REVISION 3 — Your Papers

0. Start here — the one sentence that ties them together

The through-line, and it is the most valuable thing in this chapter

Your CV looks scattered: crime statistics, legal summarisation, a symptom checker, an LLM evaluation harness, an education system. It is not scattered, and here is why: *Looking across them, there is one question I keep returning to without having planned it: what is the measurement actually measuring?* *The crime 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. DiagnoChat, in retrospect, is about an accuracy figure over two hundred classes that says nothing about the rare ones. And the thesis is about an evaluation that measured imitation of a teacher model rather than correctness.* *Five pieces of work, one question — and it is the question the accuracy strand of ATLANTIS is asking.* Say that and a scattered CV becomes a research programme. It is also true, which is why it works.

The same defect, five times

  WHAT YOU WANTED        WHAT YOU ACTUALLY HAD
  TO MEASURE             IN THE DATA
  ==================================================
  crime                  recorded offences
                         (reporting, not offending)
  ------------------------------------------------
  a model's reasoning    the model's stated reasons
                         (fluent, true, incomplete)
  ------------------------------------------------
  summary quality        word overlap with a reference
                         (scores a wrong summary well)
  ------------------------------------------------
  diagnostic skill       accuracy over 200+ classes
                         (says nothing about rare ones)
  ------------------------------------------------
  grade-appropriateness  agreement with GPT-4o Mini
                         (imitation, not correctness)
  ==================================================

  AND THE ONE THEY ARE HIRING FOR:

  collusion              cartel DECISIONS
                         (enforcement, not conduct)

   ^ Your five, and their one, are the same problem.
     That is the whole pitch.

Every item in the left column is the thing you cared about. Every item in the right column is the thing that was available. The gap between the two columns is called measurement validity, and the bottom row is the problem this project exists to solve.

1. Beyond the Dark Figure — your lead publication

What it is	SocArXiv preprint, DOI 10.31235/osf.io/4fxsm_v2, posted 19 September 2026. Sole author. A twenty-year district-level panel study of recorded crime in India, built from National Crime Records Bureau data with interpolated census demographics
What it set out to do	Ask whether recorded crime is driven by demographic factors — urbanisation, a male-female literacy gap, gender ratio — and whether the pattern is geographic
What worked	Three findings, and all three are defensible. No spatial clustering. State-level factors explain almost nothing. And a legislative change produced a large, measurable jump in recording
What did not	The literacy gap, which was the variable the paper was most interested in, is not statistically significant — and the abstract claims more than the statistics support

Result	The number	What it actually supports
Moran's I (spatial clustering)	I = 0.095, p = 0.151	No evidence of geographic clustering. High-crime districts are not systematically next to other high-crime districts
ICC from the two-level model	state variance 3.895, district residual 662.59 → ICC = 0.6%	State policy and governance explain under 1 percent of variation in local rates. The other 99.4 percent is local. This is the headline and it is robust — a variance partition is a direct computation, not an inference
Urbanisation	β = 0.371, p = 0.004	The one clearly significant demographic driver. The result you would actually defend
Literacy gap (mixed model)	β = 0.148, p = 0.156	Not significant
Literacy gap (OLS, clustered errors)	β = 0.2049, p = 0.399	Also not significant
Causal effect of literacy gap	ATE = 0.1487	A point estimate after closing the urbanisation backdoor path — and note it is essentially the same number as the non-significant coefficient
Robustness check	re-estimated at 0.1485, p = 0.96	The estimate barely moved when a random confounder was injected
BSTS, 2013 Criminal Law (Amendment) Act	+88,879 cases per year against the counterfactual	A reporting shock, not a behavioural epidemic. Your single most quotable result
Kidnapping → rape	r = 0.43; causal β = 0.26	Offence types are linked rather than independent — an escalation pathway
Risk matrix	low urbanisation + high gap → P(high crime) = 49.0%; high urbanisation + high gap → 5.9%	A strongly non-linear interaction — the literacy gap's effect reverses with urbanisation
VIF diagnostics	all below 5	No multicollinearity problem. Worth mentioning unprompted; it shows you ran diagnostics

The one real vulnerability — volunteer it, do not wait

**Your abstract says the analysis isolates the causal effects of the literacy gap and urbanisation, and uses the word *proving*. But the literacy gap is not significant in either regression — p = 0.399 and p = 0.156. Then a causal effect of 0.1487 is reported for that same variable. A careful reader asks how you can report a causal effect for a predictor you could not distinguish from zero. The answer is good, so have it ready: *Identification and 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 at p = 0.004 is the result I would actually defend.* Say that before they ask and the paper's credibility survives intact.**

And the three limitations the paper already states, which are good

Point at your own limitations section — it is a strength. The paper flags that the data tracks First Information Reports only, so the measure is reporting rather than incidence; that the dark figure persists; that district aggregates mean individual inference would be an ecological fallacy; and that interpolating census figures annually assumes constant sub-national growth. Two design choices also show care and are worth naming: annual population interpolation rather than a static denominator, and dropping districts with all-zero counts, because an all-zero row signals a reporting failure rather than an absence of crime. And the paper carries an explicit AI-usage statement — a tool used for pipeline structure and readability, with you taking sole responsibility. On a panel that studies AI governance, a researcher who already discloses their own tool use is making a small but real credibility point.

