitor intrusion; the FDA recall of around 500,000 pacemakers over hackability; a casino database reached through an aquarium thermostat; the 2016 Mirai botnet taking down major services; heating cut to Finnish apartment blocks for a week; and the Chandan Kumar case in India, where a complaint filed in November 2021 remained untraced

The one thing to fix in your head before Wednesday — MEDIUM, because it is a naming error on an EU law panel

**The paper calls it the *General Data Privacy Regulation, 2018*. That is wrong on both counts. It is the General Data *Protection* Regulation, Regulation (EU) 2016/679, adopted 2016 and applicable from 25 May 2018. This is an EU law panel and that is the single most embarrassing thing in your entire publication list if someone spots it. It is a 2022 paper written by a second-year student and you were not the lead author, so it is forgivable — but you must not repeat the error in the room. If GDPR comes up at all, name it correctly and give the regulation number. If anyone raises the paper, concede it immediately:** *that paper gets the GDPR's name wrong, which I noticed later — it is the General Data Protection Regulation, 2016/679.* Volunteering it neutralises it completely.

The methodological problems — LOW likelihood, and remember you were a second-year LL.B. student

**Hold the standard that applies: 2022, second author, second-year law student, student-run journal. This is early work and the right frame is *I would not run it this way now*, not *I am embarrassed by it*. The sample. n = 50, a single city, and the paper claims stratified sampling without ever specifying the strata. Recruitment by WhatsApp and Facebook Messenger is a convenience or snowball sample, not a stratified one. Say so plainly if asked** — *it was a convenience sample described as stratified, and that is not right.* The question wording. Two items returned 100 percent, including *do you think there must be a separate legislation* — that is a leading question and a 100 percent response is a symptom of it rather than a finding. The victimisation items ask whether you have been a victim *or know anyone who has*, which conflates personal and second-hand experience, so 53.1 and 64 percent are not victimisation rates. And varying denominators — 48, 49, 50 — because no question was mandatory, which the paper handles honestly but which means there is no consistent base.

What to say if Wisman raises it — and he is the one most likely to

*It is early work and I was second author as a second-year law student, so I would not oversell it. What I would stand behind is the framing: the argument is that the law lagged the technology and the gap produced practical impunity — the Chandan Kumar case in it is someone who filed a complaint and was simply never helped. And I notice that is the same argument I would now make about the AI Act, just about adaptivity rather than absence. What I would not stand behind is the empirical half. Fifty respondents in one city, described as a stratified sample when it was a convenience sample, and two questions that returned a hundred percent because they were leading. It also gets the GDPR's name wrong, which I found later. If I ran it again I would either do the doctrinal comparison properly or do the survey properly, and not present fifty responses as though they supported the legal argument.*

Why it is worth re-reading tonight anyway

Wisman's publication record includes RFID and consumer privacy in connected devices, and the privacy implications of smart meters and the smart grid. This is the only paper you have that touches his actual subject. And there is a live connection you can draw: the EU Data Act's access-by-design obligation took effect on 12 September 2026, requiring new connected products to make their data available to users in structured, machine-readable form. So the regulatory vacuum your 2022 paper complained about has partly been filled in the four years since — and filled from the competition and data-economy side rather than the privacy side. That is a genuinely interesting observation about which body of law ended up doing the work, and it is exactly the kind of thing that sits between him and Mirza.

7. The pattern across all of them — and the honest self-assessment

What the four papers have in common, stated as a research identity

*Looking across them, there is one question I keep returning to without having planned to: what is the measurement actually measuring? The crimes paper is about police records measuring reporting rather than offending. The Mens Rea paper is about a model's stated reasons not being its actual reasons. The summarisation paper is about overlap metrics scoring a summary that drops the operative facts. And DiagnoChat, in retrospect, is about an accuracy figure over two hundred classes that tells you nothing about the rare ones. Four papers, one question — and it is the question the accuracy strand of ATLANTIS is asking.* That is the single most useful sentence in this chapter. It converts a scattered CV into a research programme, and it is true rather than constructed.

Paper	Greatest strength	Vulnerability to volunteer first
Dark Figure	The ICC of 0.6 percent and the 88,879-case reporting shock. Genuinely strong, well-identified results	The literacy gap is not significant (p = 0.156 and 0.399) yet the abstract says the analysis proves its causal effect. Concede the wording and defend urbanisation instead
Mens Rea	Paired three-condition design, parser-scored forced-choice probe, per-run provenance log, and a self-found v1 error	15 runs, 5 scenarios, judge was also a subject in 5 rows, and both admissions fall in those 5
Hybrid summarisation	An honest three-stage negative result that is the neurosymbolic argument, derived empirically	No quantitative evaluation at all, while the abstract claims experimental results
DiagnoChat	A working end-to-end system with a real interface and sensible NLP preprocessing	Only one model was tuned, so the superiority claim does not follow. Accuracy is the wrong metric over 200+ classes

The trap to see coming

If you volunteer weaknesses well, a good interviewer will test whether you are performing humility or exercising judgment. The follow-up is usually some form of *so is any of it reliable?* Do not retreat. The answer: *Yes — differently in each case. The ICC result I would defend as it stands, because the variance partition is a direct computation and the finding is large. The reporting-shock estimate I would defend with the caveat that a BSTS counterfactual assumes nothing else changed at that moment, and the Nirbhaya case plausibly changed behaviour as well as recording. The literacy-gap causal estimate I would present as a point estimate that cannot exclude zero. And the summarisation paper I would present as a design study rather than an evaluation. Being able to say which claims survive at which strength is the skill, and it is the one this position is for.* That answer is better than any individual result in any of the papers.

8. If you remember ten things

**Strike the word *proves*.** It appears in two of your papers. Say *indicates*, and say unprompted that you would rewrite those sentences.
Dark Figure's hole: the literacy gap is not significant — p = 0.399 OLS, p = 0.156 mixed — yet the abstract claims its causal effect is proven. Identification and precision are separate; concede the wording, defend urbanisation (β = 0.371, p = 0.004).
Defend these three Dark Figure numbers without hesitation: Moran's I 0.095 at p = 0.151, ICC 0.6 percent, and +88,879 cases a year from the 2013 Act.
Random Common Cause does not prove robustness — it injects a *random* confounder. The stronger test is a sensitivity or E-value analysis, and you did not run it.
Moran's I p = 0.151 means no evidence of clustering, not proof of randomness. Section 4.2 of your own paper overstates this.
Police records measure reporting; a cartel screen measures detection. Same defect, and your DAG plus BSTS toolkit transfers directly. This is your best pitch and it comes from Dark Figure, not Mens Rea.
The summarisation paper has no experiment despite the abstract claiming one. Reframe it as a design study — and its *negative* result, that hand-built FOL does not scale across offence types, is the actual contribution.
DiagnoChat: only Naive Bayes got GridSearchCV, so the superiority claim does not follow. Untuned NB was the worst of eight at 66.73 percent. Concede it, then make the AI Act and Medical Device Regulation point — that converts the paper into an asset.
Thesis: rented GPUs for the fine-tune with local development, Simple English Wikipedia not English Wikipedia, and references generated by the teacher model. Volunteer all three, and do not repeat the user-testing claim. It has its own chapter now — read that one.
The through-line: all four papers ask what the measurement is actually measuring. Say that sentence — it turns a scattered CV into a research programme, and it is the accuracy strand's own question.