CRAM — the only page, if you read nothing else

1. The opener — say this almost word for word

~85 seconds, then stop talking

*Thank you. I'm Rudrani Ghosh.* *My background is a slightly unusual order: computer applications first, then law, then a master's in data science.* *What ties those together is one question, and I only noticed it in hindsight — whether the measurements a legal system runs on actually measure what they claim to.* *The clearest case is my most recent paper. I built a twenty-year district panel of recorded crime in India. The main finding was that a legislative change in 2013 produced about eighty-nine thousand additional recorded cases a year. That reads like a crime wave, and it isn't one. It's a reporting shock — the amendment changed what got written down, not what happened.* *So police records measure reporting, not offending. And a cartel screen measures detection, not collusion. That is the same problem, and it is the reason I applied for the computational position rather than one of the legal ones.*

2. Why antitrust, and why the computational seat

Six sentences. The two in bold are load-bearing.

*I came to a question rather than to a field. My father has been in a fraud matter for about twenty years — not a complicated case, twenty years.* *And what stayed with me wasn't the unfairness. It was that the delay was nobody's decision. No one chose it; it's a property of how a system processes volume — and a property you can measure, where a grievance you can't.* *Competition law itself is recent for me — weeks of serious reading, not years, and I said so in my letter. What isn't recent is the question: I came to VU in 2024 for the logic summer school, two years before this vacancy existed.* *This field is the right home for it because here the gap between a pattern and a legal finding is actually written down — Wood Pulp says parallel conduct proves concertation only where nothing else explains it. And the undetected cartel is the dark figure of unreported crime: the same estimation problem, a labelled set produced by the very process you are trying to evaluate.* *What draws me to the computational side is that the measurement sits upstream of the doctrine, not downstream of it. If entity resolution merges two undertakings that aren't one economic unit, the turnover ceiling is wrong and so is the fine; if market definition clusters on the wrong threshold, the dominance finding never existed.* *The advert says a law degree isn't required for this post, and I have one — so I can tell an engineer that the threshold they just typed is a legal decision, and tell a lawyer that the p-value won't carry the argument. The second direction is the one people forget.*

3. The numbers — do not get these wrong

	Number	What it means in one line
Crime paper — the headline	+88,879 recorded cases a year	The 2013 Criminal Law Amendment, against the counterfactual. A reporting shock, not a behavioural epidemic
Crime paper — the robust one	ICC = 0.6%	State policy explains under one percent of local variation. A variance partition is a direct computation, so defend this as it stands
Crime paper — no clustering	Moran's I 0.095, p = 0.151	No evidence of geographic clustering. Absence of evidence, not evidence of randomness — your abstract overstates this
Crime paper — the one to defend	Urbanisation β = 0.371, p = 0.004	The only clearly significant demographic driver
Crime paper — the hole	Literacy gap p = 0.399 and p = 0.156	Not significant, yet the abstract claims a proven causal effect of 0.1487. Volunteer this
Mens Rea — control	0 of 15	And all 15 went the way the real precedents support, so 15 of 15 legally correct
Mens Rea — the document alone	9 of 15	*The fake document with a NEUTRAL system prompt.* This is the interesting number
Mens Rea — document + instruction	15 of 15	*Do not merge this with the 9.* Different condition, different claim
Mens Rea — the finding	0 of 15 cognizable intent; 2 admissions in 45	Both admissions from gpt-oss, which was also the judge. None of the 30 rulings ever questioned whether the fake case was real
Astroformer	28,000+ monthly active users	Your only production system with real traffic