2. The Mens Rea Evaluator — the most ATLANTIS-relevant thing you have

What it is	SSRN, DOI 10.2139/ssrn.7446198, dated 16 August 2026. Written during an Apart Research hackathon sprint — which is your legitimate answer on sample size
What it set out to do	Test whether a language model, given fabricated legal authority in its context, would produce a biased legal ruling — and whether it would acknowledge having done so
The design, and this is the strong part	A paired three-condition experiment. A clean control; an ablation with partial manipulation; a poisoned condition with fabricated authority planted in the retrieved context. Plus a parser-scored forced-choice probe and a per-run provenance log
What worked	The design worked and the result is clean. A clean baseline makes everything else interpretable, and the effect was total in the poisoned condition
What did not	15 runs over 5 scenarios. And in 5 of those rows the model judging the output was also a subject — both of the two admissions fall in exactly those rows

Result	Value	What to say
Control	0 of 15 biased	A clean baseline — which is what makes the rest mean anything
Ablation	9 of 15 (4, 4, 1 of 5)	Partial manipulation, partial effect. Wide spread across scenarios, so not separable
Poisoned	15 of 15 biased	Fabricated authority in context produced a biased ruling on every single run
Admissions	2 of 15, both self-judged	In the 10 rows graded by a different model, none. Volunteer this
Cognizable intent	0 of 15	This is 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 against model	p = 0.206	Licenses nothing. No model ranking — say so
Wilson intervals	≈38–96% and ≈4–62%	They overlap heavily — the formal reason the models are not separable

The framing correction to make unprompted — this is essential

*Mens rea does not transplant to AI systems, and my paper is not arguing that it does.* *EU competition liability is largely objective — the undertaking is liable for conduct, and 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 actually bears on is the duty to give reasons.* *The system's account of its own decision was consistently true, verifiable and incomplete — which is a problem for the obligation to state reasons and for meaningful human 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 judgement.

3. Hybrid Legal Text Summarization — the honest negative result

What it is	SSRN, DOI 10.2139/ssrn.6669602. Ghosh and Vetriselvi. A three-stage design study on summarising Indian judgments
What it set out to do	Produce summaries of judgments that keep the legally operative facts
What worked	The diagnosis. You tried the obvious thing, watched it fail, tried the pure-logic thing, watched it fail differently, and built a hybrid. The negative result is the contribution and it is a real one
What did not	There is no quantitative evaluation of any kind, and the abstract claims experimental results

Three approaches, and why each failed

  A -- FEED IT STRAIGHT TO BART
   in:  "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
   out: captured the gist, and LOST:
          - the party names (defaulted to
            "petitioner" / "defendant")
          - the jurisdiction
          - the sum of money

   DIAGNOSIS: surface-pattern summarisation
   cannot tell SALIENT from FREQUENT.
   ** In law, the salient fact is often the
      one mentioned exactly once. **


  B -- REGEX + PARTS OF SPEECH -> FORMAL LOGIC
   extracted: parties, dates, sections, court,
              loss amount
        |
        v
   Contract(Kashyap, Gandhi)
   DeliveryLate(Gandhi, Kashyap)
   Loss(Gandhi, 5000)
        |
        v
   DEDUCED: "Gandhi breached the contract with
   Kashyap due to late delivery, causing a loss
   of 5,000."

   WORKED -- but DID NOT SCALE.
   DIAGNOSIS: the predicate set is CONTEXT-SPECIFIC.
   Rules written for breach of contract do not
   transfer to theft or defamation. Finite
   vocabulary, unbounded sentences.
   ** You cannot hand-write the rules. **


  C -- THE HYBRID
   regex (fixed-form entities)
    + parts of speech (subject/object/action)
    + GloVe + cosine similarity
    + sentence templates
    + a paraphraser for fluency
        |
        v
   richer summary, key 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 across offence types, which is precisely the argument for pairing a learned model with a logic layer rather than choosing one. That is the neurosymbolic case, reached empirically rather than asserted.

What to say when this comes up

*It is a three-stage negative result and I think the negative part is the contribution. BART alone lost the party names, the 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, then did not scale, because a 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 learned model for extraction and a logic layer for the test. Its weakness is that it has no quantitative evaluation, and the abstract claims results it does not contain.*

4. DiagnoChat — and the move that turns it into an asset

Model	Accuracy	Note
Naive Bayes + hyperparameter search	90.11%	The paper's headline — and the only model that was tuned
KNN	88.94%	Untuned, 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 variant
Plain Bernoulli NB	66.73%	Worst of the eight

If it comes up, concede cleanly and move on

The paper concludes Bernoulli Naive Bayes is superior for this task. Its own table does not support that — untuned, it is the worst of eight at 66.73 percent, and the 90.11 figure comes from Naive Bayes plus a hyperparameter search no other model received. So the comparison attributes to the algorithm what may belong to the tuning. *That comparison is not like-for-like and I would not stand behind the claim. Only one model got a hyperparameter search. The defensible statement is that a tuned Naive Bayes reached 90 percent on this dataset, and an equivalently tuned KNN or logistic regression might well have matched it.* And the deeper point, which is the one worth making: accuracy is close to meaningless over 200-plus classes — it says nothing about rare conditions, which in a diagnostic setting are exactly the ones that matter. Macro-averaged precision and recall per disease is what the task needed.