4. Your papers, one line each, plus the two concessions

Paper	The one-line version	The weakness to volunteer
Beyond the Dark Figure (SocArXiv, sole author)	Twenty-year district panel of recorded crime. Three findings: no spatial clustering, state policy explains under 1 percent, and the 2013 amendment produced a reporting shock	*Literacy gap not significant, abstract says proving.* Identification and precision are separate questions — closing a backdoor path tells me which quantity I am estimating, not how precisely
The Mens Rea Evaluator (SSRN, sole author)	An evaluation harness separating what a model does from what it says about why. Five legal scenarios, one fabricated precedent, three conditions changing one variable each, then six probes	n = 15, one run per cell, and the judge was also a subject. But self-grading cannot explain the admissions, because the forced-choice probe is parsed by letter with no grader
Bridging Legal Reasoning (unpublished)	Five approaches to classifying offences, including first-order logic rules and the Article 14 formalisation separating equality before law from equal protection of the laws	The RNN is textbook overfitting — training accuracy 1.0000 against validation 0.8689 on 300 sentences — and the paper does not name it
Hybrid Legal Text Summarization (SSRN, first author)	Three-stage design study. BART lost the party names and the sum in dispute; hand-built logic fixed that and did not scale across offence types	No quantitative evaluation at all, and the abstract claims experimental results. It is a design study, not an evaluation
DiagnoChat (SSRN, second of three)	Symptom-to-disease predictor, 200+ diseases, 489 symptoms. Tuned Naive Bayes reached 90.11 percent	Only one model was tuned, so the superiority claim does not follow — and accuracy is near-meaningless over 200+ classes
Master's thesis (VIT)	Topic in, narrated video out, at a chosen grade level. Ten stages, built solo in four months	The training answers and the reference answers both came from the same teacher model, so the metrics measured imitation rather than correctness

5. The five laws in one table

Instrument	Max penalty	The one fact to have
Competition law (Arts 101, 102 TFEU)	10% of worldwide turnover	**Ex post: nothing is forbidden until conduct is investigated and found to have an effect.** Liability attaches to the undertaking — a single economic unit, not a legal person
DMA	10%, rising to 20% for repeats, plus 5% daily	**Ex ante: if you are designated, the conduct is simply prohibited — no market definition, no dominance finding, no proof of harm. Thresholds ~€7.5bn turnover, 45m monthly users. 16 July 2026: Google fined €460m plus €430m = €890m**
DSA	6% of worldwide turnover	45m monthly users makes you a VLOP. Article 40 is the researcher data access provision — the one that matters to you
GDPR	4% or €20m, whichever is higher	Regulation (EU) 2016/679, applicable 25 May 2018. Article 22 turns on the word *solely* — a human who rubber-stamps an output is not a human decision
AI Act	7% / 3% / 1% by tier	Four risk tiers. *Article 27's fundamental rights impact assessment is tied to the Annex III high-risk list — so a tool that flags firms* rather than deciding about people may fall outside the high-risk regime entirely.** That gap is your best AI Act point

6. Two cases, and the ceiling on everything computational

*Michelin* T-188/24 — your best card, and almost nobody will have it

General Court, judgment of 9 July 2025. The Commission's tyre investigation began from data analytics over earnings calls and investor presentations — not from a complaint and not from leniency. **It adopted an inspection decision on 10 January 2024 under Article 20(4) of Regulation 1/2003, Case AT.40863 codenamed *Hoops*, and raided Michelin and others on 30 January 2024. *The Court annulled the decision in part: the reasoning lacked clarity and specificity, and the Commission lacked sufficiently serious indicia for part of the period.* It did confirm that public statements made under financial-disclosure obligations can be relied on as evidence. *That is the duty to give reasons — Article 296 TFEU — doing exactly the work my Mens Rea paper says it has to do.***