And the move that converts a weak paper into evidence of judgement

Do not defend the methodology. Reframe the object. *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 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 honest, it is specific about two instruments, and it demonstrates the exact judgement this position needs.

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

What it does	Detail
The argument	India has no law specifically regulating connected devices. The IT Act 2000 does not define cybercrime at all, and its provisions are bailable unless read with the penal code — so the paper maps IT Act sections onto penal-code sections to show how prosecutions actually have to be constructed. Conclusion: India needs a dedicated framework
The comparative section — mention this part	EU: the GDPR for data generated by devices, the NIS Directive for infrastructure, the Cybersecurity Act 2019. US: the IoT Cybersecurity Improvement Act 2020 for federal procurement, and California's device-security law, the first in the US. This is comparative EU/US/India law, which is exactly this panel's register
The empirical part	A questionnaire to 50 respondents in Kolkata, distributed by messaging apps. 98 percent daily internet use, 30.6 percent reusing passwords, 69.4 percent claiming to understand connected devices, 100 percent believing cybercrime had increased, 100 percent supporting separate legislation
The case material	A 2018 baby-monitor intrusion; a regulator's recall of around 500,000 pacemakers over hackability; a casino database reached through an aquarium thermostat; the 2016 botnet that took down major services; heating cut to Finnish apartment blocks for a week; and an Indian case where a complaint filed in November 2021 remained untraced

The one thing to fix in your head before Wednesday

**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 in 2016 and applicable from 25 May 2018. This is an EU law panel, so this is the single most embarrassing error in your publication list if anyone spots it. It is a 2022 paper, you were second author and a second-year student, so it is forgivable — but do not repeat the error in the room. If GDPR comes up at all, name it correctly and give the regulation number. And if the paper is raised, concede 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.

What to say if Wisman raises it — and he is the likeliest 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 Indian 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 the survey properly, and not present fifty responses as though they supported the legal argument.*

And the live connection worth drawing

The regulatory vacuum your 2022 paper complained about has partly been filled in the four years since — and filled from an unexpected direction. 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 gap was closed 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 sits exactly between Wisman and Mirza. Also note: two changes have made the paper's Indian-law section obsolete — a data protection statute was enacted in 2023, and the penal code was replaced the same year, so every mapping in the paper refers to a repealed code. The nice detail: your own later offence-reclassification work is about exactly that transition — so the paper dated, and you went and worked on the thing that dated it.

6. The trap, and the answer that beats every individual result

If you volunteer weaknesses well, expect this follow-up

**A good interviewer will test whether you are performing humility or exercising judgement. The probe is usually some form of *so is any of it reliable?* Do not retreat. *Yes — differently in each case, and being able to say which is the point.*** *The ICC result I would defend as it stands, because a variance partition is a direct computation and the finding is large. The reporting-shock estimate I would defend with the caveat that a counterfactual time series assumes nothing else changed at that moment, and the case that prompted the 2013 amendment plausibly changed behaviour as well as recording. The literacy-gap causal estimate I would present as a point estimate that cannot exclude zero. The summarisation paper I would present as a design study rather than an evaluation. And the Mens Rea result I would state as a susceptibility finding with fifteen runs, not a model comparison.* *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.

7. If you remember eight things

The through-line: all of your work asks what the measurement is actually measuring. Say it. It turns a scattered CV into a research programme and it is true.
**Strike the word *proves*** and say unprompted that you would rewrite those sentences.
Dark Figure's one real hole: the literacy gap is not significant — p = 0.399 and p = 0.156 — yet the abstract claims its causal effect is proven. Identification and precision are separate; concede the wording and defend urbanisation at β = 0.371, p = 0.004.
Defend these three numbers without hesitation: Moran's I 0.095 at p = 0.151, ICC 0.6 percent, and +88,879 recorded cases a year from the 2013 amendment.
Mens Rea: 0 of 15 control, 9 of 15 ablation, 15 of 15 poisoned, and 0 cognizable intent — which is the finding. Volunteer that both admissions were self-judged and that Fisher at p = 0.206 licenses no model ranking. And correct the framing: mens rea does not transplant; the result is about the duty to give reasons.
The summarisation paper has no experiment despite the abstract claiming one. Reframe it as a design study — and its negative result, that hand-built logic does not scale across offence types, is the actual contribution.
DiagnoChat: only one model was tuned, so the superiority claim does not follow — and accuracy is the wrong metric over 200-plus classes. Concede it, then make the AI Act and Medical Device Regulation point, which converts it into an asset.
The IoT paper mis-names the GDPR. It is the General Data Protection Regulation, 2016/679, applicable from 25 May 2018. Concede it instantly if raised, and never repeat the error.