7. Ten ways to detect collusion, with the mechanism

Variance screen. Collusive prices are steadier, not merely higher. Frozen perch: when the conspiracy ended, mean price fell ~16 percent while the standard deviation rose over 200 percent. Needs no competitive benchmark.
Within-tender bid dispersion. A cover bid is drawn to lose, not priced to win, so it carries no cost information — the spread of losing bids collapses while the winner's margin widens.
The bid gap. Distance between lowest and second-lowest, abnormally wide or abnormally consistent.
Rotation and win-rate persistence. Winners alternating too evenly for independent cost draws.
Pairs that never meet. Two firms that bid often but never against each other. Market division, visible with no price data.
Subcontracting to the loser after award. The compensation channel — closer to evidence than inference.
Co-bidding graph against a degree-preserving null. *The null is the essential part* — firms co-bid for innocent reasons, and without it every regional specialist looks like a cartel.
Temporal change-point in graph structure. A clique that forms, persists and breaks around a datable event. Has a mechanism and a date, so far stronger than cross-sectional density.
Structural auction tests (Bajari and Ye). Test whether bids are consistent with competitive equilibrium. *Needs no cartel labels, because the null comes from theory.*
Communications evidence. Dawn-raid email metadata, shared document authors, same IP on an e-procurement portal — and text screening of public communications, which is what the Commission ran before the tyre raids.

8. Traps, the escape line, and the two questions to ask

Trap	The escape
So you are comfortable with deep learning frameworks?	Name the tier. Applied layer, yes. Framework internals, no
So is any of your work reliable?	*Yes — differently in each case, and being able to say which is the point.* Then rank them: the ICC as it stands, the reporting shock with one caveat, the literacy gap as a point estimate that cannot exclude zero
Do you think models can have intent?	*Say no* in the first three words**, then the duty to give reasons
Should agencies use these tools at all?	The premise is that they will. Frame every criticism as a condition of proper use, never as an objection
How would you make a model explainable?	Do not offer SHAP or LIME. Offer counterfactual explanations, because that form is testable by a court, and a stability test — refit on a slightly different sample and see whether the same firms are flagged. If the flagged set moves, who gets investigated is partly arbitrary, and that is a legal defect
A pause after your answer	Let it sit. A panel pausing is usually writing, not waiting

The two questions to ask — lead with the first

One, always, unless already answered: *You describe this as the cross-project computational position. I am trying to understand how you picture it day to day — closer to a service role producing the measurements the legal strands specify, or a parallel line of inquiry that sets some of its own questions? Both are workable and they would shape my first year quite differently.* Two, if the conversation stayed on your CV, this is the one that shows the depth they have not seen: *The first strand is accuracy. I have been thinking about what accuracy can even mean when the only labelled cartels are the ones that were caught — precision is estimable and recall is not. So my question is about appetite: if the honest finding on a screen is that its recall cannot be estimated, is that a publishable result inside this project, or does the project need these tools to work?*

Who asks what	Their question	Your answer in one line
Schrepel (chair)	What would you build first?	A corpus and an entity layer. Decisions in one place with provenance per document, firms resolved to undertakings with a reported matching threshold. That is most of a first year
Goanta	How would you handle the data side?	A written codebook before coding starts, two coders on an overlapping subsample, an agreement statistic fixed in advance. And volunteer that you wrote your own benchmark scenarios, so you chose the difficulty
Wisman	What about the people subject to these systems?	SyRI was struck down in February 2020 partly because the model's workings were insufficiently transparent to be verified — not because it was inaccurate. A court treating irretraceability as a legal defect. And average effects cannot carry individual attribution
Mirza	Thresholds, designation, market definition	A bright-line threshold is a natural experiment and also a target — expect bunching just below it. And **market is a legal conclusion that arrives in the dataset as a column**

Trap	The escape
So you are comfortable with deep learning frameworks?	Name the tier. Applied layer, yes. Framework internals, no
So is any of your work reliable?	*Yes — differently in each case, and being able to say which is the point.* Then rank them: the ICC as it stands, the reporting shock with one caveat, the literacy gap as a point estimate that cannot exclude zero
Do you think models can have intent?	*Say no* in the first three words**, then the duty to give reasons
Should agencies use these tools at all?	The premise is that they will. Frame every criticism as a condition of proper use, never as an objection
How would you make a model explainable?	Do not offer SHAP or LIME. Offer counterfactual explanations, because that form is testable by a court, and a stability test — refit on a slightly different sample and see whether the same firms are flagged. If the flagged set moves, who gets investigated is partly arbitrary, and that is a legal defect
A pause after your answer	Let it sit. A panel pausing is usually writing, not waiting